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.0005364 | Multibacillary leprosy by population groups in Brazil: Lessons from an observational study | Leprosy remains an important public health problem in Brazil where 28,761 new cases were diagnosed in 2015, the second highest number of new cases detected globally. The disease is caused by Mycobacterium leprae, a pathogen spread by patients with multibacillary (MB) leprosy. This study was designed to identify population groups most at risk for MB disease in Brazil, contributing to new ideas for early diagnosis and leprosy control.
A national databank of cases reported in Brazil (2001–2013) was used to evaluate epidemiological characteristics of MB leprosy. Additionally, the databank of a leprosy reference center was used to determine factors associated with higher bacillary loads.
A total of 541,090 cases were analyzed. New case detection rates (NCDRs) increased with age, especially for men with MB leprosy, reaching 44.8 new cases/100,000 population in 65–69 year olds. Males and subjects older than 59 years had twice the odds of MB leprosy than females and younger cases (OR = 2.36, CI95% = 2.33–2.38; OR = 1.99, CI95% = 1.96–2.02, respectively). Bacillary load was higher in male and in patients aged 20–39 and 40–59 years compared to females and other age groups. From 2003 to 2013, there was a progressive reduction in annual NCDRs and an increase in the percentage of MB cases and of elderly patients in Brazil. These data suggest reduction of leprosy transmission in the country.
Public health policies for leprosy control in endemic areas in Brazil should include activities especially addressed to men and to the elderly in order to further reduce M. leprae transmission.
| Leprosy is caused by Mycobacterium leprae, a bacillus transmitted by patients with multibacillary (MB) disease. Men and elderly are more likely to have MB leprosy, which has been attributed to an increased exposure to infection by male sex, decreased access of men to healthcare resulting in delayed diagnosis and long incubation period for MB disease. In this study, we found that the odds ratio for MB leprosy is two-fold higher for men than for women in all Brazilian states, independent of their endemic level. The same was observed for patients aged 60 or older compared to younger cases. Detection rates for MB leprosy remained higher for men and elderly patients timely detected (without physical disabilities), showing that late diagnosis is not enough to explain this association. Additionally, we showed that M. leprae load is higher in men than in women, despite early detection. These findings are relevant, because main activities to diagnose new cases of leprosy in Brazil have been focused on school surveys, detecting children who most likely have paucibacillary leprosy, which is non-contagious. To prevent transmission within the community, additional activities need to include groups at greater risk for MB leprosy. As such, we suggest that specific strategies for disease control should be adopted to effectively reach males and the elderly in endemic areas for leprosy.
| Leprosy is a chronic disease caused by Mycobacterium leprae infection [1]. In 2015, 210,758 new cases of leprosy were diagnosed in 136 countries and territories worldwide, of which 8.9% were in children [2]. Between 2007 and 2015, more than 110,000 people diagnosed with leprosy had physical deformities at diagnosis, an indicator of delayed detection [2,3].
The global number of leprosy new cases sharply dropped by 75% during the period of 2000–2006, which reflects especially an abrupt fall of cases detected in India, a country that contributes more than 50% of the world new cases annually [2,3]. Some authors have attributed this decrease to a reduction in active case finding activities and the adoption of new strategies to effectively reduce transmission of M. leprae has been emphatically recommended [4,5].
Leprosy presents a wide range of clinical manifestations, but for treatment purposes a simplified field operational classification based on the number of skin lesions is available: paucibacillary (PB) leprosy has one to five skin lesions, and multibacillary (MB) patients have six or more skin lesions [6]. Bacillary index based on the slit-skin smear analysis is a more accurate means of classification, but usually it is available only in more specialized health centers [6,7].
MB leprosy occurs in people with weak cell-mediated immune response against M. leprae, who develop high bacillary load and become the main source of infection [8]. Therefore, strategies to stop transmission should include measures to diagnose and treat MB cases. However, screening campaigns generally search for skin lesions with loss of sensation, even though in 30% of patients, especially in those with MB leprosy, this finding can be absent [9]. Surveys in schools are used as a strategy for early diagnosis [10], but MB cases are less common in children [11]. Thus, a question remains: Are current control strategies successful in effectively stopping M. leprae transmission?
In 2015 Brazil detected 28,761 new cases of leprosy, the second highest number of cases in the world [2,12]. The disease is still highly endemic in different areas of the country where its control must be improved. For these reasons, this study aimed to analyze epidemiologic patterns of MB cases in Brazil, in order to gather data to develop additional strategies to decrease M. leprae transmission.
This is an observational analytic non-concurrent secondary data study. Data of the national government database of leprosy cases diagnosed in Brazil (2001–2013) were used for the epidemiological trend evaluation of MB cases. The database is generated in all Brazilian States that have diverse capacities and expertise for leprosy diagnosis and classification. Therefore, to have a homogeneous group classified under the same standards, we included an analysis of bacillary loads obtained from a database at a leprosy referral center. At Souza Araújo Outpatient Clinic (Ambulatório Souza Araújo, ASA/Fiocruz, Rio de Janeiro, Brazil), slit skin smears are collected and read under standard protocols and only cases with positive bacilloscopy are classified as multibacillary. For this analysis we chose a different period of time (1990 to 2014) to include the largest possible number of cases.
Brazil monitors mandatory notifiable conditions using SINAN (National Notifiable Diseases Information System). For each new case, a report form is generated at local health units, and sent to the central health services where data are checked, digitized and transferred from municipalities to States, and to the national database [13]. Until mid-2015 de-identified information in the national databases was freely accessible through the Informatics Department of the Public Health Care System (DATASUS).
All data were obtained from DATASUS on June 2, 2015. New cases, defined by the Ministry of Health as cases of leprosy with no previous treatment [7], were included. Sex, age group and operational classification were used as classifying variables. Cases with missing information regarding one of these variables were excluded. Physical disabilities caused by leprosy were classified as grade zero when patients had no problem in eyes, hands or feet; grade one when there was anesthesia in hands, feet or eyes; and grade two when there were visible deformities in hands, feet or eyes, or severe visual impairment [14]. As disability grade is an indirect indicator of diagnosis delay [9], rates were calculated independently for patients without physical impairment (disability grade 0) to reduce the influence of late diagnosis in progression to MB leprosy. Given that detection rates for MB leprosy have been related to leprosy endemicity, which is variable in Brazil, we looked for the relation of sex, age groups and MB rates by state.
To evaluate bacillary load variations, all new cases treated at ASA/Fiocruz (1990 to 2014) were analyzed using de-identified data on sex, age, disability grade and bacillary index (BI) at diagnosis. BI is calculated using Ridley’s logarithmic scale for number of acid-fast bacilli per microscopic field in smears collected from ear lobes, elbows, and leprosy skin lesions. In this scale, each additional unit indicates a 10-fold increase in the number of bacilli and is scored from 0 to 6+, ranging from no bacilli per 100 microscopic fields to 1,000 or more bacilli per microscopic field. Final BI results are calculated as the arithmetic mean of skin sites collected in each patient [15]. In ASA/Fiocruz, four or six sites are examined, generally two ear lobes, one elbow and a cutaneous lesion. Cases classified as indeterminate or pure neural leprosy, which present higher probabilities for misdiagnosis, and those with missing information regarding any variable were excluded.
National data on leprosy were categorized by year of diagnosis, sex and age group. Population data were obtained from the IBGE (Brazilian Institute of Geography and Statistics) [16]. Age was assigned with DATASUS age group definitions: 0–4, 5–9, 10–14, 15–19, 20–39, 40–59, 60–64, 65–69, 70–79, and 80 or more years of age.
New case detection rate (NCDR) is the number of new cases per 100,000 people, per year. Sex ratio is the quotient of NCDR in men to NCDR in women. Mean NCDRs were calculated as the arithmetic mean of annual new cases from 2001–2013, divided by the population in mid-2007. Mean NCDRs were calculated by sex and by age group using arithmetic mean of cases (2001–2013) and corresponding population at mid-2007. Mean NCDRs were also calculated by operational classification.
Mean NCDRs by age groups were calculated separately for three time periods: 2002–2005; 2006–2009 and 2010–2013, to ascertain whether there were changes in age distribution related to reduction in NCDRs.
Data were exported to Microsoft Excel worksheets (version 14.0.4760.1.000/2010). Statistical analysis was performed in Openepi (version 3.03a, CDC, Atlanta, Georgia, USA, available at http://www.openepi.com) and SPSS (version 22, IBM Corp, Armonk, NY, USA) using chi (χ)2 test (Pearson) to compare rates between selected groups. Annual coefficient means were compared between men and women for all age groups using analyses of variance (ANOVA). Linear regression was used to test gender differences for the NCDR curves over time. BI was compared among study groups using Kruskal Wallis test. Confidence interval was established at 95%, thus a p-value less than 0.05 was considered to be statistically significant.
The protocol for this study was reviewed and approved by the Universidade Federal do Rio Grande do Norte Ethical Committee (CAA 06189612.9.0000.5537).
A total of 543,677 new cases of leprosy were reported in Brazil from 2001–2013, of which 99.5% (n = 541,090) had complete data available and were included in this study. Of those leprosy cases, 54.8% were men, 6.4% were children under 15 years and 17.5% were 60 or more years of age. Most cases (89%) were evaluated for disability grade at diagnosis, of whom 64% (n = 307,834) had no disability and 6% had disability grade 2. NCDR increased from 2001 to 2003, and then progressively decreased to 15.68 new cases/100,000 population in 2013. The decrease in NCDR for both women and men was similar over the 13 years (p = 0.061). The proportion of newly diagnosed leprosy cases who were MB increased 11.6%. The percentage of new cases with 60 or more years of age increased 6% while the proportion of new cases in children under 15 years of age was variable (Table 1).
During the study period, the number of cases in males was significantly higher than in females (p<0.0001). The odds of presenting with MB leprosy was twice as high in men as in women (OR = 2.36, CI95% = 2.33–2.38), which was similar for MB leprosy patients detected earlier, with disability grade zero at diagnosis (OR = 2.22; CI95% = 2.19–2.25). Patients aged 60 years old or more had twice the odds of being classified as MB compared with those under 60 years of age (OR = 1.99, CI95% = 1.96–2.02); a similar pattern was observed for MB leprosy patients without disabilities (OR = 1.69; CI95% 1.65–1.72).
Mean leprosy NCDRs reached high levels of endemicity in 18 out of 27 Brazilian States, with lower levels observed in the South and Southeast Regions (Fig 1). A significant association between MB cases and age above 59 years was seen in all Brazilian states (S1 Table). This pattern was similar in all states, either with high NCDR such as Mato Grosso (OR = 1.94, CI95% = 1.83–2.07), with a mean NCDR of 104.5 cases/100,000 residents or in states with a low NCDR as Rio Grande do Sul, (OR = 1.86, CI95% = 1.49–2.32), with a mean NCDR of 1.7 cases/100,000 population. When comparing with patients’ sex, the odds of being MB were significantly higher in males than in females regardless of the state’s level of leprosy endemicity. For example, in the North Region, where mean NCDR was 55.6 cases/100,000 population, men had odds of 2.45 (CI95% = 2.39–2.51) higher than women of being MB, whereas in the South Region (mean NCDR = 6.3/100,000 population) the odds were 2.13 (CI95% = 2.01–2.25).
Mean NCDRs by age group increased from 1.05 in 0–4 years to a peak of 45.22 cases/100,000 population in 65–69 years (p<0.0001), and was higher for people between 60–69 years of age in the three time periods evaluated (2002–2005, 2006–2009 and 2010–2013). NCDR overall have significantly decreased in all age groups (p<0.0001) (Fig 2). NCDR in both sexes was similar for children and adolescents, but NCDR was significantly higher in adult males than in adult females (Fig 3A), except for those aged 40–59 (p = 0.069). The largest difference in NCDR between males and females was in the 60 years or older age group (p<0.0001).
Analysis of the mean NCDR (2001–2013) by sex and operational classification according to age group (Fig 3B) indicated that the general mean NCDR was mainly influenced by male MB cases. For females (PB or MB) and for males with PB leprosy, mean NCDR remained under 20 cases/100,000 population in all age groups. However, the NCDR of males with MB, increased steeply above 19 years of age and peaked at 44.8/100,000 population in 65–69 years age group. These differences were also observed for leprosy cases detected with disability grade zero, ie. earlier diagnosed patients, with a progressive increase of NCDR with age for men with MB leprosy, while detection rates remained stable in PB leprosy with predominance of women in all age groups (Fig 4).
Bacillary index (BI) data from 2,253 leprosy cases diagnosed at the ASA/Fiocruz reference center were included (Fig 5A). BI was significantly higher in patients 20–39 years old compared with other age groups (p<0.0001), and it was higher in male than female patients (p<0.0001). Although patients with disability grade 1 or 2 had significantly higher BI than patients with disability grade zero (median test = 90.821, p<0.0001), BI remained significantly higher in males than in females independent of disability grade (median test = 14.423, p<0.0001 for grade zero; median test = 85.383, p<0.0001 for grade 1 or 2). When analyzing only MB cases (Fig 5B), BI was significantly higher in men than women in the 20–39 and 40–59 age groups (p<0.0001).
Decrease in NCDR in countries where leprosy was previously endemic, such as Norway, United States and Japan, was accompanied by an increase in sex ratio and in the percentage of elderly and MB patients among new cases [17,18,19]. In our study, sex ratio was stable over time, but the progressive reduction of annual NCDR together with progressive increase in percentage of elderly and MB cases could indicate decrease in leprosy transmission in Brazil.
Despite the hypothesis that global reduction of leprosy detection rates can reflect a decrease of active case finding activities, an abrupt fall in NCDR was not observed in Brazil as it was in India, where a reduction of 420,000 new cases was observed comparing 2000 to 2006 data, i.e. a reduction of 75% in only six years [4]. In Brazil, the decrease of 20,000 new cases observed between 2003 and 2013 (a 40% reduction), occurred during a period with significant decentralization of leprosy control activities in the country. From 2000 to 2011 it was reported a 284% expansion in the number of health centers that registered patients under treatment (from 3,327 to 9,445) [20].
It is worth noting that from 1980 to 2011 the reduction of leprosy cases in Brazil has occurred as a parabolic curve, without drastic decreases [21]. Moreover, grade 2 disability rate was reduced from 1.40 to 0.99 new leprosy cases per 100,000 population during our study period (2001–2013) [12], suggesting no increase in diagnosis delay. An analysis of leprosy detection rates from 1980 to 2009 in Amazonas State showed that cohorts born in more recent years had smaller risks of leprosy infection than older cohorts, with a declining trend of the relative annual reduction in children that disappeared after 29 years of age [22]. In 2013 and 2014 the Brazilian Ministry of Health introduced specific leprosy detection campaigns for school children from endemic municipalities that involved over than 6 million of students [23], but even though the national NCDR in children was lesser in those years than annual rates observed between 2000 and 2011 [12].
All of these findings may indicate an actual decline in national leprosy incidence, rather than a reduction of new case detection activities. Notwithstanding this decrease in leprosy national NCDR, Brazil is still endemic for leprosy, with hyperendemic or very high NCDR observed in 37% of States in 2015 [24], but with a profile similar to that considered to be of low endemic countries. Therefore, the predominance of multibacillary leprosy in elderly men seems to be a common characteristic to both levels of endemicity.
Male predominance in leprosy is reported in different parts of the world [1,25–27]. This association could be related to greater exposure of men to the M. leprae bacillus or by the lack of women undergoing a full physical examination in some cultures [26]. There are also reports of more MB leprosy in men [18,25–27]. In Brazil, this has been documented, and hypothesized to be related to men being less attentive to their health, contributing to later diagnoses with progression to MB leprosy [28]. However, our results show that rates of MB leprosy were higher in men than in women in leprosy per se as well as in early diagnosed cases. This association was observed in different states within Brazil regardless of background NCDR of leprosy.
A recent study using random-effects models to evaluate the outcome of infection with ten major human pathogens, including M. leprae, concluded that differences in response to infection were more associated with physiologic than with behavioral risk factors [29]. Interestingly, in our study sex ratio for leprosy was similar in people less than 20 years of age and then became progressively larger with age, a finding reported by other authors [18,26]. This may be related to a greater exposure of men to the bacillus after childhood, but could be also related to physiologic changes at adolescence, such as estrogen and testosterone levels.
Studies in experimental models of other intracellular pathogens, such as Leishmania sp [30], Paracoccidioides brasiliensis [31], Mycobacterium marinum [32], M. avium [33] and M. tuberculosis [34], demonstrated that while estrogen stimulates cellular immune response to these pathogens, with increase in antigen-specific CD4 T cells and IFN-γ secretion, testosterone stimulates production of anti-inflammatory cytokines associated with Th2 response, such as IL-10 and IL-4, and increase in antibody levels. In all experimental models tested more serious lesions and higher parasite burdens were observed in males or in females previously treated with testosterone [30–34].
Our study with data from the ASA/Fiocruz reference center showed that BI was higher in males, even when considering only patients without disability at diagnosis. This suggests that factors other than decreased opportunities for diagnosis may be involved in the development of MB leprosy and higher BI in men. This study found higher BI in men from the age of 20, with significantly higher bacterial load in men 20–59 years of age. A possible explanation for this finding is that testosterone which change during adolescence may be involved in creating an environment that could facilitate M. leprae growth and yields a higher bacillary load in men. Additional research on sex hormones and leprosy are very important to better discuss these findings.
In Norway, the peak of leprosy NCDR moved from 15–29 years of age in mid 1800’s to people older than 50 years in early 1900s [17], paralleling the overall decrease in the incidence of the disease. At the same time, there was an increase in the percentage of MB cases among the elderly. This is explained by less exposure to M. leprae among younger people and by a longer incubation period for MB leprosy. Thus, only the previously infected individuals developed especially this type of disease at later ages. This was also observed in other areas where leprosy declined [17–19,27,35], but we found a significant association between MB leprosy and elderly patients in all Brazilian States, including hyperendemic areas from the North and Center-West Regions, indicating that reduced exposure to infection may not be enough to explain this pattern of leprosy.
In addition, it is meaningful to observe that life expectancy in Europe which used to be only 36.3 years in 1850 [36], increased to 50 years of age in Norway at the end of the 1800s [37]. Leprosy incidence peaked in adults in Congo [38] and in children and adolescents in the Philippines [39] when life expectancy in those countries was under 41 years of age [36,40]. Our results in Brazil showed a higher NCDR in elderly in all 3 time periods studied (2002–2005, 2006–2009 and 2010–2013), where life expectancy has been above 70 years of age since the 2000’s [40]. Similar results were observed more recently in Japan [19], Mexico [27] and Korea [41]. These findings may suggest that differences of leprosy NCDR by age could be related to overall life expectancy of the population.
Usually the association between leprosy and the elderly is discussed only as it relates to decreased leprosy transmission and longer incubation period for MB disease, but other potential contributing factors could include age-related immune changes and increased susceptibility to infectious diseases [42]. For example, the elderly can have impaired monocyte and neutrophil function, decrease in CD28 co-stimulatory molecule expression [43], reduction in phagocytic capacity, decreased antigen presentation and a change in cytokine profile from Th1 to Th2 [42]. Thus, key steps required for defense against M. leprae are potentially compromised with aging, and may be involved in the higher incidence of leprosy in older age groups.
As MB patients transmit M. leprae, we advise targeting specific case detection strategies to men and elderly people to further reduce transmission of leprosy, in addition to preventing disability with early diagnosis and treatment.
In Brazil, the National Health System is based on family health teams, who are responsible for leprosy control among other health issues. These teams are multidisciplinary and usually composed of doctors, nurses and community health workers. The team visits household units for general medical care [44]. This opportunity should be considered an important means to exam elderly people, who sometimes have less access to health services due to decreased mobility and independence [45]. In addition, this knowledge can be useful for active search of the source of infection for leprosy cases under 15 years of age among the elderly men in the households.
Rapid serological tests with high sensitivity to detect MB leprosy are currently available [46, 47]. They are rapid immunochromatographic flow tests that detect IgM antibodies against specific M. leprae antigens and could be performed as point-of-care serologic evaluations during leprosy active search activities. Although they are not diagnostic tests because they can be positive in healthy people who may never develop leprosy, sera and blood from MB leprosy cases show strongly positive results. Thus, they can be used as an adjunctive tool in diagnostic campaigns, especially for specific population groups where MB leprosy is more frequent.
These strategies can be employed to improve MB leprosy detection and to accelerate the reduction of this important public health problem in Brazil.
This study utilized data generated and entered by hundreds of different people, which needs to be considered in interpretation of the results. An analysis of the leprosy database in Brazil (SINAN) has identified problems linked to updating of data after case notification to the system, such as the date of completion of leprosy therapy; however, no problems where observed related to the software or data transfer [13]. Considering that we used secondary data produced in different settings nationwide, it is important to keep in mind that leprosy classification methodology may vary according to the diagnostic capacity of the center. In primary care settings, diagnosis and operational classification are mainly based on clinical findings, while in references centers, confirmatory and complementary tests, such as skin biopsy and slit-skin smear, can be performed. Nevertheless, during the study period, there were no major changes in the National Program’s recommended diagnosis and classification criteria. Although this may be considered as a limitation, the large number of cases included in the analysis should reduce the impact of differences in classification criteria between centers.
|
10.1371/journal.pcbi.1003377 | Interference and Shaping in Sensorimotor Adaptations with Rewards | When a perturbation is applied in a sensorimotor transformation task, subjects can adapt and maintain performance by either relying on sensory feedback, or, in the absence of such feedback, on information provided by rewards. For example, in a classical rotation task where movement endpoints must be rotated to reach a fixed target, human subjects can successfully adapt their reaching movements solely on the basis of binary rewards, although this proves much more difficult than with visual feedback. Here, we investigate such a reward-driven sensorimotor adaptation process in a minimal computational model of the task. The key assumption of the model is that synaptic plasticity is gated by the reward. We study how the learning dynamics depend on the target size, the movement variability, the rotation angle and the number of targets. We show that when the movement is perturbed for multiple targets, the adaptation process for the different targets can interfere destructively or constructively depending on the similarities between the sensory stimuli (the targets) and the overlap in their neuronal representations. Destructive interferences can result in a drastic slowdown of the adaptation. As a result of interference, the time to adapt varies non-linearly with the number of targets. Our analysis shows that these interferences are weaker if the reward varies smoothly with the subject's performance instead of being binary. We demonstrate how shaping the reward or shaping the task can accelerate the adaptation dramatically by reducing the destructive interferences. We argue that experimentally investigating the dynamics of reward-driven sensorimotor adaptation for more than one sensory stimulus can shed light on the underlying learning rules.
| The brain has a robust ability to adapt to external perturbations imposed on acquired sensorimotor transformations. Here, we used a mathematical model to investigate the reward-based component in sensorimotor adaptations. We show that the shape of the delivered reward signal, which in experiments is usually binary to indicate success or failure, affects the adaptation dynamics. We demonstrate how the ability to adapt to perturbations by relying solely on binary rewards depends on motor variability, size of perturbation and the threshold for delivering the reward. When adapting motor responses to multiple sensory stimuli simultaneously, on-line interferences between the motor performance in response to the different stimuli occur as a result of the overlap in the neural representation of the sensory stimuli, as well as the physical distance between them. Adaptation may be extremely slow when perturbations are induced to a few stimuli that are physically different from each other because of destructive interferences. When intermediate stimuli are introduced, the physical distance between neighbor stimuli is reduced, and constructive interferences can emerge, resulting in faster adaptation. Remarkably, adaptation to a widespread sensorimotor perturbation is accelerated by increasing the number of sensory stimuli during training, i.e. learning is faster if one learns more.
| Transformations that map sensory inputs to motor commands are referred to as sensorimotor mappings [1]. While sensorimotor mappings are already formed at early stages of development [2], they are subject to modifications, since the brain, the body and/or the environment are constantly changing. Plasticity in sensorimotor mappings has been extensively studied in situations where subjects receive sensory feedback during the task, allowing them to correct their motor actions and to adapt to the induced perturbation. These include visuomotor rotation [3], reaching movements under forcefields [4], adaptation in a smooth pursuit eye movements [5], prism adaptation [6], and pitch perturbation in songbirds [7] and in humans [8].
Although these studies involve different sensory modalities and different effectors, they are similar in the sense that they all have sensory goals (targets) and a motor gesture is made to reach the target. They consist of three phases namely a standard phase, in which subjects perform the task under regular conditions followed by an adaptation phase, where subjects perform the same task under the perturbed condition and a washout phase during which the perturbation is removed, and the subject readapts toward baseline. Remarkably, in all these three phases, movements display substantial trial to trial variability. Recent theoretical as well as experimental studies suggested that this variability plays a crucial role in sensorimotor learning and adaptation processes [9]–[11].
Another issue concerns the ability of subjects to generalize the adaptation from one context condition to a different context. This has been investigated by testing how subjects perform upon presentation of sensory stimuli that were not present during the adaptation phase [12], [13]. Generalization is usually good for sensory stimuli that are similar to the one used during adaptation and degrades as the sensory stimuli become different [3], [14]. Remarkably, subjects can even perform worse than in baseline (negative generalization) for sensory stimuli which are very different from those which was presented to the subject during adaptation. This has been observed, for instance, in motor reaching tasks, when the tested stimulus is presented in a direction which is opposite to the adapted direction [4], [14].
The above mentioned studies implicitly assumed that the neural mechanisms for adaptation are driven by a sensory feedback, which supplies a continuous error signal to the subject. Yet, recent studies show that adaptation is possible even without any sensory feedback, when only a binary reward that informs on a success or a failure of a trial is provided to the subject [15]–[17]. Moreover, recent experimental works suggest that reward based mechanisms also affect the adaptation dynamics in sensorimotor tasks even when a sensory feedback is available [18], [19].
However, and not surprisingly, adaptation relying solely on rewards at the end of a trial is more difficult than when a sensory feedback on the performance is provided continuously during the task, as adapting with sensory feedback conveys more information regarding errors. For instance, when visual feedback is available in visuomotor rotation tasks, subjects adapt to large perturbation (e.g. 30 degrees) in a few dozen trials [3], [20], while in the absence of such feedback, but with binary (success or a failure) reward feedback, subjects find it notoriously difficult to adapt. Recent studies, nevertheless, have shown that it is possible to adapt to large perturbations relying solely on rewards if the size of the perturbation is slowly increased between rewarded blocks of trials [17], [21]. The fact that progressively increasing the amount of perturbation makes it possible to adapt, even when the perturbation is large, is reminiscent of the classical shaping strategy [22]. In shaping, the difficulty of the task is increased gradually in order to accelerate learning, or to even make it possible. Although shaping is routinely used in laboratories when training animals to perform complex sensorimotor and cognitive tasks [23]–[25], it is only in recent years that it started to be explored in a theoretical framework [26]–[28].
What neural mechanisms could be involved in this reward based learning? Recent experimental evidence [29]–[31] indicates that rewards modulate local synaptic plasticity via global neuromodulatory signals, e.g. dopamine. When combined with the popular idea that synapses are modified according to Hebbian rules, this leads to the hypothesis that reward signals interact with local neuronal activity to modulate synaptic efficacies [32], [33]. This theoretical paper aims to provide qualitative as well as quantitative insights into the conditions in which sensorimotor adaptation relying solely on rewards can take place. More specifically, we assume that a local learning rule based on the coactivation of pre and postsynaptic neurons is gated by a binary reward signal is the neural basis for modifications of synaptic efficacies [32], [34], [35].
We focus here on adaptation to a rotation during reaching movements where subjects are asked to move a cursor on a screen to bring it within a circular target while the cursor trajectory is rotated (perturbed) by some angle with respect to the hand trajectory. These perturbation tasks are classically used in behavioral studies of sensorimotor adaptation [3]. We consider a simplified network model of this task where adaptation relies solely on binary rewards [17]. The simplicity of the model allows us to analytically study several aspects of the adaptation dynamics. Combining these results with numerical simulations enables us to investigate the ways in which the learning dynamics depend on the model parameters. The key question is how the dynamics of adaptation are affected when the task involves multiple targets. Four main findings are reported: interferences can occur when adapting to multiple stimuli, interferences can slow down the adaptation dynamics dramatically, this depends on the (binary, stochastic) reward, and the slow down can be overcome by using shaping strategies.
We consider the classical rotation experiment [3] in which a subject has to move a cursor on a screen to bring it within a circular target with a radius of ; see Figure 1A. At the beginning of the experiment there is no discrepancy between the movement of the hand and the movement of the cursor. We assume that the subject is able to generate the appropriate hand movement to perform the task correctly. A perturbation is then introduced, so that the cursor trajectory is rotated by an angle γ with respect to the hand trajectory. The subject has to adapt his movements to this new condition.
In the present work, we focus on the case where the subject receives no visual feedback about the trajectory of the cursor. The only information on performance is a reward provided by the experimentalist at the end of a trial, according to the location of the cursor with respect to the desired target.
Our simplified model for a network which generates the reaching movement is depicted in Figure 1B. Its input layer consists of sensory neurons tuned to the location of the target. It has the geometry of a ring: the preferred direction (between and ) of a neuron corresponds to its location on the ring (see Eq(2)). Hence, when a target appears, the population activity profile in the input layer peaks around a location which is also the target direction. For simplicity we assume that the tuning curves of all the neurons have the same shape. Therefore, the shapes of the population activity profile and the tuning curves are identical. In particular, the tuning width, , is also the width of the activity profile.
The output layer consists of two linear units. Their activity encodes the coordinates of the endpoint of the hand movement in the two dimensional environment. The connectivity matrix implementing the sensorimotor mapping between the input and the output layer is denoted by . In addition to their feedfoward inputs from the first layer, the output units also receive a Gaussian noise, (see Eq(4)), where is the of the noise (also referred to hereafter as the noise level). The vector representing the endpoint of the cursor is obtained by rotating the output vector of the second layer, , by an angle (2×2 rotation matrix- ).
The reward, R, delivered at the end of the movement, depends on the distance between the cursor and the target. Unless specified otherwise it is binary: for a successful trial, i.e. if the squared distance is smaller than the target size, and , otherwise. The target size is controlled by the parameter and therefore is referred to as the target size in the text.
Following trial t, the network adapts to the rotation by modifying the connectivity matrix, , according to the reward-gated synaptic plasticity rule [32], [36]–[38]:where η is the learning rate, is the noise in the output layer and is the activity of the input layer in response to the presentation of a target in direction . We will assume that the initial value of the connectivity matrix is such that without noise, the network performs the task perfectly for all target directions when (See Eq(9)). More details about the model are given in Materials and Methods.
The simplicity of the model allows for analytical calculations in the limit of small targets and a better understanding of the learning dynamics. However, the results reported here are grounded on the assumption of a reward-modulated learning rule and are qualitatively independent of the simplifying assumptions used to construct the model. For instance, as shown in Figure S2, the results still hold qualitatively in a more complicated network architecture with a different decoding scheme.
We first consider the case where the network has to adapt to a rotation of the cursor when only one target is presented. Figure 2A (left) plots the evolution of the error (see Eq.(5)) with the number of trials, hereafter referred to as the learning curve, while the network adapts to an imposed rotation with an angle . On the right panel we plotted for the same parameters the learning curve of the directional error, which takes into account only the direction of the movement.
The error is large at the beginning of the process and decreases with the number of trials. Importantly, the dynamics strongly depend on the noise. For a low noise level (Figure 2A, ), the error remains large for many trials and learning is slow. When the noise level is higher (Figure 2B, ) the error declines faster. However, this comes at the cost of increasing the error after learning: the median of this error, called hereafter the final error (see Materials and Methods), is larger when the noise level is larger. Similarly, the probability that the network will perform the task successfully, improves more rapidly with the number of trials for than for , but at very long time it is larger in the latter () than in the former () case.
The learning curves plotted in Figure 2A–B were obtained for particular realizations of the noise, . To provide a statistical characterization of these dynamics, we estimated the distributions of the logarithm of the learning duration () over many realizations of the noise (see Materials and Methods). As shown in Figure 2D, this distribution shifts toward longer learning duration as the noise level decreases.
Figures 2A and 2C plot the learning curves for and for the same noise level. The learning is substantially faster for but the final error is larger in this case. This is because when the target size is large, a reward might also be delivered for less precise movement, i.e., for large errors. Figure 2E plots the log learning duration and the final error averaged over realizations vs. the target size: when increasing the target size, the learning duration rapidly decreases, whereas the final error increases.
When the noise level or the target size are increased, the dynamics are typically faster because the probability of generating rewarded trials at the beginning of the learning is larger. As this probability increases, the time for the network to generate a rewarded trial decreases, leading to more updates in the connectivity matrix ; hence the probability of the following trials to be rewarded increases further. This argument can be made more quantitative if one considers how the time to get the first reward depends on and . It has a geometrical distribution with a parameter (see Eq.(10)), which is the probability to get the first reward. Lower values of increase the expectation time to the first reward, and thereby the learning duration. When the noise level is low and the initial error is larger than the target size, the network explores a small region of the two dimensional space and the probability of getting a reward is small. In contrast, for very large noise the target is missed most of the time. The probability therefore varies non-monotonically with the noise level (Figure 2F). The dependency on target size is simpler: increases monotonically with target size, as it is more likely to reach a larger target.
What is the learning dynamics when the subject has to perform the task for two targets ? How does learning the task for one of the targets affect learning the other one? We addressed these questions in numerical simulations, in which two targets were presented at an angular distance, , at consecutive times. Similar results were obtained when the targets were presented in a random order with equal probability.
How does the learning duration, i.e., the time to learn the task for all the presented targets, vary with the number of targets? We simulated the learning of m targets, whose directions were evenly distributed between and . We took a small target size (), so that up to non-overlapping targets could be considered (for targets presented on a circle with radius 1).
Figure 11A plots the average time to learn the entire task in terms of the total number of target presentations for a fixed noise level and different values of tuning widths. It shows a non-monotonic dependency with the number of targets. This contrasts the monotonically increasing learning duration when targets are learned independently with the same noise level and target size (dashed line).
We explored the reward-based component in adaptation processes in a setting in which a subject has to adapt reaching movements to a rotation when the only information available is the location of the target and a binary reward signal indicating success or failure on a trial [17]. The subject thus has to adapt to the perturbation by relying solely on the reward. In the framework of a simplified model of a neural network learning the task, we investigated the ways in which the adaptation dynamics depend on the noise level in the network, the target size, the size of the perturbation and the shape of the reward function. The key finding is that if the network has to adapt simultaneously to several target locations, constructive or destructive interferences between the different movements may occur. Such destructive interferences may result in a severe slowdown in the adaptation process, but this slowdown can be mitigated if the reward changes more gradually from a success to a failure around the target.
If the motor variability is not large enough with respect to the target size and the amount of perturbation (Figure 2), it takes a long time for the network to generate rewarded trials and to update its connectivity matrix. This results in slow adaptation and may be the reason why adaptation in the absence of visual feedback is notoriously difficult for subjects when the rotation angle is too large. For example, at the noise level and target size reported in [17], the probability to generate a rewarded trial in less than trials for a rotation of is essentially zero.
The time to adapt also depends on the size of the change in synaptic strength on each rewarded trial; i.e., on the learning rate parameter. We showed that perfect adaptation to one target (i.e. performance in the absence of noise) is possible only when the (normalized) learning rate is smaller than 1. A high learning rate leads to decreased performance and eventually fully impedes adaptation (Figure 3). Therefore, the extent to which adaptation can be accelerated by choosing a large learning rate is limited.
Adaptation is faster for large noise. On the other hand, if the noise is too large, final performance is impaired. Interestingly, motor areas display high variability at the early stages of learning, which becomes smaller afterward. This has been observed in reaching tasks in monkeys [39], as well as in song acquisition in songbirds [40]. Our study suggests that this change in noise level during learning can be functionally important to making a compromise between fast adaptation and good performance.
We showed that when adapting to multiple targets, learning the task for one target can impair performance on other targets due to destructive interference. As a result, the probability that the network will generate a rewarded trial for these targets decreases. Therefore, in this case the same noise level that allows exploration of one movement direction is insufficient when adapting to two or more targets, resulting in a delayed learning effect. Interestingly, when the network starts to adapt to the perturbation to the second target, it does not deteriorate the performance of the network on the first target that was already learned. This is because the network keeps generating rewarded trials for the first target and prevents the connectivity matrix from changing in the wrong direction for the first target.
We also showed that there are cases where the interference that occurs when multiple targets are presented is constructive. In fact, the strength and the nature of the interference depend on the similarities in the distance between the targets (the physical stimuli) and in the overlap of the tuning curves (the neural representations of the stimuli). Adding more targets creates constructive interference and therefore can accelerate adaptation.
Models of sensorimotor control and learning frequently assume minimizing a squared error function. This is convenient because of analytical or computational simplicity [13], [14]. However, it was shown that although these models can be a good approximation they tend to penalize large errors excessively [41]. In contrast, we chose to explore adaptation with a binary reward function, as used in experiments. Our results and predictions stem from the shape of the reward function. Specifically, they do not depend qualitatively on the specific choice of the distance error used, but are based primarily on the fact that the reward function varies sharply with the distance to the target center. The dynamics of the adaptation to more than one target depend on the overlap between the tuning curves of the input neurons. However, the precise shape of the tuning curves is not crucial and the results are unchanged if one replaces the Von Mises function we used here with any other tuning curve function, such as a cosine tuning curve (see e.g. Eq. 23).
As a matter of fact, the results we describe are the outcome of the following: 1) the same system is used to learn the task for several targets, leading to interference which depends on the way in which the targets differ physically as well as in their neuronal representation and 2) learning the task for one target can deteriorate performance on another target such that the information provided by the reward when attempting to learn the task for it becomes small, thereby delaying the learning. These two properties of the learning process are not specific to the simple model we investigated here.
In our model, the latter property stems from the fact that the reward varies sharply with the error. The learning rule we used is part of a general family of gradient-like reinforcement learning rules; i.e., learning rules that on average form a gradient ascent on the reward function [35]–[37]. In fact, learning with an on-line Gradient Ascent algorithm with a sigmoidal cost function can result in similar effects (Text S1; Figure S1). It might be claimed that plasticity also occurs when no reward is delivered [42]. Therefore, we also verified that the phenomenology of the model remains qualitatively the same when instead of using a reward function (unpublished data). Note that to avoid a drift of the output vector which occurs when , the synaptic weights must be normalized in this case after each trial. Another extension of our model would be to use a reward prediction error instead of an instantaneous reward; e.g., by subtracting a running average of the reward from the instantaneous reward. Delayed learning also occurs with this type of learning rule (results not shown). In fact, previous works have argued that this modification does not affect most of the qualitative behavior of the algorithm [32], [36]. However, it should be noted that in the case of multiple targets, computing the running average of the rewards over all targets is an additional source of interference, as shown recently in [35]. To avoid this, the running average of the reward needs to be monitored for each target separately.
We focused on the learning dynamics in a feed-forward network of linear neurons with only two layers. We chose this architecture for the sake of simplicity. However, we verified that similar qualitative behaviors in terms of interference and delayed learning occur in a network model in which an intermediate layer consisting of nonlinear neurons was added, and in which a decoder provides the angle of reach movement instead of a vector (Text S1, Figure S2 and unpublished data). Note that in the framework of this more complex model, the noise can be unambiguously related to neuronal variability whereas in the simplified two-layer model considered in our paper, the noise is in the decoder.
One limitation of our work is that we did not model the trajectory of the movement and/or the muscle activation patterns needed to produce movements [43]. However, we expect that delayed learning and interferences also occur in a more detailed model of movement production, such as the one used in Legenstein et al. [34].
A reward-based component in a sensorimotor task was shown to be involved in adaptation to rotations even when detailed spatial information regarding the error was provided to the subject [18], [19]. We investigated the ways in which neural possible mechanisms that reinforce successful actions affect adaptation dynamics. This type of reward-based mechanism was also studied in [17]. In this experiment, subjects adapted without visual feedback to a gradually increasing rotation of every 40 trials, up to an rotation. Our modeling results are in line with these experiments (Figure 4B). We thus predict that shaping the reward also accelerates adaptation.
Besides demonstrating that adaptation relying on rewards is possible by utilizing a gradual rotation paradigm, the Izawa and Shadmehr [17] results suggested that there is no change in the perceived sensory consequences of the motor commands; i.e., there should be no change in a “forward model”. Therefore, in [17] adaptation was modeled by an action selection rule. Our model is similar to the latter, as we focused on the reward-based component during adaptation. However, our model differs in that it is value-free, whereas in [17] it involved value-based reinforcement learning. Nevertheless, our model can also account for the experimental results reported in [17] for one target (see Text S1, Figure S3). Moreover, it allowed us to investigate the generalization curve and possible interference during adaptation for multiple targets.
The key finding of this theoretical work is that if a reward-modulated learning rule underlies adaptation, interferences are likely to be observed when learning multiple targets with a binary reward. It would be valuable to explore whether such effects occur in reward-based sensorimotor adaptation experiments with multiple sensory stimuli. We predict that for a binary reward function, destructive interferences will be observed if the neurons that encode the stimuli have broad tuning curves. These interferences are a dynamical counterpart of the generalization function and might result in a dramatic slowdown because of the abrupt change in the reward from success to failure around target size. We also predict that adding more targets should accelerate adaptation (Figure 11). From the learning curve of adaptation to one target, the rate and variability in which subjects adapt can be estimated. We predict that at parity of variability, subjects with larger learning rates will tend to display more destructive interferences and therefore slower adaptation to two targets (see Eq. (23)). By contrast,if the tuning curves are very narrow, destructive interferences are unlikely to be found. However, even in this case, when the stimuli are sufficiently close, constructive interferences should be observed. In this case as well, adding more targets should accelerate the adaptation.
Another prediction is that if adaptation is driven by reward modulated plasticity rules similar to the one we used here, smoothing the reward function should reduce interferences. In our model, this stems from the assumption of a reward modulated learning rule and not from the simplifying assumptions we made in constructing the model. Therefore, we suggest that testing this prediction could shed light on the synaptic mechanisms underlying adaptation tasks.
Finally, the location of the reward-based mechanism involved in adaptation could be the cortex-basal-ganglia network. As a matter of fact, there is evidence for the involvement of this network in pitch shift adaptation in songbirds. Although the neural correlates for adaptation in songbirds are unknown, when an auditory feedback is available to songbirds (by using miniature headphones [7]), the anterior frontal pathway, which is the avian homologue of the cortex-basal-ganglia network [50], is essential for adaptation based solely on binary rewards [15], [16]. Thus, exploring the behavioral and neural differences in auditory feedback versus binary reward adaptations in pitch shift experiments in songbirds may help reveal the neural mechanisms for reward-based adaptation.
We consider a motor reaching task (see Figure 1A) in which a subject manually controls the location of a cursor on a screen to bring it within a circular target of radius [16]. The target location is characterized by a two dimensional vector of norm 1 (we assume that the target is always at distance 1 from the center of the screen) and direction . In a standard block of trials, the direction of motion of the cursor and the hand are the same. We assume that the subject is able to perform the task perfectly in this case. In a rotation block of trials a perturbation is introduced: the movement of the cursor on the screen is now rotated by an angle with respect to the hand movement. To overcome this perturbation the subject must move his hand in a direction with respect to the target. Here we focus on the case where there is no visual feedback (the cursor is not on the screen): the only information the subject receives about his performance is provided by a reward signal delivered by the experimentalist [17].
We consider a simplified computational model of a network performing this reaching task, see Figure 1B. The input layer of the network encodes the sensory information regarding the direction of the target, . It is composed of directionally tuned neurons labeled by their preferred direction, . For simplicity, we assume that the shape of the tuning curves is the same for all neurons: upon presentation of a target in direction the activity of neuron is . We take:(2)where characterizes the width of the tuning curve and is the peak response of a neuron. The width of the tuning curves at half of its maximal activity relative to the baseline (half bandwidth) in this case is:(3)
The second layer of the network encodes the location of the endpoint of the hand movement. It consists of two output units whose activities, and , represent the two components of the hand position, . Upon presentation of a target in direction :(4)where is the connectivity matrix between the two layers, denotes the N dimensional vector of the input layer with components , and is a Gaussian noise. The location of the cursor at the end of the movement is related to by a rotation matrix, , of angle . Therefore, the squared distance between the endpoint location of the cursor and the center of the target is:(5)where . This quantity will be used to measure the error with which the network performs the reaching task. It is also useful to define the noiseless error:(6)where . This quantity measures the error if the noise is suppressed.
Upon presentation of a target in a direction θ at trial t, the network performs the task and a reward R is delivered according to the outcome:(7)
The matrix is then modified according to a reward-modulated learning rule:(8)where is the learning rate. This learning rule can be derived in a REINFORCE framework [36].
We assume that at the beginning of learning , when there is no rotation, the network is able to perform the reaching task with zero noiseless error for all targets. When all the Fourier components of are non-zero, this constraint fully determines :(9)where is the first Fourier component of the tuning curves. To get Eq. 10, one needs to calculate the Fourier expansion of by using the constraint:for each of the possible target directions, . When some of the Fourier coefficients of the tuning curve function are zero, e.g. when using a cosine tuning curves, is determined up to the Fourier coefficient that are irrelevant to the above constraint. This does not affect the learning dynamics.
In the numerical simulations described in this paper, the input layer consists of neurons. We normalized the tuning curves (parameter in Eq. (2)) such that remains constant () when changing . This was done to guarantee that the time to learn one target does not depend on the tuning width.
|
10.1371/journal.ppat.1002725 | Shedding of TRAP by a Rhomboid Protease from the Malaria Sporozoite Surface Is Essential for Gliding Motility and Sporozoite Infectivity | Plasmodium sporozoites, the infective stage of the malaria parasite, move by gliding motility, a unique form of locomotion required for tissue migration and host cell invasion. TRAP, a transmembrane protein with extracellular adhesive domains and a cytoplasmic tail linked to the actomyosin motor, is central to this process. Forward movement is achieved when TRAP, bound to matrix or host cell receptors, is translocated posteriorly. It has been hypothesized that these adhesive interactions must ultimately be disengaged for continuous forward movement to occur. TRAP has a canonical rhomboid-cleavage site within its transmembrane domain and mutations were introduced into this sequence to elucidate the function of TRAP cleavage and determine the nature of the responsible protease. Rhomboid cleavage site mutants were defective in TRAP shedding and displayed slow, staccato motility and reduced infectivity. Moreover, they had a more dramatic reduction in infectivity after intradermal inoculation compared to intravenous inoculation, suggesting that robust gliding is critical for dermal exit. The intermediate phenotype of the rhomboid cleavage site mutants suggested residual, albeit inefficient cleavage by another protease. We therefore generated a mutant in which both the rhomboid-cleavage site and the alternate cleavage site were altered. This mutant was non-motile and non-infectious, demonstrating that TRAP removal from the sporozoite surface functions to break adhesive connections between the parasite and extracellular matrix or host cell receptors, which in turn is essential for motility and invasion.
| Malaria infection begins with the bite of an infected mosquito which inoculates sporozoites into the skin. Sporozoites then go to the liver where they invade hepatocytes and replicate, ultimately leading to the blood stage of infection. Sporozoites are motile and actively invade hepatocytes using a unique form of motility called gliding motility. The mechanism by which the parasite moves forward is somewhat similar to a treadmill and the sporozoite protein TRAP, is key to this process. Its extracellular portion binds to host proteins while its intracellular portion binds to the parasite's motor. As the motor moves the protein rearwards, the sporozoite moves forward. It follows that the extracellular adhesive interactions of TRAP must ultimately be disengaged for forward movement to occur. We have generated mutant sporozoites that can only partially disengage these parasite-host adhesive interactions and find that these sporozoites have a halting, constipated movement. Following this, we generated a mutant that cannot disengage these interactions at all and these sporozoites are nonmotile and noninfectious. Lastly we found that a parasite rhomboid protease, ROM4, is on the surface of the sporozoite and thus may be responsible for TRAP cleavage and shedding from the sporozoite surface. Overall, our results demonstrate that robust gliding motility requires the disengagement of adhesive interactions.
| Malaria is one of the most important infectious diseases worldwide, causing an estimated 500 million clinical cases and 800,000 deaths annually [1]. Plasmodium species, the causative agents of malaria, belong to the phylum Apicomplexa, whose members include other human pathogens such as Toxoplasma gondii and Crytosporidium species. The Apicomplexans are obligate intracellular parasites and the invasive stages of these protists, called zoites, actively enter host cells using a unique form of locomotion called gliding motility.
Gliding motility is a substrate-dependent form of locomotion that does not involve significant change in cell shape and is powered by a subpellicular actomyosin system linked to the zoite surface through one or more members of the Thrombospondin Related Anonymous Protein (TRAP) family (reviewed in [2], [3]). TRAP family members are type I transmembrane proteins bearing extracellular adhesive domains and a cytoplasmic domain that recruits the glycolytic enzyme aldolase which in turn binds to F-actin and hence connects to myosin A [4], [5]. The forward locomotion of the zoite results from the posterior translocation of TRAP-aldolase-actin assembly. In the rodent malaria parasite, Plasmodium berghei, deletion of TRAP or mutations in its cytoplasmic domain that abrogate its interaction with aldolase result in non-motile sporozoites [6], [7].
Generation of nonmotile sporozoites linked gliding motility to host cell invasion, supporting earlier findings in Toxoplasma that apicomplexan zoites actively invade host cells [6], [8]. Zoites also require motility to reach their target cell and vary tremendously in the degree to which they are reliant on motility in this regard. Plasmodium merozoites, for example, are released in close proximity to their target cell and although they possess all of the motor components and likely use this machinery to invade cells [9], they are not capable of gliding motility in vitro. In contrast, Plasmodium sporozoites develop in oocysts on the mosquito midgut wall, far from their ultimate target, the mammalian liver. They must enter mosquito salivary glands, from where they are inoculated into the mammalian dermis, exit the dermis to enter the blood circulation and finally penetrate the sinusoidal barrier of the liver to reach the hepatocytes. In vitro they display a robust gliding phenotype that parallels their need to move longer distances compared to merozoites, ookinetes and zoites of other Apicomplexan genera.
Proteolytic cleavage of surface proteins is a central feature of invasion by apicomplexans (reviewed in [10]). Cleavage occurring in the amino-terminus exposes critical adhesive motifs [11] whereas carboxy-terminal cleavage is thought to disengage adhesive interactions between parasite ligands and host cell receptors (reviewed in [10]). Carboxy-terminal cleavage can occur either extracellularly, close to the plasma membrane, or within the transmembrane domain of the protein, with distinct classes of serine proteases being responsible in each case. In Plasmodium, removal of adhesins from merozoites has been studied in some detail. Two of the most abundant merozoite surface proteins, merozoite surface protein 1 (MSP1) and apical membrane protein 1 (AMA1) are removed by a subtilisin-like protease called SUB2 which cleaves its substrates in a juxtamembrane location [12], [13] whereas the invasion ligand EBA-175 is cleaved within its transmembrane domain [14]. Intramembraneous cleavage is accomplished by rhomboid proteases, a family of serine proteases initially described in Drosophila that require helical instability in the transmembrane domain and have specific residue requirements in their P1, P4 and P2′ positions [15], [16]. Initial studies demonstrating a role for intramembraneous cleavage of zoite adhesins were carried out in Toxoplasma where it was shown that the microneme proteins TgMIC2, TgMIC6 and TgMIC12 were shed from the zoite surface after cleavage within their transmembrane domain [17], [18], [19]. More recently, a conditional knockout of the rhomboid protease TgROM4 demonstrated that this protease plays a critical role in cleavage of TgMIC2 and TgAMA1 with downstream effects on motility, invasion and parasite replication [20], [21]. Importantly, a conserved rhomboid substrate motif is found in all TRAP family members [17] .
To date, the role of cleavage and shedding of surface adhesins in Plasmodium sporozoites has not been addressed. However, a previous study has shown that TRAP is cleaved and shed into the supernatant after incubation of sporozoites at 37°C [22]. Moreover, when expressed in heterologous systems, TRAP can be cleaved by a rhomboid protease [18], [23]. Considering that an ortholog of TgROM4 is found in all malaria parasite genomes and TRAP contains a canonical rhomboid cleavage site in its transmembrane domain, it is plausible that TRAP is shed from the sporozoite surface by the action of a rhomboid. In this study, we generated a series of TRAP mutants in the rodent malaria parasite, P.berghei, and performed functional and biochemical assays to elucidate the importance of TRAP cleavage and to characterize the nature of the protease responsible for this event.
To analyze TRAP processing, we performed pulse-chase metabolic labeling experiments with sporozoites, and immunoprecipitated TRAP from the sporozoite pellet and supernatant using antibodies specific for the repeat region (α-Rep, Fig. 1A). An ∼83 kD species was associated with the sporozoite pellet whereas TRAP processing led to the release of a ∼76 kD species into the supernatant (Fig. 1B). The recognition of the 76 kD product released into the supernatant by anti-repeat antibodies suggests that the extracellular domain of TRAP is shed, in agreement with previous findings [22]. When antisera recognizing the cytoplasmic tail of TRAP (α-CT, Fig. 1A) was used to immunoprecipitate TRAP, full-length TRAP associated with the sporozoite pellet could be detected but the cleaved portion was not detected in the supernatant, indicating that TRAP is shed without its C-terminal domain (Fig. 1B). These data suggest that TRAP is cleaved either within its transmembrane domain or in the juxtamembrane region.
To determine the nature of the protease responsible for TRAP cleavage, we examined the effect of a variety of protease inhibitors on this process. As shown in Figure 1C, TRAP cleavage was inhibited by a subset of serine proteases inhibitors, namely TLCK, PMSF, and DCI but was not affected by EDTA, cysteine and aspartyl protease inhibitors, or the serine protease inhibitors leupeptin and aprotinin. Overall these data suggest that TRAP is cleaved by a calcium-independent serine protease.
Since TRAP was previously shown to play a critical role in gliding motility [6], we determined if these protease inhibitors also impact on motility. When sporozoites were preincubated with protease inhibitors and then kept in their presence during a gliding motility assay, we found that TLCK and PMSF blocked gliding. DCI, which has a short half-life, had a moderate effect on motility, however, when it was replenished during the assay the inhibition was stronger (Fig. 1D). Protease inhibitors that had no effect on TRAP processing, namely, pepstatin, leupeptin, and E-64, also had no inhibitory effect on motility. Since the same subset of serine protease inhibitors had inhibitory effects on both TRAP cleavage and gliding motility, our data suggest that the removal of TRAP from the sporozoite surface is required for gliding motility.
In order to elucidate the function of TRAP cleavage and to better define the nature of the responsible protease, we generated sporozoites expressing mutated forms of TRAP in which point mutations were introduced in the transmembrane domain to disrupt the putative rhomboid substrate motif. We created two rhomboid cleavage site mutants based on previously published studies in Toxoplasma and Plasmodium in which these mutations led to aberrant or impaired release of the rhomboid protease substrate from the zoite surface [14], [18], [19]: One in which the canonical rhomboid motif AGGIIGG was changed to VALIIGV (TRAP-VAL; Fig. 2A) and another in which it was changed to FFFIIGG (TRAP-FFF; Fig. 2A). Targeting plasmids were designed to replace the endogenous locus via double-cross-over homologous recombination (Fig. S1A). A recombinant control line (TRAP-rWT) was generated using a plasmid containing a wild type copy of the TRAP open reading frame. After transfection and cloning, a series of diagnostic PCRs and sequencing was used to verify integration into the correct genomic locus and the presence of the desired mutations (Fig. S1B). TRAP-rWT parasites were similar to wild type P. berghei ANKA parasites and were used throughout this study for comparison to mutant lines. Western blot analysis of the rhomboid cleavage site mutants demonstrated that the parasites express normal amounts of TRAP compared to controls (Fig. 2B).
We performed pulse-chase metabolic labeling experiments to assess TRAP processing in the rhomboid cleavage site mutants and found that cleavage of TRAP was severely impaired in the two mutant lines (Fig. 2C). We then examined the cellular localization of TRAP in the mutant sporozoites by immunofluorescence microscopy. Total overall TRAP staining in permeabilized sporozoites was similar in mutants and controls (data not shown), consistent with the Western blot results. However, when only surface TRAP was stained, a striking difference was observed between mutant and control sporozoites. In wild type sporozoites, there is typically only a small amount of TRAP found on the sporozoite surface, and the staining pattern can be described as a “faint dusting” [24]. In contrast, the majority of rhomboid-cleavage site mutants displayed larger amounts of TRAP on their surface with bright staining along most of their surface (Figs. 2D & 2E), indicative of an absence of shedding. Quantitative measurement of fluorescence intensity of these stained sporozoites was revealed a statistically significant difference between the rhomboid cleavage site mutants and TRAP-rWT sporozoites (p<.0001).
Given the impaired TRAP processing and shedding in the rhomboid cleavage site mutants, we set out to analyze the phenotype of these mutants. Since TRAP is not expressed in erythrocytic stages and a previous study in which the TRAP gene had been deleted did not show an altered phenotype in these stages [6], we did not expect altered growth of asexual erythrocytic stages or gametocyte production in these mutants and this was indeed the case (data not shown). To study the mosquito stages, Anopheles stephensi mosquitoes were allowed to feed on infected mice and sporozoites were isolated from mosquito midguts, hemolymph, and salivary glands for analysis. Sporozoites develop in oocysts on the mosquito midgut wall and reach maximum numbers on day 14 post blood-meal. They are then released into the hemolymph from where they specifically bind to and invade salivary glands, where they reach maximal numbers on day 18 post blood-meal. As shown in Table 1, numbers of oocyst and hemolymph sporozoites were comparable in mosquitoes infected with TRAP-rWT, TRAP-VAL or TRAP-FFF parasites, indicating that the introduced mutations had no effect on sporozoite development in the oocyst or their release into the hemolymph. However, when salivary gland sporozoite populations were examined, both TRAP-VAL and TRAP-FFF clones had five times fewer salivary gland sporozoites compared to the TRAP-rWT clones, suggesting a defect in invasion (Table 1).
To determine the role of TRAP cleavage and shedding on motility, we performed gliding motility assays with the rhomboid cleavage site mutant sporozoites. Initial experiments assayed motility by staining and counting trails left by gliding sporozoites. Staining trails for CSP or TRAP showed that mutants were capable of gliding and leaving trails in their wake (data not shown and Fig. S3). When these trails were counted, we found only a small decrease in the percentage of mutants that were non-motile (Fig. 3A, pie charts). However, when comparing the number of trails produced by each parasite line, rhomboid-cleavage site mutants produced fewer trails: Whereas over 40% of control sporozoites produced 31–50 circles and 25% produced greater than 50 circles, only 10% of mutant sporozoites produced 31–50 circles and none produced more than 50 circles (Fig. 3A bar graph).
To further analyze gliding motility of mutant sporozoites we performed live imaging studies. Like control sporozoites, the rhomboid cleavage site mutants moved in circles, however, they moved at a slower rate and frequently appeared stuck, attempting to move forward but unable to (Videos S1, S2 and S3). The mutants also displayed patterns of non-productive motility, such as bending, flexing, waving, and pendulum-like movements, which consists of moving one-third of a circle and then returning to the starting position (Fig. 3B, Videos S1, S2 and S3). When we calculated their speed, control sporozoites glided with an average speed of ∼2 µm/s, whereas mutant sporozoites glided at a rate of ∼0.5 µm/s (Fig. 3C). Overall these data indicate that the rhomboid cleavage site mutant sporozoites are impaired in motility, traveling shorter distances and at a reduced speed.
Since gliding motility is required for host cell invasion [6], [8], we examined the infectivity of the rhomboid cleavage site mutants in a number of in vitro assays. First we determined the invasion rate by counting the number of intracellular and extracellular sporozoites after their incubation with the hepatocyte cell line, Hepa1-6. The mutant sporozoites displayed a marked decrease in invasion with 20% of mutant sporozoites compared to 60% of control sporozoites being found intracellularly (Fig. 4A, left axis). One limitation of this assay, however, is that it fails to distinguish between sporozoites that have productively invaded the cells, i.e. with the formation of a parasitophorous vacuole (PV) versus sporozoites that are only migrating through, a process that is distinct from productive invasion and results in wounding of the traversed cell [25]. To address this issue, we performed invasion assays and stained intracellular sporozoites with antisera to UIS4, a protein that is localized to the PV membrane [26]. Only a small proportion of the intracellular TRAP-VAL and TRAP-FFF sporozoites stain with UIS4 (Fig. 4A, right axis), indicating that the majority of the mutants are not able to productively invade hepatocytes. This was confirmed when we tested the ability of the rhomboid cleavage site mutants to develop into exoerythrocytic stages (EEFs) in vitro. There was an approximately 10-fold reduction in the number of EEFs produced by the mutant sporozoites compared to controls, correlating with the small percentage of mutant sporozoites that productively invade hepatocytes (Figs. 4 A&C). However, the few EEFs that are formed are similar in size to control EEFs (data not shown), indicating that TRAP cleavage does not play a role in EEF development.
Since the gliding motility studies demonstrated that the rhomboid cleavage site mutants moved more slowly, we also examined the kinetics with which these mutants invade hepatocytes. The number of sporozoites in the process of entering hepatocytes after 15, 30, and 45 minutes was quantified by counting sporozoites that were half in and half out. The highest percentage of control TRAP-rWT sporozoites in the process of entering was seen at 15 minutes after their addition to cells and this number decreased over time. Conversely, the percentage of TRAP-VAL and TRAP-FFF sporozoites entering hepatocytes was lowest at 15 minutes and increased slightly at later time points (Fig. 4B). As stated above, this inside/outside assay cannot distinguish sporozoites that are productively invading versus those that are migrating through. Nonetheless, these data indicate that cell entry for either of these processes, is slower for the rhomboid cleavage site mutants, a finding that is consistent with their gliding phenotype.
Infectivity of the rhomboid-cleavage site mutants was evaluated in mice after both intravenous (i.v.) and intradermal (i.d.) inoculation. We first determined the time to detection of blood stage parasites (prepatent period) after i.v. inoculation into Swiss Webster mice and found that the rhomboid cleavage site mutants had a delay of one to two days in the prepatent period compared to controls (Table 2). Since each day delay is correlated with a 10-fold decrease in infectivity [27], our results indicate that the mutant sporozoites were ∼10 to 100-fold less infective than control sporozoites in these mice. We also examined infectivity of these mutants in C57BL/6 mice, which are highly susceptible to P. berghei infection, and found a one day delay in the prepatent period compared to controls (Table 2). When we tested the infectivity of the rhomboid cleavage site mutants after i.d. inoculation, the decrease in infectivity was more pronounced. In all of the Swiss Webster mice and the majority of the C57BL/6 mice, blood stage parasites were not observed after i.d. inoculation of the mutant sporozoites, with monitoring up to 21 days post injection (Table 2). The one C57BL/6 mouse in each group that did become positive for blood stage parasites had a significant delay in patency. Since sporozoites are inoculated by mosquitoes into the dermis of the mammalian host [28], [29] and migration through cells and tissues is critical for sporozoite exit from the dermis [30], [31], [32], this dramatic decrease in infectivity after i.d. inoculation suggests that robust/fast gliding motility is particularly important for dermal exit. To further evaluate the ability of the rhomboid cleavage site mutants to migrate through cells, we performed an in vitro assay in which sporozoites are added to cells in the presence of a fluorescent cell-impermeant dye. Wounded cells take up the dye and can then be counted [25]. As shown in Figure 4D, mutant sporozoites exhibited a reduction in cell traversal activity compared to control sporozoites.
Although the point mutations introduced within the putative rhomboid cleavage site of TRAP resulted in impaired cleavage, parasite motility and invasion, these activities were not completely abolished. It is conceivable that a small amount of TRAP is still cleaved, at an alternate site, allowing for the slow, halting movement we observed. Indeed, overexposure of the film from the pulse-chase experiment shown in Figure 2 indicated residual processing of TRAP (Fig. 2C, black arrow). Although we cannot isolate sufficient amounts of protein to map either the canonical or the alternate cleavage sites, the size of the shed form of TRAP-VAL and TRAP-FFF was not significantly smaller than the shed form of TRAP-rWT, suggesting that mutant TRAP was being cleaved close by, possibly in a juxtamembrane location.
In Plasmodium merozoites, subtilisin-like protease PfSUB2 has been implicated in shedding of adhesins at a juxtamembrane position [13]. The known substrates for Plasmodium subtilisins harbor no obvious sequence similarity. Instead, it has been proposed that subtilisin-substrate recognition involves stretches of conformationally unrestrained peptides around the target peptide bond such that the degree of “disorder” rather than a primary amino acid sequence serves as a determinant for cleavage [12]. The juxtamembrane region of P. berghei TRAP, which starts downstream of the proline based repeats, consists of approximately 160 amino acids (Fig. S2) and is predicted to be disordered and with low complexity. To test whether this region can serve as a substrate for cleavage, we generated two additional TRAP mutants. TRAP-DMut corresponds to a double mutant (Fig. 5A and Fig. S2) in which 140 amino acids of the juxtamembrane region were deleted and the rhomboid substrate motif, AGGIIGG was changed to FFFIIGG. While no function has been ascribed to the juxtamembrane region of TRAP, it is possible that deletion of this large portion of the protein may affect protein stability, expression and function. Therefore, we also generated a control mutant line, TRAP-JMD, in which the same 140 amino acids of the juxtamembrane domain were deleted but the rhomboid substrate motif was maintained (Fig. 5A and Fig. S2). Any defects associated solely with the deletion of the juxtamembrane region could be analyzed with this mutant. Similar to the generation of the rhomboid cleavage site mutants, targeting plasmids containing TRAP-JMD and TRAP-DMut genes were constructed to replace the endogenous TRAP locus and a series of diagnostic PCRs and sequencing was used to verify integration into the TRAP locus and the presence of the desired mutations in the cloned transfectants (Fig. S1 and data not shown).
Both mutants expressed normal amounts of TRAP by Western blot, however, its molecular weight was significantly lower due to deletion of the juxtamembrane region (Fig. 5B). TRAP localization in permeabilized TRAP-DMut and TRAP-JMD sporozoites was analyzed by immunofluorescence and was found to be similar to previous controls (data not shown). In contrast, surface staining of TRAP gave strikingly different results for the two mutants. Whereas TRAP-JMD sporozoites were similar to controls, the majority of TRAP-DMut sporozoites expressed large amounts of TRAP on their surface similar to the rhomboid cleavage site mutants (Figs. 5C & 5D). This result suggests that removal of TRAP-DMut from the sporozoite surface is considerably inhibited. Unfortunately, a more quantitative assessment of TRAP cleavage by pulse-chase metabolic labeling was technically not possible due to the limited number of TRAP-DMut salivary gland sporozoites (see below).
Sporozoite development in mosquitoes of TRAP-JMD and TRAP-DMut parasites was similar to wild type parasites, with normal numbers of midgut and hemolymph sporozoites (Table 1). Furthermore, TRAP-JMD mutants had normal numbers of salivary gland sporozoites, indicating that infectivity in the vector was not altered by deletion of the juxtamembrane region of TRAP. In contrast, there were very low numbers of salivary gland TRAP-DMut sporozoites, approximately 30-fold lower than controls and 4-fold lower than TRAP-VAL and TRAP-FFF mutants (Table 1). To examine the phenotype more in depth, we determined the percentage of salivary gland sporozoites that were inside the glands for each mutant. Salivary glands were incubated with trypsin after which sporozoites remaining with the glands and those released into the supernatant were counted. As shown in Table 1, TRAP-JMD parasites and the rhomboid cleavage site mutants were not significantly different from controls, with approximately 77% of salivary gland associated sporozoites found inside the glands. In contrast, there was a dramatic decrease in the proportion of TRAP-DMut sporozoites inside the salivary glands with only 15% found inside (Table 1). Taken together, these data indicate that TRAP-DMut sporozoites are severely impaired in their ability to invade salivary glands.
We performed a number of functional assays to determine the phenotype of TRAP-DMut and TRAP-JMD sporozoites in the mammalian host. In gliding motility assays the majority of TRAP-JMD sporozoites were motile and produced greater than 10 circles, similar to controls (Fig. 6 A&B; controls shown in Fig. 3A). In contrast, the majority of TRAP-DMut sporozoites were completely non-motile and the small number that moved produced only 1 to 2 circles (Fig. 6 A&B).
In vitro infectivity assays revealed that TRAP-JMD sporozoites were able to invade hepatocytes with rates similar to those observed with TRAP-rWT sporozoites whereas TRAP-DMut sporozoites did not invade hepatocytes at all (Fig. 6C). When we examined the infectivity of these mutants in mice after i.v. inoculation, we found that TRAP-JMD mutants had similar infectivity to controls in both Swiss Webster and C57Bl/6 mice whereas i.v. inoculation of high doses of TRAP-DMut sporozoites never resulted in a blood stage infection in either mouse strain (Table 2). This complete lack of infectivity of TRAP-DMut sporozoites correlates the absence of gliding and lack of shedding of TRAP from the zoite surface.
The phenotype of the rhomboid cleavage site mutants suggests that a rhomboid protease is primarily responsible for removal of TRAP from the sporozoite surface. The identity of this rhomboid protease remains unknown, however, the transcriptome database shows that two rhomboid proteases are highly expressed in sporozoite stages, namely ROM1 and ROM4 [33]. Previous investigators have demonstrated that ROM1 does not cleave TRAP [34] and that ROM4 cannot be conventionally deleted because it is vital in the blood stages [14]. We therefore made antisera specific for ROM4 to determine whether the protein was expressed in the sporozoite stage. Polyclonal antisera to the exposed C-terminal portion of P. falciparum ROM4 cross-reacted with P. berghei ROM4 by Western blot, recognizing a ∼69 kD species which corresponds to the predicted size of PbROM4 (Fig. 7A). The specificity of this antiserum was confirmed using PbROM4 conditional knockout parasites (Fig. S4). Furthermore, immunofluorescence experiments showed that PbROM4 was abundantly expressed in sporozoites and co-localized with CSP found on the sporozoite surface (Fig. 7B). These data demonstrate that ROM4 is expressed during the sporozoite stage and future experiments should determine whether it is responsible for TRAP cleavage and shedding from the sporozoite surface.
TRAP provides a link between the extracellular surface and the motor of the parasite, thus allowing the force generated by the motor to be transmitted to the sporozoite surface. This motile force results in posterior translocation of TRAP and the current model predicts that shedding of TRAP would be essential to disengage interactions between the extracellular matrix or host cell receptors and the motor assembly [2], [3]. Here we demonstrate that removal of TRAP from the sporozoite surface is critical for sporozoite motility which in turn is crucial for exit from the inoculation site in the mammalian host and cell invasion in both mosquito and mammalian hosts. Our data suggest that removal of TRAP is accomplished by a rhomboid protease and that when canonical rhomboid activity is prevented, either another protease can modestly compensate for its activity or the rhomboid protease can cleave, albeit less efficiently, at a juxtamembranous site
The TRAP cleavage site mutants that we generated exhibit two distinct phenotypes: The rhomboid cleavage site mutants, TRAP-VAL and TRAP-FFF, displayed an intermediate gliding/infectivity phenotype and TRAP-DMut exhibited an almost complete abrogation of motility and infectivity. Our data suggest that the intermediate gliding phenotype of the rhomboid cleavage site mutants is due to a defect in TRAP shedding from the sporozoite surface. Indeed, both pulse-chase metabolic labeling experiments demonstrate a dramatic decrease in TRAP cleavage and immunofluorescence studies indicate an accumulation of TRAP on the surface of the mutant sporozoites. Furthermore, live imaging studies showed that rhomboid cleavage site mutants could move the length of a sporozoite but then appeared stuck at their posterior end, eventually disengaging and moving forward another sporozoite length only to become stuck again. This ‘constipated’ movement paralleled the inefficient removal of TRAP from their surface. Nonetheless, both the partial phenotype of these mutants and the small amount of cleaved TRAP observed in pulse-chase metabolic labeling experiments suggested that TRAP could be removed from these parasites, albeit slowly and inefficiently. We hypothesized that inefficient TRAP cleavage was being performed by another protease and focused on the subtilisin-like proteases because of their role as sheddases in the erythrocytic stage of the parasite [13]. The genomes of rodent and human Plasmodium parasites contain three subtilisin-like protease genes, sub1, sub2 and sub3, all of which are transcribed in sporozoites [33]. Since SUB2 removes adhesins from the Plasmodium merozoite surface by juxtamembranous cleavage [13], we indirectly tested whether shedding of mutant TRAP was being performed by a subtilisin-like protease by generating a mutant in which both the juxtamembrane domain was deleted and the rhomboid cleavage site was altered (TRAP-DMut). The low numbers of salivary gland sporozoites produced hampered a direct quantitative assessment of the mutations on TRAP cleavage. Nonetheless, immunofluorescence assays indicated that TRAP-DMut sporozoites accumulated TRAP on their surface and functional assays demonstrated that this mutant was non-motile and not infectious in the mammalian host. Since TRAP-JMD sporozoites, in which the putative subtilisin cleavage site was deleted but the rhomboid cleavage site was left intact, had a phenotype similar to wild type sporozoites, the phenotype of TRAP-DMut sporozoites cannot be attributed to the large deletion in the juxtamembrane region. Overall, these data support the hypothesis that disengagement of adhesive interactions through removal of TRAP from the zoite surface is essential for gliding motility. These findings logically complement a recent study in which sporozoite gliding motility was studied by reflection interference contrast microscopy and found to consist of a series of adhesion de-adhesion events in which TRAP plays a critical role [35]. Taken together, our study and the Munter study [35] highlight the delicate balance between adhesion and de-adhesion that must be achieved for fast and effective gliding motility.
The impairment of TRAP removal from the sporozoite surface and its effect on gliding motility had significant downstream effects on target cell invasion in both the mosquito and mammalian hosts. Furthermore, the degree to which motility was affected correlated with the impairment in infectivity. Less expected and equally important, the intermediate gliding phenotype of the rhomboid cleavage site mutants highlights the critical role of fast (and robust) gliding motility for exit from the dermal inoculation site. In comparison to zoites of other Plasmodium life cycle stages and other Apicomplexan genera, sporozoites move longer distances and display a more pronounced gliding phenotype that fuels their need to get from the dermis, where they are deposited by an infected mosquito, to the liver. The rhomboid cleavage-site mutants whose motility was slower and appeared to get stuck due to their inability to disengage adhesive interactions, were significantly less infectious after i.d. inoculation compared to i.v. inoculation, demonstrating the importance of robust gliding for tissue migration.
Both rhomboid cleavage site and TRAP-DMut sporozoites accumulated large amounts of TRAP on their surface in contrast to wild type sporozoites which shed most of their surface TRAP. This is similar to the accumulation of TgMIC2 on the Toxoplasma tachyzoite surface following depletion of TgROM4 [20]. It has been postulated that during motility, adhesins are translocated along the zoite surface and removed at the posterior end of the zoite. However if TRAP were only removed from the posterior end of the sporozoite, we would expect to see accumulation at this location. In contrast, we observed large amounts of TRAP all over the zoite surface. In addition, PbROM4, the protease that may be responsible for TRAP shedding, decorates the entire sporozoite surface and is not concentrated at the posterior pole. Our hypothesis is that TRAP removal occurs in a stochastic manner as it is translocated posteriorly and that this pattern of removal results in the smooth gliding pattern observed with wild type sporozoites. In contrast, the rhomboid cleavage site mutants have a slow ‘constipated’ movement and appear to be stuck at their posterior end. These data suggest that the alternate protease responsible for TRAP removal in these mutants may be located posteriorly, resulting in a back-log of adhesin removal and in the observed gliding phenotype. It is not yet known how rhomboid proteases are regulated, however, in the case of Drosophila ROM1, there is no evidence of regulation and the rhomboid is active wherever it is present [36]. Thus it is possible that rhomboids are regulated by their spatial localization and perhaps the localization of PbROM4 along the sporozoite surface allows it to act in a constitutive manner much like Drosophila ROM1. This is supported by studies with Toxoplasma which found that TgROM4 is distributed along the entire parasite surface and disruption of this gene leads to decreased MIC2 processing and accumulation of MIC2 along the tachyzoite surface [17], [19], [20]. Thus in both Plasmodium and Toxoplasma, ROM4 may function to remove adhesins as they translocate to the rear of the parasite, enabling smooth forward gliding.
Our data suggest that ROM4 may be the protease that is responsible for TRAP shedding from the sporozoite surface. It is expressed in sporozoites and its localization along the sporozoite surface correlates with the accumulation of TRAP-VAL and TRAP-FFF along the entire surface of the rhomboid cleavage site mutants. There are eight rhomboid-like genes in the P. falciparum genome and they are numbered to reflect their homology with their counterparts in T. gondii. Of these, only PfROMs 1, 3, 4 and 6 are clear orthologs of the Toxoplasma ROMs [17]. PfROM1 and PfROM4 are highly expressed in sporozoites [33], making them the most likely candidates for TRAP cleavage. A previous study examining the enzymatic properties of the Plasmodium rhomboids in a heterologous mammalian cell system found that co-expression of TRAP with PfROM4 resulted in its release from the cell surface whereas co-expression of TRAP with PfROM1 did not [23]. Recently, ROM1 was disrupted in P. berghei and TRAP processing was unaffected in these sporozoites [34]. These data, together with the data from our study suggest that PbROM4 is the primary candidate for TRAP removal from the sporozoite surface. Nonetheless, definitive proof that PbROM4 is the TRAP sheddase awaits the establishment of experimental conditions to conditionally deplete PbROM4 in the mosquito stages since the gene cannot be disrupted in intra-erythrocytic stages [14]. Overall these findings reinforce the notion that targeting serine proteases such as the rhomboids and subtilisins, may constitute novel chemotherapeutic targets for malaria. While it has been demonstrated that rhomboid proteases play a role in the erythocytic stage of the parasite life cycle [14], the findings presented here implicate a role for rhomboid proteases, and potentially subtilisins during the pre-erythrocytic stages of malaria. While the erythrocytic stages of the parasite life cycle are responsible for the clinical manifestations of the disease, the sporozoite stage is responsible for establishing infection in the mammalian host. Hence, targeting the rhomboids would not only provide a mechanism for treating malaria, but also for prevention of disease by inhibiting sporozoite infectivity. Although rhomboids are found throughout all kingdoms of life, there is a significant level of diversity among the different classes of rhomboids with potential differences in structure, function and substrate recognition. Perhaps these differences can be exploited to generate specific rhomboid inhibitors that can lead to a new class of anti-malarial drugs.
All animal work was conducted in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the NYU School of Medicine Institutional Animal Care and Use Committee (protocol #080413) and by the Johns Hopkins University Animal Care and Use Committee (protocol #M011H467), which is fully accredited by the Association for Assessment and Accreditation of Laboratory Animal Care. All efforts were made to minimize suffering.
Primer sequences used in the generation and verification of mutants are listed in Table S1. Transfection plasmids designed to replace the endogenous locus with mutant or wild type TRAP were generated using pDEF-hDHFR (www.malaria.mr4.org) containing the human dihydrofolate reductase (hDHFR) selection cassette (Fig. S1). A 576 bp fragment of 5′UTR located 1000 bp upstream of TRAP was amplified by PCR using template genomic DNA from P.berghei ANKA parasites, with the primer pair, PbTRAP5′UTR-FWD and PbTRAP5′UTR-REV and this was cloned into pDEF-hDHFR upstream of the selection cassette. A second fragment, consisting of 1062 bp of 5′UTR directly upstream of the TRAP gene, 1821 bp of TRAP open reading frame, and 2057 bp of TRAP 3′UTR was amplified by PCR using the primer pairs PbTRAPFWD and PbTRAPREV and Pfu polymerase (Stratagene). This was cloned into pDEF-hDHFR downstream of the selection cassette to generate the plasmid pTRAP, which was then used to generate the TRAP mutants using the Quick Change Mutagenesis kit (Stratagene). Generation of each of the rhomboid cleavage site mutants required 2 steps. Primers MUT-TRAP-VAL1FWD and MUT-TRAP VAL1REV were used to generate pTRAP-VALIIGG and then this plasmid was mutated to VALIIGV using primers, MUT-TRAP-VALGV-2FWD and MUT-TRAP-VALGV-2REV. For the TRAP-FFF mutant, first AGGIIGG was mutated to AFFIIGG using primers MUT-TRAP-FF1 FWD and MUT-TRAP-FF1 REV to generate pTRAP-AFF and this was used to generate pTRAP-FFFIIGG using primers MUT-TRAP-FFF2 FWD and MUT-TRAP-FFF2 REV. Mutagenesis of pTRAP to delete 420 bp of the juxtamembrane region of TRAP was performed using primer pairs Juxtamem-MUT1FWD and Juxtamem-MUT2REV to generate pTRAP-JMD. To generate the double mutant containing the juxtamembrane deletion and an altered rhomboid cleavage site, pTRAP-FFF was used to delete 420 bp of the juxtmembrane of TRAP using the same primer pairs. All constructs were sequenced to confirm the presence of the desired mutations.
P. berghei ANKA GFP 507clone 1 parasites, which express GFP under the control of the ef1a promoter were used for transfection [37]. Each targeting plasmid was digested with EcoRV and XhoI to liberate the fragment and transfections were performed as previously outlined [37] using 10 µg of digested plasmid DNA and the Amaxa Nucleofector (program U33). Transfected parasites were injected i.v. into Swiss Webster mice and drug resistant parasites were selected using pyrmethamine in the drinking water. Once a parental population was obtained, cloning by limiting dilution was performed in mice [37].
For each clone, integration of the DNA fragment used for transfection at the correct location was confirmed by PCR using 300 ng of genomic DNA isolated from recombinant parasites. To confirm integration at the 5′ end, primers TX-1TRAP5′INT-FWD and 5UTRhDHFRseqREV were used; at the 3′ end, primers hDHFR-3UTRseq and TX-2TRAP3′INT-REV were used; and to verify that there was no contamination with wild type parasites, primers TX-1 TRAP5′INT-FWD and 5′UTRPbTRAP-REV were used. To amplify the TRAP open reading frame, primers SEQPbTRAP2-FWD and SEQPbTRAP3-REV were used and the resulting PCR product was sequenced to confirm the presence of the desired point mutations and/or deletions.
Monoclonal antibody (mAb) 3D11 directed against the repeat region of P.berghei CSP [38] and mAb 2E6 directed against P.berghei Hsp70 [39] were used to stain sporozoites and EEFs respectively. UIS-4 polyclonal antiserum, specific for the hepatic stage parasitophorous vacuole [26] and antiserum specific for the cytoplasmic domain of P. berghei TRAP [7] were gifts from Dr. Stefan Kappe and Dr. Ali Sultan, respectively. Antiserum to the repeat region of P.berghei TRAP was generated in rabbits using the repeat peptide AEPAEPAEPAEPAEPAEPCNH2 synthesized and purified by Anaspec Incorporated. The peptide was conjugated to keyhole limpet hemacyanin and the rabbit was immunized and boosted as previously outlined [40]. For generation of polyclonal antisera against the C-terminus of PfROM4, the last 49 amino acids of the protein was amplified from a plasmid containing a synthetic PfROM4 gene (pHAROM4synth [14]) using forward primer 5′-GGATCCTATAGCC CCCTCGGCCAGATCAAG-3′ and reverse primer 5′-CTCGAGCTTGTTGCAGTAA TACCGAGTGGCTTC-3′, and cloned into pGex4T1 vector (Amersham Bioscience) for protein expression. Protein was purified using the QIAGEN Ni-NTA superflow resin under denaturing conditions according the manufacturer's instructions. Antiserum was raised in rabbits by Eurogentec S.A. according to their standard protocol. IgG fraction of PfROM4 C-terminal antiserum was purified using a Protein G agarose column according the manufacturer's instructions (Pierce).
Anopheles stephensi mosquitoes were fed on mice infected with recombinant parasite lines once abundant gametocyte stage parasites were observed. Days 14, 16 and 18 post-blood meal, mosquito midguts, hemolymph and salivary glands were harvested for determination of sporozoite numbers, respectively. Although there is some variation among different institutions as to the optimal time to harvest sporozoites, at our facility these are the times when sporozoite numbers in each respective compartment are at their maximum. For midgut and salivary gland sporozoite counts, organs from 10 to 15 mosquitoes were pooled and homogenized and released sporozoites were counted using a hemocytometer. For hemolymph sporozoite counts, the hemocoel was perfused with DMEM and the first two drops of perfusate collected from 10 to 15 mosquitoes was pooled and sporozoites were counted as above. To determine the proportion of salivary gland sporozoites that were inside the glands, day 18 salivary glands were dissected, incubated with 50 µg/ml trypsin for 15 min at 37°C and centrifuged at 100× g for 5 min at 4°C. The supernatant was collected and the salivary glands were homogenized to release internalized sporozoites and the sporozoites in each compartment were counted using a hemocytometer.
Experiments with wild type sporozoites were performed with 106 sporozoites per condition whereas experiments with recombinant sporozoites (TRAP-rWT, TRAP-VAL and TRAP-FFF) were performed with 2.5×105 sporozoites. Sporozoites were metabolically labeled in DMEM without Cys/Met, containing 0.2% BSA, and 400 µCi/ml [35S]-Cys,Met for 1 hour at 28°C and then kept on ice or chased in DMEM with Cys/Met and 0.2% BSA for 2 hours at 28°C in the absence or presence of the indicated protease inhibitors. The concentrations of inhibitors used were: 100 µM TLCK, 75 µM leupeptin, 0.3 µM aprotinin, 100 µM 3,4 DCI, 5 mM EDTA, 10 µM E64, 1 mM PMSF, and 1 µM pepstatin. Metabolically labeled and chased sporozoites were spun at 16,000× g for 4 min, and the pellets and supernatants were separated. The sporozoite pellets were lysed in SDS/Urea lysis buffer [1% SDS, 4 M Urea, 150 mM NaCl, 50 mM Tris-HCl pH 8.0, 1X Protease Inhibitor Cocktail, (Roche)] for 1 hr at 4°C and TRAP was immunoprecipitated with TRAP repeat antiserum or TRAP C-terminal antiserum conjugated to Protein A agarose beads overnight at 4°C. The beads were then washed with lysis buffer (1% Triton X-100, 150 mM NaCl, 50 mM Tris-HCl, pH 8.0) followed by lysis buffer containing 500 mM NaCl and pre-elution buffer (0.5% Triton X-100, 10 mM Tris-HCl, pH 6.8). TRAP was eluted with 1% SDS in 0.1 M glycine, pH 1.8, neutralized with 1.5 M Tris-HCl, pH 8.8, and run on a 7.5% SDS–polyacrylamide gel under non-reducing conditions using 18×16 cm gels (Hoefer SE600 system). Gels were fixed, enhanced with Amplify (GE Biosciences), dried and exposed to film.
Experiments with TRAP-FFF and TRAP-VAL mutants used salivary gland sporozoites whereas those with TRAP-JMD and TRAP-DMut mutants utilized midgut sporozoites due to the low numbers of salivary gland sporozoites in the TRAP-DMut parasite. Sporozoites were lysed in 6X SDS-PAGE sample buffer and 3×104 sporozoite equivalents were loaded per lane of a 18×16 cm, 7.5% SDS-polyacrylamide gel and separated under non-reducing conditions. The proteins were then transferred to a PVDF membrane, and incubated with either TRAP repeat antiserum (1∶100) or mAb 3D11 (4 µg/ml), followed by either anti–rabbit (1∶50,000) or anti–mouse Ig (1∶200,000) conjugated to HRP. Bound antibodies were visualized using the enhanced chemiluminescence detection system (GE Biosciences). Western blot analysis for PbROM4 expression was performed by loading lysates of Nycodens purified P. berghei asexual blood stage schizonts and dissected P. berghei wildtype sporozoites on an SDS-polyacrylamide gel, proteins were transferred as outlined above and the blot was incubated with anti-ROM4 antisera and developed as above. PbROM4 conditional knockouts were used to demonstrate the specificity of this antisera. These parasites were generated by a double crossover strategy positioning the transactivator under control of the endogenous P. berghei ROM4 promoter, while PbROM4 expression was controlled by the inducible tet-operator containing promoter, resulting in an inducible copy of PbROM4. Several independent transgenic parasite pools were obtained, cloned and verified by PCR (P. Pino and D. Soldati-Favre, unpublished data). Asexual stage transgenic parasites were inoculated into mice which were treated or not with anhydrotetracycline (ATc) in the drinking water for 36 hrs, parasites were collected and allowed to mature to schizonts in vitro for 12 hrs in the presence or absence of ATc, schizonts were purified and lysates were evaluated by western blot as outlined above.
Wild type or mutant salivary gland sporozoites were centrifuged onto glass 8-chambered Lab-Tek wells at 300× g for 2 min at 12°C, and then fixed with 4% PFA for 1 hr at RT. For total TRAP staining, sporozoites were also permeabilized with cold methanol for 15 min at −20°C following fixation with PFA. Sporozoites were stained with TRAP repeat antiserum (1∶100) diluted in 1% BSA and 5% goat serum in PBS for 1 hr at 37°C. Following this, wells were incubated with anti-rabbit IgG Alexa Fluor 594 for 1 hr at 37°C and coverslips were then mounted in Citifluor (Ted Pella). Sporozoites were visualized by phase and fluorescence microscopy with a Nikon 100X PlanApo objective on a Nikon E600. Images for quantitative analysis were acquired using a DS-Ri1 digital camera with identical resolution, gain and color settings in which parasites were exposed for 75 ms in the fluorescence channel. Intensity measurements were calculated using the NIS Elements Br 3.2 software (Nikon) and statistical analysis was performed using the Student's unpaired t-test.
Glass 8-chambered Lab-tek wells were coated with 10 µg/ml mAb 3D11 in TBS overnight at 25°C. Salivary gland sporozoites were dissected in 3% BSA/DMEM and 2×104 sporozoites were added to each well and incubated for 1 hr at 37°C. Wells were then fixed in 4% PFA and stained with biotinylated 3D11 for 1 hr at 25°C, followed by incubation with Streptavidin-FITC (1∶100, GE Biosciences) for 1 hr at 25°C. Slides were mounted using Citifluor mounting medium and visualized as above. The number of sporozoites with and without trails was counted. Assays in which TRAP was visualized in the trails were performed in the same way except that Lab-tek wells were not precoated with antibody and trails were visualized using TRAP repeat antiserum followed by anti-rabbit IgG conjugated to FITC. For experiments with protease inhibitors, the sporozoites were pre-incubated with indicated inhibitor for 1 hr at 28°C, and then added to the wells in the continued presence of the inhibitor. The concentrations of inhibitors used were: 100 µM TLCK, 75 µM leupeptin, 100 µM 3,4 DCI, 10 µM E64, 1 mM PMSF, and 1 µM pepstatin.
104 salivary gland sporozoites were dissected in 3% BSA/DMEM, spun at 16,000× g for 4 min, and the pellets were resuspended in 3% BSA/DMEM at 4°C for 1 hr. Sporozoites were incubated at 37°C for 5 min before being added to a 14 mm glass bottom dish (MatTek) and then visualized using a Zeiss LSM 510 confocal microscope or Leica Inverted Laser Scanning Confocal Microscope (Model Number TCS SP2 AOBS), with a stage heated to 37°C. Image files were processed using Leica LCS software and Image J. Manual tracking was performed using the Image J Manual Tracker plug-in, and the data was compiled in Microsoft Excel and Sigma Plot.
2×104 salivary gland sporozoites in 1%BSA/DMEM were added to the monolayers of Hepa 1,6 cells in the presence of 1 mg/ml TOTO-1 (Invitrogen), a dimeric cyanine nucleic acid dye, for 1 hr at 37°C. Cells were washed with DMEM, fixed in 4% PFA and the number of TOTO-1 positive cells in 50 fields was counted. When indicated, sporozoites were pre-incubated with 1 mM cytochalasin D for 10 min at 28°C and added to cells in the continued presence of the compound.
5×104 salivary gland sporozoites were added to coverslips of semi-confluent Hepa 1–6 cells in DMEM with 10% fetal calf serum and 0.1 mM glutatmine (DMEM/FCS) for 1 hr at 37°C and cells were then fixed with 4% PFA and stained with a double staining technique that distinguishes extracellular from intracellular sporozoites [41], [42]. To determine the kinetics of cell entry, sporozoites were added to cells as outlined above and fixed with 4% PFA at 15, 30 or 45 minutes after their addition. Sporozoites were stained with the double staining assay and those sporozoites that were in the process of entering, i.e. half stained as an intracellular sporozoite and half stained as an extracellular sporozoite, were counted and compared to the total number of sporozoites. To assess productive invasion, sporozoites were incubated with cells for 6 hrs at 37°C, washed, fixed with methanol and stained with UIS-4 antiserum (1∶500 dilution) and mAb 3D11 (1 µg/ml). To quantify EEF development, cells with sporozoites were grown for 40 hrs after which they were fixed with methanol and stained with mAb 2E6 followed by goat anti-mouse IgG conjugated to rhodamine.
Swiss Webster or C57BL/6 mice were injected with the indicated number of salivary gland sporozoites either by i.v. or i.d. inoculation. The onset of blood stage infection was determined by observation of Giemsa-stained blood smears beginning on the third day after sporozoite inoculation.
|
10.1371/journal.pbio.2002912 | Hey1- and p53-dependent TrkC proapoptotic activity controls neuroblastoma growth | The neurotrophin-3 (NT-3) receptor tropomyosin receptor kinase C (TrkC/NTRK3) has been described as a dependence receptor and, as such, triggers apoptosis in the absence of its ligand NT-3. This proapoptotic activity has been proposed to confer a tumor suppressor activity to this classic tyrosine kinase receptor (RTK). By investigating interacting partners that might facilitate TrkC-induced cell death, we have identified the basic helix-loop-helix (bHLH) transcription factor Hey1 and importin-α3 (karyopherin alpha 4 [KPNA4]) as direct interactors of TrkC intracellular domain, and we show that Hey1 is required for TrkC-induced apoptosis. We propose here that the cleaved proapoptotic portion of TrkC intracellular domain (called TrkC killer-fragment [TrkC-KF]) is translocated to the nucleus by importins and interacts there with Hey1. We also demonstrate that Hey1 and TrkC-KF transcriptionally silence mouse double minute 2 homolog (MDM2), thus contributing to p53 stabilization. p53 transcriptionally regulates the expression of TrkC-KF cytoplasmic and mitochondrial interactors cofactor of breast cancer 1 (COBRA1) and B cell lymphoma 2–associated X (BAX), which will subsequently trigger the intrinsic pathway of apoptosis. Of interest, TrkC was proposed to constrain tumor progression in neuroblastoma (NB), and we demonstrate in an avian model that TrkC tumor suppressor activity requires Hey1 and p53.
| Tropomyosin receptor kinase C (TrkC) is a transmembrane receptor at the cell surface and has been described to work paradoxically both as an oncogene and as a tumor suppressor. We partly solved this paradox in a previous study, demonstrating that TrkC is a double-facet receptor: Upon interaction with its ligand neurotrophin-3 (NT-3), TrkC has a tyrosine kinase activity and induces survival and proliferation of the cell; conversely, in the absence of the ligand, TrkC is cleaved and releases a "killer-fragment" that triggers apoptosis. In this study, we analyze the fate of this fragment and show that TrkC killer-fragment is translocated to the nucleus, where it stabilizes the apoptosis inducer p53. We further find that p53 activates the transcription of cytoplasmic molecular partners, which interact with TrkC killer-fragment and induce apoptosis. We also demonstrate that alteration of this mechanism favors tumor growth in neuroblastoma (NB), an avian tumor progression model for a pediatric cancer.
| The neurotrophins nerve growth factor (NGF), brain-derived neurotrophic factor (BDNF), neurotrophin-3 (NT-3), NT-4/5, and their respective receptors neurotrophin receptor p75 (p75NTR) and tropomyosin receptor kinases (TrkA), B, and C have been notably studied for their critical role in neurodevelopment [1]. Yet as TrkA, B, and C are tyrosine kinase receptors (RTKs), their deregulated functions in cancer have been investigated [2]. The overall view is that their kinase activity confers them the ability to activate mitogen-activated protein kinase (MAPK) and phosphoinositide 3-kinase (PI3K)/AKT pathways known to promote cell survival, proliferation, and differentiation under physiological conditions and to contribute to tumor progression when constitutively activated in cancers [2]. The kinase domains of TrkA, B, and C are indeed involved in oncogenic translocations or mutated in cancers (for review [2]). In line with the pharmaceutical rush to design antitumoral treatments based on RTK inhibition, drugs targeting TrkA, B, and C have been under development [3]. Nevertheless, TrkC expression has been paradoxically associated with favorable outcome in pediatric neoplasia, namely neuroblastoma (NB) and medulloblastoma, and was more recently shown to act as a tumor suppressor in colon cancer ([4] and for review [5–8]). We and others have indeed proposed that TrkC has a dual functionality: (i) In presence of its ligand NT-3, TrkC behaves as a classical RTK, transducing positive signals; (ii) in absence of NT-3, TrkC does not stay inactive but rather triggers apoptosis [9, 10]. TrkC thus belongs to the functional family of "dependence receptors." These receptors play a crucial role in constraining the adequate number of cells in a tissue in which the ligand is expressed in a limited amount during neurodevelopment but also during tumorigenesis: Cells in excess that carry an unbound dependence receptor undergo apoptosis [11]. It was demonstrated in different types of tumors that (i) the silencing of the dependence receptor by epigenetic mechanisms or genetic alterations or (ii) the overexpression of the ligand confers to the tumor cells a survival selective advantage: The dependence receptor is then no longer able to trigger apoptosis. TrkC expression was indeed shown to be epigenetically silenced in colon tumors [4, 6]. Along the same line, we also demonstrated that a large proportion of high-grade NB tumors shows an autocrine production of NT-3 as a mechanism to constitutively block TrkC proapoptotic function. It was thus proposed that interfering with ligand–receptor (NT-3/TrkC) interaction, either by gene silencing or the use of a blocking antibody, is associated in different animal models with the inhibition of tumor growth and metastasis [12].
The mechanism for TrkC proapoptotic activity has been investigated in recent years [9, 10, 13]. Upon ligand withdrawal, TrkC appears to be cleaved by caspase-like proteases at 2 sites (D495 and D641) within its intracellular domain, and the released fragment (TrkC 496–641, called the "killer-fragment" [TrkC-KF]) is necessary and sufficient to promote apoptosis. We demonstrated recently that this fragment interacts with cofactor of breast cancer 1 (COBRA1), which shuttles TrkC-KF to the mitochondria [13]. Once at the mitochondria, TrkC-KF and COBRA1 activate B cell lymphoma 2–associated X (BAX) and induce mitochondrial outer membrane permeabilization (MOMP), the release of cytochrome c, and the subsequent apoptosome activation [13].
Here, we show that TrkC-KF is not only cytoplasmic as described previously but is also observed in the nucleus. TrkC-KF is translocated to the nucleus by importins. A 2-hybrid screen allowed us to identify that TrkC-KF then interacts with Hey1, a basic helix-loop-helix (bHLH) transcription factor originally described as an effector of the NOTCH pathway. Hey1 and TrkC-KF bind on mouse double minute 2 homolog (MDM2) promoter and negatively regulate MDM2 transcription. This decrease of MDM2 expression favors p53 stabilization, which triggers the transcription of TrkC proapoptotic partners acting at the mitochondria. We finally show in an avian model of NB tumor progression that Hey1- or p53-silencing abrogates TrkC tumor suppressor activity.
We have previously shown that in absence of its ligand NT-3, TrkC is cleaved by caspase at 2 sites (D495 and D641) within its intracellular domain, leading to the release of several intracellular fragments. This caspase-dependent cleavage can be detected both in vitro and in vivo and is required for apoptosis induction, since the mutation of the caspase sites inhibits apoptosis induced by TrkC [9, 12, 13]. TrkC cleavage by caspases leads to the generation of 3 fragments: TrkC 1–495, TrkC 496–641, and TrkC 642–825 (Fig 1A). In various cell lines, including the murine Neuro2a (N2A) and human SHEP NB cell lines enforced expression of the internal caspase-generated fragment TrkC 496–641 (named TrkC-KF) was associated with cell death induction, while TrkC 1–495 and TrkC 642–825 displayed no proapoptotic activity [9, 13]. In addition to its mitochondrial localization described earlier [13], the green fluorescent protein (GFP)–tagged TrkC-KF (TrkC-KF-GFP) was detected in the nucleus of N2A cells (Fig 1B). As a control, full-length TrkC (TrkC-FL-GFP), the C-terminal cleavage fragment (TrkC-642-825-GFP), and the intracellular fragment of an unrelated receptor—Neogenin (Neo-IC-GFP)—were mostly detected outside the nucleus of transfected cells (Fig 1B). We used GFP-fused fragments as none of the commercial antibodies or antibodies we generated were able to detect endogenous TrkC-KF. We verified that Flag-tagged TrkC-KF was also observed both in the cytoplasm and in the nucleus upon cellular fractionation (S1A Fig). In a yeast 2-hybrid screen using TrkC-KF as bait and a mouse embryonic cDNA library as prey, we identified importin-α3 (karyopherin alpha 4 [KPNA4]) (S1B Fig) [13]. Importins are cargo proteins shuttling cytoplasmic proteins into the nucleus [14, 15]. A proximity ligation assay (Duolink) using a pan-importin antibody and an anti-GFP antibody allowed us to detect a close interaction between TrkC-KF-GFP and endogenous importins (Fig 1C and 1D), suggesting that TrkC-KF is interacting with importins. We thus treated N2A cells with Ivermectin, a pan-importin inhibitor, and performed a fractionation experiment (Fig 1E and S1C Fig). As a control, we used a version of Neo-IC deleted for its nuclear export sequence (Neo-IC-ΔNES) but with an intact nuclear localization sequence (NLS), which is mostly localized in the nucleus [16]. We observed that the amount of TrkC-KF was greatly reduced in the nucleus upon treatment with Ivermectin, while the cytoplasmic pool was not significantly affected (Fig 1E and S1C Fig). As a positive control, Neo-IC nuclear translocation was also affected by Ivermectin treatment. TrkC-KF thus appears to be shuttled in the nucleus by importins. Importins need to first recognize NLSs in the proteins they are supposed to shuttle [14, 15]. Two putative NLSs could be mapped in TrkC-KF sequence, and we thus generated constructs bearing 1 (KFΔNLS1) or the 2 (KFΔNLS1/2) mutations of these putative sites (Fig 1F). While the mutation of NLS1 had no effect on TrkC-KF nuclear translocation, mutation of NLS1/2 greatly reduced the amount of TrkC-KF in the nuclear fraction of transfected SHEP cells (Fig 1G and S1D Fig). Furthermore, the mutation of NLS1/2 (TrkC-KFΔNLS1/2) is sufficient to partially but significantly inhibit TrkC proapoptotic activity (Fig 1H) without affecting its functionality. Indeed, TrkC-KFΔNLS1/2 is able to bind COBRA1, its cytoplasmic partner, as wild-type TrkC-KF does when overexpressed in cells (S1E Fig), suggesting that this mutant is still functional. TrkC-KF nuclear translocation seems therefore necessary for its proapoptotic activity in SHEP cells. In addition, no nuclear export sequence (NES) could be found in TrkC-KF, suggesting that once in the nucleus, TrkC-KF does not return in the cytoplasm. To monitor the role of TrkC-KF in the nucleus, we investigated whether it is able to transactivate gene transcription. To first assay this, TrkC-KF was fused to a Gal4 DNA-binding domain (DBD), and SHEP cells were forced to express TrkC-KF-Gal4DBD together with a construct encoding a luciferase reporter gene under the control of the upstream activating sequence (UAS)-GAL4 promoter. As shown in Fig 1I and S1F Fig, TrkC-KF-Gal4DBD is unable to transactivate the UAS-GAL4, unlike deleted in colorectal cancer intracellular domain (DCC-IC), as shown previously [17]. This result suggests that TrkC-KF has no intrinsic transcriptional activity per se.
As TrkC-KF does not seem to have a transcriptional activity, we looked for putative nuclear interacting partners in the 2-hybrid screen mentioned in Fig 1, using TrkC-KF as bait. We focused on Hey1, which was identified as a putative partner of TrkC-KF in the screen (S1B Fig). Hey1 is a transcription factor that belongs to the bHLH-Orange (bHLH-O) family of transcriptional repressors, together with Hey2 and HeyL [18]. NOTCH pathway activation increases Hey1 expression, leading to the transcriptional inhibition of downstream targets. Therefore, Hey1 is a critical effector of the NOTCH pathway, being involved in cardiac and vascular development [19]. Hey1 is mostly nuclear but has also been detected in the cytoplasm [20]. We showed by confocal microscopy that TrkC-KF-GFP colocalizes with Hey1 tagged with red fluorescent protein (Hey1-RFP) in the nucleus of N2A cells (Fig 2AB). Silencing Hey1 with a designed small interfering RNA (siRNA) was associated with a strong reduction of Hey1 mRNA (S2A Fig) or protein (S2B Fig) without affecting the level of other members of the bHLH-O transcription factor family members Hey2 and HeyL (S2A and S2C Fig). We further showed in SHEP cells that TrkC-KF-GFP and endogenous Hey1 are interacting, by using an anti-Hey1 antibody in a proximity ligation assay (Duolink) (Fig 2C and 2D). No signal of interaction was detected when TrkC 642–825 was used instead of TrkC-KF or when Hey1 was silenced with an siRNA (Fig 2C and 2D). Furthermore, TrkC-KF-GFP nuclear localization was not affected by invalidation of endogenous Hey1 with the siRNA (S2D and S2E Fig). Thus, Hey1 specifically interacts with TrkC-KF and does not seem to be required for TrkC-KF nuclear translocation.
We next confirmed by coimmunoprecipitation performed in human embryonic kidney 293 T (HEK293T) cells that Hey1 interacts with TrkC-KF-GFP and full-length TrkC (TrkC-FL-GFP) (Fig 2E). Furthermore, HeyL, a transcription factor closely related to Hey1, fails to interact with TrkC-KF of TrkC-FL, showing that TrkC specifically interacts with Hey1 only (Fig 2E).
Interestingly, when cotransfected with Hey1, TrkC-KF-GFP was markedly detected in the nucleus in the DNA-bound fraction (Fig 2F). Thus, Hey1 interaction with TrkC-KF allows their joint binding to DNA complex.
In a previous study, we set up conditions in order to transiently express TrkC-FL without inducing an artefactual dimerization and subsequent activation of the kinase domain, as it may be the case with overexpressed RTKs [12, 13]. In this setting, we observed in absence of NT-3 that the expression of TrkC-FL or TrkC-KF induces cell death in various cancer cell lines, including N2A NB cells [13]. We show here that silencing of Hey1 by RNA interference abrogates cell death induced by TrkC-FL or TrkC-KF in N2A cells (Fig 3A). As a control, the apoptosis induced by another dependence receptor, Patched (Ptc), is not affected by Hey1 silencing (Fig 3A). Rather than forcing TrkC expression, we investigated whether cell death induced upon NT-3 withdrawal is also dependent on Hey1. We have previously shown that SHEP cells are expressing both NT-3 and TrkC and that silencing of NT-3 is associated with TrkC-induced apoptosis in these cells [12]. We therefore silenced Hey1 in NT-3-depleted SHEP cells. As shown in Fig 3B, caspase-3 activation induced by NT-3 deprivation is similarly abrogated by cosilencing of endogenous Hey1. We had also previously demonstrated that various other NB cell lines overexpress NT-3 [12]. Among them, CLB-Ga [21] and LAN6 [22] express NT-3 within the same range as SHEP cells (S3A Fig and [12]). Despite the fact that LAN6 also expresses another Trk receptor, TrkA (S3A Fig), we demonstrate that interfering with endogenous NT-3 by RNA interference also triggers caspase-3 activity in LAN6 cells and that this caspase-3 activity can also be abrogated by cosilencing of Hey1 (S3B and S3C Fig). As p75NTR has been suggested to mediate TrkC-induced apoptosis [10], we wondered whether its cosilencing could abrogate the apoptosis induced by NT-3 deprivation in SHEP cells as well as Hey1 did. As shown in S3D and S3E Fig, the caspase-3 activity measured by endogenous NT-3 invalidation is not altered upon treatment with an siRNA targeting p75NTR. p75NTR may thus be dispensable, at least in our model, for the apoptosis triggered by TrkC via Hey1 in absence of NT-3.
In order to work with cells constitutively knock-out for Hey1, we then invalidated endogenous Hey1 in SHEP cells by clustered regularly interspaced short palindromic repeat (CRISPR)/CRISPR-associated protein 9 (CAS9) editing and obtained 2 independent clones in which Hey1 expression was fully abolished. As a control, we used clones that had undergone the same selection process but without the transfection of the guide RNA (gRNA) (Fig 3C). We labeled the various SHEP clones and parental cells with an anticleaved caspase-3 antibody after siRNA NT-3 or siRNA control treatment. Hey1 knock-out clones displayed a much-reduced staining compared to the clones still expressing Hey1 (Fig 3D and 3E). Along this line, in a WB with an anticleaved poly [ADP-ribose] polymerase (PARP) antibody, we observed that PARP cleavage (cPARP) is greatly reduced in Hey1 knock-out clones upon siRNA NT-3 treatment (Fig 3F and S3F Fig).
Altogether, these data indicate that the transcription factor Hey1 acts as a specific proapoptotic partner/effector of endogenous TrkC.
Of interest, Hey1 has been previously identified in a screen aimed at finding new activators of p53 [23]. This study indeed demonstrated that Hey1 stabilizes p53 by down-regulating the expression of the p53 antagonist, MDM2. Along this line, we were able to detect an increase in the amount of p53 protein in SHEP cells forced to express Hey1, and this increase was more important when Hey1 was coexpressed with TrkC-FL or TrkC-KF (Fig 4A and S4A Fig). Conversely, coexpression of Hey1 with the intracellular uncleavable form of TrkC (TrkC-IC-D495N/D641N [TrkC-IC-DM]) did not increase p53 protein amount in SHEP cells (Fig 4A and S4A Fig). In order to determine whether p53 could be involved in the apoptosis mediated by TrkC, we silenced p53 by siRNA and assessed whether TrkC-FL or TrkC-KF could still be proapoptotic. Of interest, silencing p53 abrogates TrkC-FL- or TrkC-KF-mediated apoptosis (Fig 4B and S4B Fig). We then took advantage of HCT116 cells, which have been knock-out for p53, and their parental wild-type counterparts [24]. In p53 constitutively knock-out HCT116 cells, TrkC-FL and TrkC-KF were both unable to trigger apoptosis compared to what is seen in p53 wild-type HCT116 cells (Fig 4C).
Again, rather than forcing TrkC expression, we silenced NT-3 in SHEP cells. As shown in Fig 4D and S4C Fig, silencing of NT-3 by siRNA is associated with an increased p53 protein level, whereas this is not the case when NT-3 and Hey1 are cosilenced. Similarly, cosilencing of NT-3 and p53 by RNA interference blocked NT-3 deprivation-induced apoptosis, demonstrating that p53 is necessary for TrkC apoptotic signaling (Fig 4E). As shown in S4D and S4E Fig, we also confirmed that Hey1 and p53 are necessary to unliganded TrkC–induced apoptosis in another NB cell line that expresses both NT-3 and TrkC: CLB-Ga cells (S3A Fig and [12]).
We then used a chemical inhibitor of p53-dependent transcriptional activation, pifithrin-α [25], and treated SHEP cells with an siRNA targeting NT-3 or an siRNA control. We detected an increased amount of the apoptotic cPARP fragment in siRNA NT-3-treated cells, and this effect was reversed upon treatment with pifithrin-α (Fig 4F and S4F Fig). This result suggests that p53 transcriptional activation is required to mediate TrkC-induced apoptosis.
Finally, in order to determine whether p53 stabilization is mediated by Hey1 transcription repressor function, we coexpressed TrkC-KF with a mutant version of Hey1 bearing 3 point mutations (Hey1-RK3: R50K, R54K, R62K). This triple mutation has been shown to affect Hey1 DNA-binding basic domain and consequently Hey1 transcriptional activity [26]. We observed that TrkC-KF when expressed with Hey1-RK3 is no longer able to induce p53 stabilization, supporting the view that Hey1 transcriptional activity is required for p53-dependent TrkC-KF proapoptotic activity (Fig 4G and S4G Fig).
Because we (i) failed to detect any interaction between MDM2 protein and neither Hey1 nor TrkC-KF by proximity ligation assay (S5A Fig), (ii) observed that Hey1 DBD appears to be important for TrkC-KF/Hey1-mediated stabilization of p53 (Fig 4G), and (iii) identified MDM2 in a chromatin immunoprecipitation sequence (ChIP-Seq) aimed at screening Hey1 binding sites in Hey1 overexpressing cells (Gene Expression Omnibus [GEO] accession number GSE60699 [27]), we hypothesized that Hey1/TrkC-KF may transcriptionally regulate MDM2 expression. We first measured MDM2 expression by quantitative real-time PCR (RT-QPCR) in SHEP cells transfected with various expression plasmids. As described previously by Huang and colleagues, forced expression of Hey1 is able to deregulate MDM2 expression [23]. Of interest, TrkC-FL itself is also able to deregulate MDM2 expression (Fig 5A). As a control, the TrkC 642–825 fragment, the uncleavable form of TrkC (TrkC-DM), TrkC-FL in presence of caspase inhibitor z-vad, or an unrelated overexpressed receptor Ptc does not significantly alter MDM2 expression (Fig 5A). We further assessed the importance of Hey1 in TrkC-mediated MDM2 repression. As shown in Fig 5B, silencing of endogenous Hey1 by an siRNA in SHEP cells, forced to express either TrkC-FL or TrkC-KF, restores MDM2 expression. TrkC-FL- and TrkC-KF-mediated down-regulation of MDM2 was also observed at the protein level, as shown by western blot on transfected SHEP cells (Fig 5C and S5B Fig). Again, as a control, TrkC 642–825, TrkC-DM, or Ptc had no effect on MDM2 protein level (Fig 5C and S5B Fig).
As illustrated in Fig 5D, the MDM2 gene has 2 promoters and an enhancer box (E-box) described as a putative binding site for various transcription factors, including bHLH-O factors like Hey1 [28]. We designed various primers all along the MDM2 promoter region and performed chromatin immunoprecipitation (ChIP) on SHEP cells expressing Flag-tagged versions of TrkC-KF and/or Hey1. ChIP with an antibody targeting endogenous Hey1 resulted in a slight enrichment of the promoter region amplified by primers located in close proximity to the E-box (Fig 5E). As a negative control, no enrichment was observed after the use of primers designed in 5′ or in 3′ of MDM2 promoter region (Fig 5E). Interestingly, TrkC-KF favors endogenous Hey1 binding to MDM2 promoter, as observed by the increased DNA enrichment in TrkC-KF-Flag transfected cells when compared to nontransfected cells expressing endogenous Hey1 only (Fig 5E). When SHEP cells were forced to express TrkC-KF-Flag or Hey1, a similar enrichment of the same promoter region was observed when chromatin was pulled down with either an anti-Flag antibody (targeting TrkC-KF) or an anti-Hey1 antibody (targeting Hey1) (Fig 5F). Together, these results support the view that TrkC-KF and Hey1 interact and bind to the same promoter region near the E-box of MDM2 promoter. To more formally address this question, we silenced Hey1 in TrkC-KF-Flag-expressing SHEP cells. Silencing of Hey1 fully reversed the DNA enrichment observed, indicating that TrkC-KF binding on MDM2 promoter is dependent on its interaction with Hey1 (Fig 5G and 5H). Finally, to assess direct binding of Hey1 and TrkC-KF to the MDM2 promoter E-box, we proceeded to an oligonucleotide pull-down assay using biotin-labeled double-stranded oligonucleotides homologous to the promoter region spanning MDM2 E-box. Oligonucleotides were mutated (mut) or not (WT) on the E-box sequence (Fig 5I) and incubated with lysates of SHEP cells expressing Hey1-Flag and TrkC-KF-GFP. As illustrated in Fig 5J and 5K, we could demonstrate the association of Hey1 (Fig 5J) and TrkC-KF (Fig 5K) with the oligonucleotide containing the wild-type E-box and much less with the mutated E-box oligonucleotide control. The number of bound oligonucleotides is increased when both Hey1 and TrkC-KF are expressed (Fig 5J and 5K). Conversely, silencing of endogenous Hey1 strongly inhibits TrkC-KF binding to the oligonucleotides corresponding to MDM2 promoter (Fig 5L). These data further confirm the association of Hey1 with TrkC-KF in the promoter region spanning MDM2 E-box. Together, these results show that TrkC-KF and Hey1 interact on MDM2 promoter and inhibit MDM2 transcription.
We described in a previous study the shuttling of TrkC-KF at the mitochondria by COBRA1, in which both partners activated BAX and induced MOMP, the subsequent release of cytochrome c, and apoptosome activation [13]. What is then the role of the nuclear pathway and p53 stabilization by TrkC-KF/Hey1 interaction? Are the mitochondrial and the nuclear pathways redundant, or are they sequentially activated? We first expressed TrkC-FL in N2A cells and invalidated Hey1 by siRNA to abrogate cell death. We observed that in this setting, transient overexpression of COBRA1 is sufficient to largely restore apoptosis (S5C Fig). Conversely, Hey1 expression does not significantly restore cell death upon COBRA1 silencing (S5D Fig). These results suggest that the nuclear Hey1/p53 pathway is acting upstream the COBRA1/BAX mitochondrial pathway.
We made the hypothesis that p53 activation may transcriptionally supply the proteins that are essential for the mitochondrial pathway triggered by TrkC. Indeed, COBRA1 promoter has been previously identified as a target of p53 in a genome-wide ChIP assay [29]. We measured COBRA1 expression by RT-QPCR and observed that invalidation of endogenous NT-3 by RNA interference (i.e., the activation of TrkC/Hey1/p53 pathway) increases the amount of COBRA1 mRNA in SHEP cells (S5E Fig). This effect was reversed upon coinvalidation of NT-3 with Hey1 or p53 (S5E Fig). This result suggested that, indeed, p53 is important to allow the expression of COBRA1.
In order to determine whether p53 is responsible for this transcriptional up-regulation of COBRA1, we identified 2 putative p53 binding sites [30] on the COBRA1 promoter and designed various pairs of primers spanning different regions of the COBRA1 promoter (S5F Fig). The chromatin of SHEP cells transfected with either control or TrkC-KF and Hey1 was pulled down with an anti-p53 antibody, and the region encompassing the 2 putative p53 sites was more amplified than the 5′ or 3′ region of the promoter (S5G Fig). These results suggest that p53 indeed binds to the COBRA1 promoter and contributes to the supply of COBRA1 proteins in the cytoplasm so that a pool of TrkC-KF fragments produced by the caspase cleavage in the cytoplasm can be shuttled at the mitochondria by COBRA1 proteins.
We also showed in our previous study that TrkC-KF and COBRA1, once anchored at the mitochondrial membrane, activate BAX—but not B cell lymphoma 2 killer (BAK)—to trigger MOMP [13]. BAX is a well-characterized target of p53 [31], so we also designed pairs of primers among which 1 pair spanned the p53 binding site (S5H Fig). The chromatin region amplified by this pair of primers was greatly amplified in SHEP cells transfected with TrkC-KF and Hey1 (S5I Fig). Conversely, BAK promoter, another p53 target [32], was not amplified when this experiment was repeated with primers spanning the p53 binding site on the BAK promoter (S5J and S5K Fig). These results are consistent with the fact that the TrkC proapoptotic pathway does not require BAK but requires BAX and COBRA1.
Together, these data support the idea that the nuclear apoptotic pathway triggered by TrkC-KF with Hey1 and p53 is essential to provide the adequate number of TrkC proapoptotic partners in the cytoplasm to finally induce MOMP and apoptosome activation.
We previously demonstrated that TrkC-mediated apoptosis constrains tumor growth in NB [12] and proposed that some NB cells escape from TrkC-induced apoptosis by up-regulating NT-3. We thus took advantage of an avian model in which we showed that interference with NT-3/TrkC is associated with NB growth inhibition [12, 33]. SHEP cells were inoculated on the highly vascularized chorioallantoic membrane (CAM) of E10 chicken embryos (Fig 6A). Five days later, a primary tumor was formed at the inoculation site. When we inoculated parental SHEP cells or the control clone of SHEP cells, silenced for NT-3 by RNA interference, the size and weight of the tumors were reduced in comparison with tumors generated by scramble siRNA-transfected cells (Fig 6B and 6C). We inoculated 2 independent clones knock-out for Hey1 (CRISPR/CAS9 edited as shown in Fig 3). The weight of the tumor did not vary significantly in Hey1 knock-out clones upon NT-3 invalidation by siRNA, suggesting that Hey1 is necessary for TrkC to limit tumor progression in absence of NT-3 (Fig 6B). We also observed that cosilencing of NT-3 with either Hey1 or p53 by RNA interference in SHEP cells also reverses siRNA NT-3-induced tumor suppressive effect (Fig 6C), as it had previously been demonstrated upon NT-3 and TrkC cosilencing [12]. The reduction in size of tumors formed by NT-3-silenced SHEP cells was associated with high apoptosis, as shown by TUNEL staining performed on tumor cryosections (Fig 6D and 6E). As expected, this induction of apoptosis triggered by NT-3 silencing is reversed when Hey1 or p53 is invalidated (Fig 6D and 6E). These results demonstrate in vivo that TrkC tumor suppressor activity requires Hey1 and p53.
To analyze whether silencing of the proapoptotic pathway induced by TrkC/Hey1 in absence of NT-3 could be associated with patient poor prognosis, we analyzed various transcriptomic analyses performed on human NB tumors (S6 Fig). It had previously been shown in a limited number of human samples that TrkC expression is associated with favorable outcome in NB [34] and that Hey1 expression is greatly reduced in NB when compared with benign tumors [35]. We made the same observation on a larger cohort, published by T. Wolf on the National Center for Biotechnology Information (NCBI) GEO, analyzed by Agilent-Microarray 44K (GSE45480, 649 samples) for both TrkC and Hey1 (S6A and S6B Fig). TrkC and Hey1 expression is significantly lower in aggressive stage 4 NB tumors than in stage 1–3 NB tumors. We further calculated the intergrade median of expression for NT-3, TrkC, and Hey1. We then selected 3 profiles of tumors based on the mode of action of NT-3 inhibiting the death induced by TrkC and Hey1: (i) tumors which express low levels of NT-3 (beyond the intergrade median of NT-3 expression), high levels of TrkC, and high levels of Hey1—i.e., tumors in which TrkC is prone to induce apoptosis (NT-3low, TrkChigh, Hey1high); (ii) tumors in which the 3 genes are expressed at a low level or silenced—i.e., this death pathway is blocked (NT-3low, TrkClow, Hey1low); and (iii) tumors with other profiles. As shown in S6C Fig, the proportion of tumors in which the TrkC death pathway is "ON" (NT-3low, TrkChigh, Hey1high) is high in low-grade NB1–3 tumors but decreases when the grade increases (22% in NB1–3 and 12% in NB4). Conversely, the percentage of tumors in which the death pathway is "OFF" (NT-3low, TrkClow, Hey1low) is limited in low-grade tumors and increases with tumor aggressiveness in NB (S6C Fig). This result is in agreement with our hypothesis that a functional TrkC proapoptotic pathway is associated with a favorable outcome, whereas the silencing of this pathway confers a selective advantage to NB tumors. We observed the same trend when analyzing 3 microarrays performed on other cohorts (L. Shi [36], O. Delattre [37], and R. Versteeg [35]). Finally, a Kaplan-Meier analysis performed on the survival data of the T. Wolf cohort indicated that tumors having a functional TrkC death pathway (Group A: NT-3low, TrkChigh, Hey1high) are significantly associated with a better prognosis than tumors with silenced TrkC proapoptotic pathway (Group B: NT-3low, TrkClow, Hey1low) or other types of tumors (Group C) (S6D Fig).
Altogether, these data support the view that TrkC constrains tumor growth via Hey1- and p53-mediated apoptosis in vivo, and this proapoptotic pathway is affected in patients with high-grade tumors.
We previously demonstrated that, upon NT-3 deprivation, TrkC-KF is released and shuttled to the mitochondria by its proapoptotic partner, COBRA1. Once at the mitochondria, TrkC-KF and COBRA1 activate BAX and induce the MOMP and the subsequent activation of the apoptosome (S7 Fig and [13]). With this work, we decipher the complex upstream mechanisms involved in TrkC-induced cell death. TrkC-KF is translocated in the nucleus via importins and interacts there with Hey1. Hey1 and TrkC-KF interact and jointly bind to MDM2 promoter E-box, in which TrkC-KF favors Hey1 repressor function on MDM2 transcription. MDM2 transcriptional inhibition promotes p53 stabilization and thus apoptosis. p53 target genes include COBRA1 [29] and BAX [31]. We show here that forced expression of TrkC-KF and Hey1 is associated with enhanced p53 binding to COBRA1- and BAX-respective promoters. Furthermore, we show that COBRA1 expression is enhanced by the activation of the TrkC/Hey1/p53 pathway. When this pathway is altered by Hey1 silencing, the supply of COBRA1 by transient transfection is sufficient to restore TrkC-induced apoptosis. Therefore, the nuclear function of TrkC-KF may not only lead to apoptosis through classic p53 effectors but also through the enhancement of the TrkC mitochondrial pathway by the transcriptional supply of its interactors. Similar fine regulation of proapoptotic protein amounts released in the cytoplasm has indeed already been described for p53 [38].
NB tumors derive from the sympathoadrenal lineage originating from the neural crest cells (NCCs) [39]. NCCs contribute to the formation of the peripheric ganglia, the sympathetic and sensory ganglia, and the medullary region of the adrenal gland. The adequate number of neuronal precursors in the forming ganglia during peripheric nervous development is tightly regulated by peaks of programmed cell death controlled by neurotrophin amounts in the close proximity of the precursors that express at their surface the corresponding neurotrophin receptors (for review [40]). This mechanism of programmed cell death is crucial during gangliogenesis. Indeed, an aberrant number of neuronal precursors in ganglia favors the development of NB, as already observed in mice bearing NB-driving mutations. As a first example, MycN is expressed by migrating NCC [41], and mice having a forced expression of MycN in the sympathetic lineage (TH-MycN) display a hyperplasia of paravertebral ganglia neuroblasts, which are resistant to NGF deprivation–induced apoptosis [42, 43]. In parallel, mice bearing a mutation identified in sporadic and familial cases of NB, anaplastic lymphoma kinase F1178L (ALKF1178L), present a higher number of sympathetic neuroblasts per ganglion than wild-type mice [44]. These studies illustrate the crucial need to clearly identify the actors, which define the adequate number of precursors in the peripheric ganglia. It has been well established that NT-3 and TrkC control the adequate number of precursors in the developing sensory ganglia [45]. We have shown in primary sensory neurons and in the chick embryo model that part of the apoptosis occurring upon NT-3 deprivation during neurodevelopment is actively triggered by TrkC itself [9] via COBRA1 [13]. Interestingly, Kessler and collaborators observed that the constitutive up-regulation of Hey1 expression in mutant mice results in a significant loss of TrkC positive sensory neurons. Conversely, Hey1 mutant mice display an increased number of TrkC-positive sensory neurons [46]. These observations are consistent with the fact that TrkC may trigger apoptosis via Hey1 in supernumerary neurons in a setting where the amount of NT-3 is limited. We decipher in this study the mechanisms that may underlie this process.
Along this line, Barde and collaborators demonstrated in murine models that TrkA and TrkC constrain the adequate number of peripheric neurons during development by actively triggering apoptosis when deprived of their respective ligands, NGF and NT-3. Conversely, TrkB has no proapoptotic activity in this context [10]. Interestingly, TrkA and TrkC expression have long been associated with regressing NB tumors, whereas TrkB expression is a marker of poor prognosis [2]. We have demonstrated here and in a previous study [12] that TrkC proapoptotic activity controls NB tumor progression. It would be of interest to determine whether TrkA and TrkC proapoptotic activity also controls NB tumor initiation in eliminating supernumerary neuroblasts or neurons in the peripheric ganglia.
In their study, Barde and colleagues suggested that p75NTR intracellular domain mediated TrkA and TrkC proapoptotic activity [10]. In our N2A cellular model, p75NTR is not required by TrkC to trigger apoptosis. Moreover, we were able to trigger apoptosis in 3 independent NB cell lines (SHEP, LAN6, and CLB-Ga) displaying various patterns of expression of neurotrophins and their receptors. Further investigations would be needed to investigate the putative interactions between neurotrophin receptors in the control of NB tumor progression.
With this work, we have confirmed in vivo that the TrkC/Hey1/p53 proapoptotic pathway indeed limits NB tumor growth. p53 and its inhibitor MDM2 have been particularly studied in NB (for review [47]). However, while mutations in p53 are generally considered to affect half of human adult cancers, pediatric cancers are characterized by the lack of p53 mutations [48–50]. More specifically, in NB, p53 is mutated in less than 1% of the tumors at diagnosis [51]. Tumors with wild-type p53 probably rely on other mechanisms to inactivate p53, and it is thus of interest to note that in pediatric tumors, and more specifically in NB, MDM2 is frequently up-regulated [52]. In the present study, we analyzed the transcriptomic public data sets available and showed that the silencing of TrkC proapoptotic pathway (NT-3low, TrkClow, HEYlow) is also associated with poor patient outcome (S6 Fig). In parallel, TrkC expression has been shown to be epigenetically controlled in various cancers [4, 6]. It is thus tempting to investigate whether reactivating TrkC proapoptotic activity in these patients with p53 wild-type tumors may constitute an interesting therapeutic strategy.
N2A, SHEP, and HEK293T were described previously [13]. WT HCT116 and p53-KO HCT116 were kindly provided by B. Vogelstein (Ludwig Center at Johns Hopkins, Baltimore, MD) [24]. SHEP, LAN6, and CLB-Ga were kindly provided by V. Combaret (CRCL, Lyon)[12]. N2A cells were grown in DMEM/F-12, GlutaMAX (Life Technologies), supplemented with 10% FBS (Lonza); SHEP, LAN6, and CLB-Ga cells were grown in RPMI1640, GlutaMAX (Life Technologies), supplemented with 10% FBS (Lonza); and HEK293T, WT HCT116, and p53-KO HCT116 cells were grown in DMEM (Life Technologies), supplemented with 10% FBS (Lonza). The plasmid constructs and siRNA were transfected using JetPrime (PolyPlus) for cell death assays and Lipofectamine RNAiMAX (Life Technologies) for RT-QPCR assays following manufacturer’s instructions. Caspase activity was inhibited in SHEP cells by treatment with 20 μM of z-VAD-fmk (Merck-Millipore), a general caspase inhibitor. The pan-importins inhibitor Ivermectin (Sigma I8898) was added to the cells at a final concentration of 10 μM 2 h before cell collection. Pifithrin-α (Sigma P4359), p53 inhibitor was used at a concentration of 20 μM for 30 h.
The plasmids encoding full-length TrkC, TrkC-KF, TrkC-642-825, TrkC495-825, TrkC-DM (TrkC-D495N/D641N), Ptc, DCC-IC, and Neogenin IC GFP were described elsewhere [9, 16]. The plasmids encoding Hey1, HeyL, and Hey1-RK3 were a kind gift of M. Gessler (University of Wuerzburg, Germany) [26]. TrkC-KF-ΔNLS 1 and 2 were generated by site-directed mutagenesis using QuickChange kit (Stratagene) and the following primers: TrkC-KF-ΔNLS1 forward: 5′-CATATGTTCAACACATCGCCGCCGCCGACATCGTGTTGAAG-3′, reverse: 5′-CTTCAACACGATGTCGGCGGCGGCGATGTGTTGAACATATG-3′. TrkC-KF-ΔNLS1/2 forward: 5′-GAGAGACATCGTGTTGGCCGCCGCCGCCGGTGAGGGAGCCTTT-3′, reverse: 5′-CAAAGGCTCCCTCACCGGCGGCGGCGGCCAACACGATGTCTCT-3′.
The plasmid encoding the sgRNA targeting Hey1 (TGACGCGCACGCCCTTGCTA) cloned into the pSPCAS9 BB-2A-GFP (PX458) was generated by GenScript.
siRNAs were purchased from Sigma-Aldrich for siRNA control (siRNA Universal Negative Control #2 SIC002_10Nmol), siRNA huHey1 (NM_001040708; SASI_Hs02_00309099), siRNA hup53 (NM_000546; SASI_HS02_00302766), siRNA huCOBRA1 (NM_015456; SASI_hs01_00236976), siRNA muCOBRA1 (SASI_Mm01_00110121), and from Santa-Cruz for siRNA control (sc-37007), siRNA mHey1 (sc-42126), siRNA huNGFR p75 (sc-36057), and siRNA huNT-3 (sc-42125).
The 2-hybrid screen was performed by Hybrigenics (Paris, France) using the Mouse Embryo Brain RP2 library as a prey and pB27 (N-LexA-bait-C fusion) and pB66 (N-GAL4-bait-C fusion) vectors. TrkC-KF construct was used as bait.
N2A and SHEP cells were transfected with the indicated constructs: After 4 h, the medium was replaced by medium without serum for 24 h to 72 h. Caspase-3 activity was measured as described in [12], using the Ac-DEVD-AFC substrate assay (Biovision, K105-400). Cell death percentages were assessed by trypan blue exclusion, as described in [13].
HEK293T cells were lysed in 50 mM HEPES pH 7.6, 125 mM NaCl, 5 mM EDTA, and 0.1% NP-40 in the presence of protease inhibitors and were further incubated with an anti-GFP antibody (A11122, Life Technologies) and then with protein G sepharose (Sigma Aldrich) to pull down proteins of interest. Western blots were performed using an anti-Flag antibody (F3165, Sigma Aldrich).
HEK293T, SHEP, and N2A cells were lysed in 50 mM HEPES pH 7.6, 125 mM NaCl, 5 mM EDTA, and 0.1% NP-40 in the presence of proteases inhibitors. For p53 western blots, SHEP cells were lysed in 50 mM Tris–HCl pH 7.5, 100 mM NaCl, 10% glycerol, 0.1% NP-40, and 0.2 mM EDTA in the presence of proteases inhibitors. Western blots were quantified using the ImageJ64 software.
We used the following antibodies: anti-GFP (A11122, Life technologies), anti-GAPDH (sc-25778, Santa Cruz), anti-Histone (07–449, Millipore), anti-Flag (F3165, Sigma), anti-Actin (MAB1501R, Chemicon), anti-Hey1 (anti-HRT1, sc-16424, Santa Cruz), anti-p53 (SAPU, [53]), anti-Ptc (sc-6149, Santa Cruz), anti-TrkC (AF1404, R&D), anti-MDM2 (VMA00406, BioRad), anti-COBRA1 (F7E4, GeneTex), anti-BAX (sc-526, Tebu Bio), anti-BAK (G-23, Santa Cruz), anti-NT-3 (sc-547, Santa Cruz), and anticleaved PARP (9541T, Cell Signaling). The following antibodies were used for proximity ligation assays (DuoLink): anti-GFP (TP401, Biolabs), anti-Flag (F3165, Sigma), anti-Hey1 (anti-HRT1, sc-16424, Santa Cruz), anti-p53 (sc-126, Santa Cruz), and anti-importins (I1784, Sigma Aldrich).
N2A, SHEP, LAN6, and CLB-Ga cells were cultured on coverslips, transfected with indicated plasmids using JetPrime, and then fixed 20 min in 4% paraformaldehyde and permeabilized in PBS/0.2% Triton. Nuclei were stained using DAPI. Images were obtained by confocal microscopy and analyzed using ImageJ64. For cleaved caspase-3 staining, after permeabilization, SHEP, LAN6, and CLB-Ga cells were incubated in blocking solution (PBS-BSA2%-normal serum 2%) for 1 h before incubation overnight with anticleaved caspase-3 antibody (9661, Cell Signaling) diluted to 1:1,000 in PBS. After incubation with secondary antibody (Alexa Fluor Donkey anti-Rabbit IgG 555, Invitrogen A31572) diluted to 1:2,000 in PBS for 1 h, slides were mounted in DAPI-fluoromount G (17984–24, EMS) and imaged using a Zeiss AxioImager microscope. Quantification was performed using ImageJ64.
SHEP cells were transfected using JetPrime (PolyPlus). When indicated, cells were treated with Ivermectin (Sigma I8898) 10 μM 2 h before collection. After 4 h, the medium was replaced with medium without serum. 24 h after transfection, cells were harvested, and nuclei were isolated from cytoplasm using the Nuclei Pure Prep Isolation kit (Sigma Aldrich). Input, cytoplasmic, and nuclear fractions were analyzed by western blot, with GAPDH as cytoplasmic marker and Histone H3 as nuclear marker. HEK293T cells were transfected with TrkC-KF-GFP and Hey1 using JetPrime. After 4 h, the medium was replaced with medium without serum. Twenty-four h after transfection, cells were harvested, and cytoplasmic, DNA-bound, and DNA-unbound fractions were separated using the Subcellular Protein Fractionation Kit for Cultured Cells (ThermoFisher Scientific). Fractions were analyzed by WB, using Histone as nuclear marker and actin as loading control.
To assay protein interactions in cells by fluorescence, the DuoLink PLA kit was used (Sigma Aldrich). Briefly, cells were cultured on coverslips and then fixed in 4% PFA for 30 min and washed using PBS/7.5% glycine for 5 min. Cells were then permeabilized in PBS/0.2% Triton and incubated in a blocking solution for 30 min (PBS/2% BSA). After an overnight incubation with the primary antibodies, cells were incubated with Plus and Minus PLA probes. The probes were ligated and amplified using the Duolink In Situ Detection Reagents Orange (Sigma Aldrich). After several washes with the Duolink In Situ Wash Buffers for Fluorescence (Sigma Aldrich), nuclei were stained using DAPI, and coverslips were mounted in fluoromount.
The analysis was made by fluorescence microscopy, and signal quantification was assessed using the ImageJ64 software to count the number of red fluorescence spots compared to total cell number (assessed using DAPI staining).
To assay mRNA expression, total RNA was extracted from cells using the Nucleospin RNAII kit (Macherey-Nagel). One microgram of RNA was reverse-transcribed using the iScript cDNA Synthesis Kit (Bio-Rad). RT-QPCR was performed using a Light-Cycler 480 (Roche Applied Science) and the FastStart TaqMan Probe Master Mix (Roche Applied Science). The primers and probes (Universal Probe Library, Roche Applied Science) used are indicated on S1 Table.
To assay TrkC-KF transcriptional activity, SHEP cells were transiently transfected with the indicated constructs fused to Gal4 DBD, a plasmid containing the firefly luciferase gene under the UAS-Gal4 control, and a plasmid coding for the Renilla luciferase gene under the CMV promoter as a control. To assess Firefly and Renilla luciferase activities, Dual-Luciferase Reporter Assay System was used following manufacturer’s instructions (Promega). Data represent Firefly value over Renilla value, indexed to control (Gal4).
SHEP cells were transiently transfected with the plasmid encoding SpCAS9, Hey1-targeted gRNA, and GFP (Genscript, target sequence: GATAACGCGCAACTTCTGCC) using JetPrime. Two d after transfection, GFP-positive cells were sorted as single cells in 96-well plates for clonal selection. Hey1 mRNA expression level was measured by RT-QPCR for all obtained clones, and 2 clones with significant decrease in Hey1 mRNA level compared to the parental SHEP cell line were selected for further analysis. Editing of the Hey1 gene was confirmed by sequencing for both clones. Control clones were obtained using a plasmid encoding SpCAS9 and GFP (Addgene). GFP-positive sorted clones were analyzed by RT-QPCR to confirm no change in Hey1 mRNA expression level as compared to the parental SHEP cell line, and 2 of them were selected for further analysis.
SHEP cells transfected with the indicated plasmids were incubated with 1% formaldehyde for cross-link: Reaction was stopped by the addition of 125 mM glycine. Cells were scraped in swelling buffer (25 mM Hepes pH 7.9, 1.5 mM MgCl2, 10 mM KCl, and 0.1% NP-40), and nuclei were isolated using dounce homogenizer. After centrifugation, nuclei were resuspended in sonication buffer (50 mM Hepes pH 7.9, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% Nadeoxycholate, and 0.1% SDS) and sonicated to obtain chromatin fragments of 400 bp to 600 bp size. Chromatin was then incubated with primary antibodies or isotypic IgG (Sigma Aldrich) overnight: anti-Flag (F3165, Sigma), anti-Hey1 (anti-HRT1, sc-16424, Santa Cruz), and anti-p53 (sc-126, Santa Cruz). Complexes were pulled down using protein G sepharose (Sigma Aldrich). After washes, immune complexes were eluted, and cross-linking was reversed at 65°C. Eluates were incubated with RNase A and proteinase K; then, DNA was recovered by phenol-chloroform extraction. DNA fragments were analyzed by RT-QPCR using a Light-Cycler 480 (Roche Applied Science) and the FastStart TaqMan Probe Master Mix (Roche Applied Science). The primers and probes (Universal Probe Library, Roche Applied Science) used are indicated on S2 Table.
To assay Hey1 and TrkC-KF ability to bind MDM2 promoter, SHEP cells were transiently transfected with Hey1-Flag and TrkC-KF-GFP-expressing plasmids: Cells were resuspended in hypotonic buffer (10 mM HEPES, pH 7.9, 1.5 mM MgCl2, 10 mM KCl, 0.5 mM dithiothreitol with protease inhibitors [Roche]) and incubated on ice for 10 min. Nuclei were isolated by centrifugation, resuspended in RIPA buffer supplemented with complete proteases inhibitors, and incubated on a rotating wheel at 4°C for 20 min to obtain a nuclear proteic extract. Meanwhile, biotinylated oligonucleotides corresponding to Hey1 and TrkC-KF binding sites on the MDM2 promoter, and containing either a WT E-box (CACGTG) or a mutated E-box (CCCGGG), were annealed to form double-stranded oligonucleotides of 80 bp size. For WT E-box, the forward oligonucleotide is 5′-biotinylated with the following 5′–3′ sequence: gggggctcggggcgcggggcgcggggcatggggcacgtggctttgcggaggttttgttggactggggctaggcagtcgcc. WT E-box reverse oligonucleotide: ggcgactgcctagccccagtccaacaaaacctccgcaaagccacgtgccccatgccccgcgccccgcgccccgagccccc.
For mut E-box, the forward oligonucleotide is 5′-biotinylated with the following 5′–3′ sequence: gggggctcggggcgcggggcgcggggcatggggcccggggctttgcggaggttttgttggactggggctaggcagtcgcc. Mut E-box reverse oligonucleotide: ggcgactgcctagccccagtccaacaaaacctccgcaaagccccgggccccatgccccgcgccccgcgccccgagccccc.
Three micrograms of double-stranded biotinylated oligonucleotides were incubated with 300 μg of nuclear protein extract for 2 h at 4°C. Complexes were pulled down using 50 μL of streptavidin-agarose beads (Sigma Aldrich) incubated for 1 h at 4°C. The protein-DNA-streptavidin-agarose complex was washed 3 times with RIPA buffer and loaded onto an SDS gel. Detection of Hey1-Flag and TrkC-KF-GFP proteins was performed by western blot as described in [13].
SHEP cells were transfected with jet prime and siRNA NT-3 (100 nM) and/or siRNA Hey1 or sip53 (50 nM) 24 h before inoculation. Five million cells were suspended in 25 μl complete medium and 25 μl matrigel (Corning 356231) and seeded on 10-d-old (E10) chick CAM. On day 15, tumors were resected and weighted. To monitor apoptosis on primary tumors, they were fixed on 4% PFA, cryoprotected by overnight treatment with 30% sucrose, and embedded in Cryomount (Histolab). TUNEL staining was performed on tumor cryostat sections (Roche Diagnostics), and nuclei were stained with DAPI. Three to eight sections were analyzed at ×20 magnification for at least 3 tumors for each condition. TUNEL and DAPI positive cells were counted by ImageJ64 software.
The expression values analyzed here are publically available in GEO database (http://www.ncbi.nlm.nih.gov/geo/). T. Wolf cohort (GSE45480[54]) analysis was performed with Agilent-020382 Human Custom Microarray 44k (GPL16876); the following data set probes were used: NT-3 (NTF3) UKv4_A_23_P360797, TrkC (NTRK3) UKv4_A_23_P205900, UKv4_A_23_P88538, and Hey1 UKv4_A_32_P83845. In this study, for TrkC and Hey1 expression values, a mean of the values obtained with the various probes was calculated. Kaplan-Meier analysis was performed in R2: Genomics Analysis and Visualization Platform (http://r2.amc.nl). The p-value is calculated to determine the optimal cutoff and is finally corrected by Bonferoni as described in [55]. A new grouping variable was made on the basis of NT-3, TrkC, and Hey1 as described in the main text.
Number of experiments and statistical tests used is indicated in figure legends. Statistical treatment of the data was performed with Prism 6.0e (GraphPad) and BiostaTGV online statistical software (http://marne.u707.jussieu.fr/biostatgv/).
|
10.1371/journal.ppat.1007378 | REM1.3's phospho-status defines its plasma membrane nanodomain organization and activity in restricting PVX cell-to-cell movement | Plants respond to pathogens through dynamic regulation of plasma membrane-bound signaling pathways. To date, how the plant plasma membrane is involved in responses to viruses is mostly unknown. Here, we show that plant cells sense the Potato virus X (PVX) COAT PROTEIN and TRIPLE GENE BLOCK 1 proteins and subsequently trigger the activation of a membrane-bound calcium-dependent kinase. We show that the Arabidopsis thaliana CALCIUM-DEPENDENT PROTEIN KINASE 3-interacts with group 1 REMORINs in vivo, phosphorylates the intrinsically disordered N-terminal domain of the Group 1 REMORIN REM1.3, and restricts PVX cell-to-cell movement. REM1.3's phospho-status defines its plasma membrane nanodomain organization and is crucial for REM1.3-dependent restriction of PVX cell-to-cell movement by regulation of callose deposition at plasmodesmata. This study unveils plasma membrane nanodomain-associated molecular events underlying the plant immune response to viruses.
| Viruses propagate in plants through membranous channels, called plasmodesmata, linking each cell to its neighboring cell. In this work, we challenge the role of the plasma membrane in the regulation of virus propagation. By studying the dynamics and the activation of a plant-specific protein called REMORIN, we found that the way this protein is organized inside the membrane is crucial to fulfill its function in the immunity of plants against viruses.
| The cell plasma membrane (PM) constitutes a regulatory hub for information processing [1]. Current knowledge suggests that PM proteins and lipids dynamically associate with each other to create specialized sub-compartments or nanodomains [2], that regulate the cellular responses in space and time [3–5]. For instance, modeling of the localization behavior of a PM-bound receptor and its downstream interactor before and after ligand perception in animal cells suggests that PM-partitioning into nanodomains improves the reliability of cell signaling [6]. In plants a recent example of PM partitioning shows that despite sharing several signaling components, the immune and growth receptors FLS2 and BRI1 are divided into context-specific nanodomains to confer signaling specificity [7]. The REMORIN (REM) family is one of the best-characterized PM nanodomain-associated proteins in plants [7–12]. The association of REMs to the PM is mediated by a short sequence at the extremity of the C-terminus of the protein, called REM-CA (REMORIN C-terminal Anchor) [13, 14]. The REM C-terminal domain contains a coiled-coil (residues 117–152, [15]) which is thought to regulate REM oligomerization [11, 14, 16] and may be involved in regulating REM spatial organization at the PM [15]. Members of the REM family have been associated with plant responses to biotic [9, 17, 18], abiotic stress [19, 20] and developmental clues [12] and current view suggests they could regulate signaling events through nanodomain association [21]. However, the molecular mechanisms leading to REM-associated downstream events remain elusive.
Several REM proteins have been identified as components of the plasmodesmata-plasma membrane subcompartment (PD-PM) [8, 22, 23]. PD are membranous nanopores, crossing the plant cell wall and enabling cytoplasmic, endoplasmic reticulum and PM continuity between adjacent cells. They regulate the intercellular transport of proteins and small molecules during development and defense [24, 25]. The PD-PM is a particular subcompartment of the PM, which displays a unique molecular composition, notably enriched in sterols [26]. The movement of macromolecules through PD can be tightly controlled through modulation of the PD size-exclusion limit (SEL) via hypo- or hyper-accumulation of callose at the PD neck region [27–29]. Overexpression of GRAIN SETTING DEFECT 1 (GSD1) encoding a phylogenetic-group 6 REM protein from rice, restricts PD aperture and transport of photo-assimilates [23].
PDs are also the only route available for plant viruses to spread from cell-to-cell. Potato virus X (PVX) promotes its cell-to-cell movement via modification of PD permeability [30] through the action of TRIPLE GENE BLOCK PROTEIN 1 (TGBp1) [31]. Overexpression of StREM1.3 (Solanum tuberosum REM from group 1b, homolog 3 [32], further referred as REM1.3) hampers TGBp1’s ability to increase PD permeability [33]. How REM1.3 obstructs TGBp1 action is still unknown. Here, we used REM1.3 and PVX pathosystem in the solanaceae Nicotiana benthamiana, because PVX cannot infect Arabidopsis [34] and N. benthamiana is a widely used model for research on plant-virus interaction [35]. We previously showed that REM1.3 lateral organization into nanodomains at the PM is directly linked with its ability to restrict PVX movement and regulate PD conductance [36].
REM1.3 was the first REM family member discovered and initially described as a protein phosphorylated upon treatment with oligogalacturonides, which are plant cell wall components and elicitors of plant defense [37, 38] The biological relevance of REM phosphorylation is not known of different REM phospho-statuses suggest that the activity of these proteins could be regulated by phosphorylation during plant-microbe interactions [16, 17, 39, 40].
In the present paper, we show that phosphorylation of REM dictates its membrane dynamics and antiviral defense by the reduction of PD permeability. Our data point towards a model in which viral proteins such as the Coat Protein (CP), TGBp1 from PVX and 30K proteins from Tobacco mosaic virus (TMV) elicit the activation of protein kinase(s), which in turn phosphorylate(s) REM1.3 at its N-terminal domain. In turn, REM1.3's phospho-status regulates its spatial-temporal organization at the PM and association with PD. The latter is associated with PD closure by induction of callose deposition at PD pit fields and restriction of viral cell-to-cell movement. Last, we further provide evidence that the membrane bound Arabidopsis CALCIUM-DEPENDENT PROTEIN KINASE 3 (CPK3) interacts with the taxonomic group 1b REMs in vivo, phosphorylates REM1.3 in vitro and restricts PVX propagation in a REM-dependent manner. Collectively, this study brings valuable information about the involvement of PM nanodomains dynamics during the establishment of membrane-bound signaling processes.
Group 1 and group 6 REM have been described as proteins regulating PD size-exclusion limit [8, 23, 33]. REM1.3 plays a role in restricting PVX passage through PD channels [8], [33] counteracting PVX movement proteins which promote PD opening [41]. To study the potential function of REM1.3 at PD in response to PVX infection, we surveyed simultaneously PD callose content and REM1.3 PD localization in healthy or PVX-infected N. benthamiana transiently expressing YFP-REM1.3 [42] (S1 Fig). Our analysis showed a significant increase in callose deposition in PVX-infected cells compared to mock conditions (Fig 1A and 1B). This finding suggests the recognition of PVX-encoded elicitors and the mobilization of a plant defense response leading to an increase of callose accumulation at PD pit fields.
Since protein activation is often linked to changes in subcellular localization [3, 44], we next examined whether PVX infection triggers changes in REM1.3 association with PD. Calculation of the PD index (ratio between fluorescence intensity of YFP-REM1.3 at the aniline-labeled PD pit fields and fluorescence at the PM around the pit fields [28], S1 Fig). Fig 1A and 1B showed that despite its role on PD regulation, YFP-REM1.3 is not enriched in the PD region of healthy N. benthamiana epidermal cells. We however reproducibly observed a slight increase of YFP-REM1.3 PD index upon PVX infection suggesting that PVX perception modulates REM1.3 localization and association with the PD pitfields (Fig 1A and 1B).
To gain further insights into REM1.3 dynamic localization at the PM upon PVX infection, we applied single-particle tracking Photoactivated Localization Microscopy in Variable Angle Epifluorescence Microscopy mode (spt-PALM VAEM) in living N. benthamiana epidermal cells [45] in absence or presence of PVX. We used the photoconvertible fluorescent protein EOS [46, 47] fused to REM1.3 to visualize, track, and characterize mobility behavior of single REM1.3 molecules. In addition, nanoscale localizations of single molecules observed overtime were computed to obtain super-resolution images and analyze REM1.3 organization at a molecular level. By this approach, we recently studied the protein organization and mobility parameters of single EOS-REM1.3 molecules in non-infected conditions and found that EOS-REM1.3 displays an immobile and confined PM localization pattern, as commonly observed for plant membrane-associated proteins (Fig 1C–1E) [48], [36]. Reminiscent of these data, previous studies using different techniques described REM-associated PM domains to be predominantly laterally static [36, 48, 49]. Analysis of PVX-infected cells demonstrated an increase of EOS-REM1.3 diffusion coefficient (D) and mean square displacement (MSD), reflecting an increase of REM1.3 mobility (Fig 1C–1E). We next apply mathematical computation (Voronoï tessellation method [36, 50]) to compare the supra-molecular organization of EOS-REM1.3 of live PALM data in mock- and PVX-infected conditions (Fig 1F and 1G). Computation of EOS-REM1.3 single molecule organization features demonstrated a modulation of REM1.3 nanodomain-organization upon PVX infection (Fig 1G). Following PVX infection, the EOS-REM1.3-formed nanodomains are bigger in size, and there is a slight decrease of the proportion of molecules that localized into nanodomains as well as a decrease in the number of nanodomains formed. Overall, in both conditions, EOS-REM1.3 nanodomains represented similar proportions of the total PM surface. Additionally, a decrease in the localization density (number of molecules observed per μm2 per s) showed that upon PVX infection, there was less REM1.3 protein at the PM level. Overall, the changes of REM1.3 distribution under PVX infection i.e. enrichment of YFP-REM1.3 in the PD pit field regions, the increase of REM1.3's mobility and the modulation of REM1.3 nanodomain organization, suggest that the plant cell modulates PD-PM and PM nanodomain dynamics to circumvent PVX infection.
REM1.3 overexpression restricts PVX local and systemic spreading in both Solanum lycopersicum [8] and Nicotiana benthamiana [33, 36] (S2A and S2B Fig). Because REM1.3 protein level is not affected by PVX infection (S2C and S2D Fig), we assumed that neither synthesis nor degradation of the protein is modified by PVX, but perhaps post-translational modifications. As REM1.3 was originally discovered as a PM-associated phosphorylated protein [38], we first asked whether REM1.3 could be phosphorylated by leaf protein extracts. Equal protein amounts of microsomal and soluble extracts from N. benthamiana leaves were used as a potential kinase source to phosphorylate affinity-purified full-length 6His-REM1.3 in an in vitro kinase assay in the presence of ATP [γ-33P]. Autoradiography revealed the presence of a clear band corresponding to a phosphorylated form of 6His-REM1.3 by kinase(s) present in the microsomal fraction (Fig 2A). The intensity of this band was completely abolished by competition with cold ATP, but not cold AMP, indicating a valid experimental set-up to study a genuine transphosphorylation event (S3A Fig). Phosphorylation of 6His-REM1.3 was almost undetectable in soluble fractions, representing cytosolic kinases (Fig 2A). In silico analysis predicted phosphorylation sites throughout REM1.3 sequence (Diphos, DEPP and NETPHOS prediction softwares). In agreement with the location of the sites presenting the highest phosphorylation potential, we experimentally found that REM1.3 was phosphorylated in its N-terminal domain (residues 1–116, hereafter 6His:REM1.3N) whereas the C-terminal domain (residues 117–198, hereafter 6His:REM1.3C) did not present any detectable phosphorylation (S3B and S3C Fig).
We next tested whether PVX activates the kinase(s) that phosphorylate(s) REM1.3. Our results unveiled that microsomal and PM fractions extracted from symptomatic PVX-infected leaves promoted higher levels of 6His-REM1.3 phosphorylation compared to non-infected plants (Fig 2B and 2C). Studies have shown that functionally different viral components, such as virus-encoded proteins and double-stranded RNA, can trigger plant defense responses [51–56]. We therefore examined whether the PVX genome in its free form was an eliciting signal for kinase activation. We found that the addition of total RNAs extracted from PVX-infected plants in the kinase reaction mix did not alter the levels of 6His-REM1.3 phosphorylation (Fig 2D). We then examined whether the sole expression of individual viral movement proteins was sufficient to trigger REM1.3 phosphorylation (Fig 2E). Importantly, our results demonstrated that the expression of TGBp1 and Coat Protein (CP) fused to GFP triggered the strongest levels of 6His-REM1.3N phosphorylation to the same extent as the full PVX-GFP construct (Fig 2F and 2G and S3E Fig for controls of viral fluorescent-tagged protein expression as described in [41]). In good agreement, expression of a TGBp1-deleted version of PVX (PVXΔTGBp1) decreased 6His-REM1.3 phosphorylation levels compared to wild-type PVX extracts (S3D Fig). Expression of TGBp2 and infiltration of the empty Agrobacterium strain alone protein also induced 6His-REM1.3N phosphorylation, albeit less effective than TGBp1 and CP proteins (Fig 2F and 2G). In accordance with previous reports suggesting REM phosphorylation during plant-microbe interactions [9], Agrobacterium infected Ν. benthamiana extracts induced much stronger REM1.3 phosphorylation than the water control condition (Fig 2F and 2G). Furthermore, we found that the 30K-RFP protein from Tobacco mosaic virus (TMV) also induces REM phosphorylation (S3D Fig). Similar to PVX-TGBp1, REM1.3 interferes with the ability of TMV-30K to increase PD permeability [33] and overexpression of REM1.3 restricts TMV-GFP cell-to-cell movement in N. benthamiana epidermal cells (S4A Fig).
Altogether our data suggest an additional role of REM-mediated plant response against TMV and possibly to bacteria. Our results also indicate that REM1.3 phosphorylation status is modulated by the perception of viral proteins by plant cells.
Since phosphorylation of REM occurs upon PVX infection, we next aimed to functionally characterize the importance of REM1.3 phosphorylation for the regulation of PVX cell-to-cell movement. Despite our efforts, the identification of in vivo phosphorylation sites of REM1.3 appeared technically challenging and remained unsuccessful. In silico predictions and in vitro kinase assays however showed that REM1.3N displays regions of intrinsic disorder and presents the highest potential of phosphorylation (Fig 2C–2F and Fig 3A). For functional characterization, we selected the three putative phosphorylation Serine(S) /Threonine(T) sites present in REM1.3N, namely S74, T86 and S91, that presented high scores of phosphorylation prediction in intrinsic disorder regions (Fig 3A). S74 and S91 are conserved across the phylogenetic group 1b of REM proteins, suggesting functional redundancy (S5A Fig) [32, 57]. S74 and S91 were the analogous residues identified as phosphorylated in vivo in the group 1b REM AtREM1.3 (At2g45820) of Arabidopsis thaliana (hereafter Arabidopsis) in a stimuli-dependent manner [39, 40, 57]. Biochemical analysis showed that α-1,4-poly-D-galacturonic acid (PGA)-induced phosphorylation of StREM1.3 occurs on T32, S74 and T86 [58]. T86 is not conserved in Arabidopsis but it is conserved in Solanaceae REM proteins, such as in N. benthamiana (S5A Fig). By an in vitro kinase assay, we show that phosphorylation occurs within three potential phosphor-residues, since mutation of S74, T86 and S91 to the non-phosphorylatable Aspartic acid (D), generating the 6His-REM1.3DDD mutant abolished REM phosphorylation by the PVX-activated kinase(s) (Fig 3B and 3C).
To discriminate which residues are functionally relevant in the context of PVX-GFP propagation, we generated RFP-tagged REM1.3 phosphomutants, individually mutated at those sites to the non-phosphorylatable Alanine. Transient expression in N. benthamiana coupled with PVX-GFP infection assays demonstrated that individual phospho-null mutations at those sites induce a loss of function of REM1.3 in restricting PVX-GFP spreading (Fig 3D). This result suggests that phosphorylation of either S74, T86 and S91 is important for REM1.3 function.
To further characterize the relevance of different REM1.3 phospho-statuses in the context of PVX-GFP propagation and PD-aperture regulation, we analysed RFP-tagged REM1.3DDD to mimic constitutive phosphorylation hereafter termed phosphomimetic mutant, or to Alanine (REM1.3AAA) hereafter termed phosphodead mutant. Infection assays in N. benthamiana confirmed that the phosphodead mutant completely lost REM1.3 ability to restrict PVX-GFP cell-to-cell movement, while the phosphomimetic mutant maintained this ability (Fig 3D). TMV-GFP propagation was similarly affected by the phospho-status of REM1.3 (S4A Fig). We then analyzed the capacity of REM1.3 phosphomutants to regulate PD aperture in the absence of viral infection. As previously described [33, 36], RFP-REM1.3 reduces the PD size-exclusion limit as measured by free-GFP cell-to-cell diffusion (Fig 3E). Detailed analysis of REM1.3 phosphorylation mutants demonstrated that the phosphomimetic mutant recapitulated REM1.3 activity towards PD-aperture regulation, while the phosphodead mutant did not (Fig 3E).
Altogether, these results provide strong evidence that REM1.3's phosphorylation state at the evolutionarily conserved positions of S74, T86 and S91 is linked to its function in controlling viral infection and PD conductance.
Both REM1.3 phosphomimetic and phosphodead mutants maintained PM localization, similarly to wild-type REM1.3, when transiently expressed in fusion with YFP in N. benthamiana (S4B Fig). Upon PVX infection we observed a modulation of REM1.3 PD-association and PM dynamics (Fig 1), linked to REM1.3 phosphorylation (Fig 2) that is required for REM1.3 function against PVX infection (Fig 3). We then asked whether different REM1.3 phospho-statuses might regulate its lateral organization at the PM and PD compartments in the absence of PVX. We examined the enrichment of REM1.3 YFP-tagged phosphomutants at the PD pit fields, previously calculated by the PD index (S1 Fig) and found that similarly to YFP-REM1.3, none of the phosphomutants appeared enriched at the pit field level (Fig 4A and 4C). The phosphodead mutant appeared statistically more excluded than YFP-REM1.3, whereas the phosphomimetic mutant displayed an increase of its PD index (Fig 4C), reminiscent of the REM1.3 localization phenotype under PVX infection (Fig 1A and 1B). Importantly, REM1.3 phosphomutants’ association with PD was directly correlated with callose content at PD (Fig 4B). These observations reinforced the hypothesis that REM1.3-mediated increase of callose levels at PD is associated with a dynamic and phosphorylation-dependent redistribution of REM1.3 to the PD surroundings.
We next used spt-PALM VAEM to characterize the localization and mobility behaviour of the EOS-REM1.3 phosphomutants in the PM plane. The analysis of reconstructed trajectories and corresponding super-resolved localization maps indicated slight modifications of lateral mobility behavior between the phosphomutants (Fig 4D and 4E). Quantification of the diffusion coefficient values (D) extracted for each individual molecule revealed that EOS-REM1.3AAA displayed a more immobile behavior than EOS-REM1.3DDD and EOS-REM1.3. Consistently, EOS-REM1.3DDD exhibited a higher mobility illustrated by higher diffusion coefficient and mean square displacement values (Fig 4D and 4E). Analysis of the supra-molecular organization of the phosphomutants by Voronoï tessellation (Fig 4F) firstly showed that all mutants displayed similar nanodomain size and localization density compared to EOS-REM1.3WT. Compared to EOS-REM1.3AAA, the EOS-REM1.3DDD nanodomains occupied a smaller area of the total PM and their density in the PM plane appeared slightly reduced (Fig 4F and 4G). A higher number of nanodomains were formed with the EOS-REM1.3AAA mutant. Hence, the phosphomimetic mutations favor a less confined and a more dynamic localization pattern of REM1.3 at the PM, reminiscent to the phenotype of EOS-REM1.3WT in the context of PVX infection (Fig 1C and 1D).
These results suggest that differential REM1.3 phosphorylation is involved in regulating REM1.3 mobility and PM domain organization and support the hypothesis that REM1.3 phosphorylation on S74, T86 and S91 reflects an ‘active form’ of the protein necessary for REM1.3-mediated defense signaling.
To gain more insights into the signaling processes leading to REM1.3 phosphorylation, we aimed to biochemically characterize the kinase(s) involved in the phosphorylation of REM1.3. Previous evidence suggested that the kinase(s) phosphorylating REM1.3 are membrane-associated (Fig 2) [38]. We therefore biochemically analyzed the localization of the kinase(s) phosphorylating REM1.3. Plant material from healthy and PVX-GFP-infected leaves was cell-fractionated to obtain crude extracts, soluble and microsomal fractions [59] to perform in vitro kinase assays on REM1.3N. Analysis confirmed a maximal kinase activity in purified microsomes (Figs 5A and 2A). Since a kinase in close proximity with its substrate would enhance reaction kinetics [60] and signal fidelity [61], and given that REM1.3 is enriched in detergent-resistant membranes (DRM) [8], we investigated whether the kinase activity towards 6His-REM1.3 is enriched in this biochemical fraction. We included “control PM” (C-PM) preparations, submitted to discontinuous sucrose gradients but in the absence of Triton-X100 treatments [62]. In vitro kinase assays on 6His-REM1.3N showed that the kinase activity in C-PM was 5 times inferior than in freshly purified PM not submitted to the sucrose gradient, suggesting that the kinase is not stable during the overnight purification procedure. Only half of the specific activity of the kinase was found in DRMs compared to the C-PM fraction, indicating that the kinase(s) phosphorylating REM1.3 is (are) only partially located in the DRM fraction (Fig 5B).
To gain more information concerning the biochemical characteristics of the kinase phosphorylating REM1.3, we analyzed its activity in the presence of known inhibitors. We firstly tested staurosporine, [63, 64] a general inhibitor that prevents ATP binding to kinases. We found an inhibition of REM1.3 phosphorylation starting at very low concentrations (30 nM) (S6A Fig). We further tested the effect of poly-L-lysine, described to stimulate the activity of the CK2 kinases and inhibit several CDPK kinases [65, 66]. No significant differences on REM1.3 phosphorylation levels were observed under increasing concentrations of poly-L-lysine (S6B Fig). The addition of the wide range of Ser/Thr phosphatases inhibitor β-glycerophosphate (BGP) [66] to the reaction mix did not alter the levels of phosphorylated 6His-REM1.3, indicating that the observed data was due to the activation of kinase activity by PVX rather than by inhibition of phosphatases (S6B Fig). Competition assays in the presence of cold AMP and GTP showed that only cold ATP even at 2 mM caused 20-fold depletion in [γ-33P] incorporation, suggesting that ATP is the major phosphoryl-donor for the kinase (S6B Fig). Addition in the reaction mix of 0,2 mM of EGTA, a chelator of Ca2+, strongly inhibited the kinase activity suggesting that the kinase(s) phosphorylating REM1.3 in healthy leaves is calcium sensitive (S6C Fig). Calcium is a conserved second messenger in signal transduction during biotic and abiotic stress. In plants, kinases harboring different calcium sensitivities can perceive calcium variations and translate them into downstream signaling activation [67, 68]. To determine whether the PVX-activated kinase phosphorylating REM1.3 is sensitive to calcium regulation, in vitro kinase assays from microsomes of healthy and PVX-infected N. benthamiana leaves were assayed in the presence of free calcium (Ca2+) concentrations ranging from 10 nM to 0,1 mM. Fig 5C shows that the kinase(s) displays a high sensitivity to calcium with an optimal activity in the presence of 10 μM of free Ca2+. At this concentration, a 5-fold increase of 6His-REM1.3N phosphorylation was observed in PVX-infected leaves (Fig 5C). These experiments allowed us to narrow-down the kinase(s) phosphorylating REM1.3 after PVX infection to the group of membrane-bound Ca2+-dependent protein kinases [67].
Plants possess three main families of calcium-regulated kinases: calmodulin-binding kinases (CBKs), calcineurin B-like-interacting protein kinases (CIPKs) and calcium-dependent protein kinases (CPKs) [67]. CPKs have the unique feature of calcium sensing and responding activities in one single polypeptide, best characterized in the model plant Arabidopsis [67]. Based on the measured calcium dose response (Fig 5C), we correlated the kinase phosphorylating REM1.3 in N. benthamiana with homologs of Arabidopsis subgroup II AtCPKs [69], and we aimed to capitalize on the knowledge of Arabidopsis CPKs to test REM1.3 phosphorylation. Among the characterized members of subgroup II AtCPKs, we selected the Arabidopsis AtCPK3 as a good candidate to test its putative role in REM1.3 phosphorylation, since previous proteomics studies in Arabidopsis have identified both AtCPK3 and AtREM1.3 as being enriched in PM, PD and DRM fractions [22, 70]. In addition, one study showed that AtREM1.3 from microsomal fractions is phosphorylated in vitro by AtCPK3 [71]. We therefore predicted that REM1.3 might share common functions with the evolutionarily conserved group 1b Arabidopsis REMs [32]. AtREM1.2 and AtREM1.3 are close homologs to REM1.3 and group 1 N. benthamiana REMs (NbREMs) in term of protein sequence [32, 36] and they conserved at least the S74 and S91 phosphorylation sites [39], [40, 57] (S5A Fig). Using super-resolution microscopy, Demir et al. showed that, when co-expressed in Arabidopsis leaves, REM1.3 and AtREM1.3 co-localized in the same PM-nanodomains [72]. Importantly, transient expression of AtREM1.2 and AtREM1.3 in N. benthamiana epidermal cells impaired PVX-GFP cell-to-cell movement, as REM1.3 does (S5B Fig), strengthening the hypothesis that the function of group 1 REMs might be conserved between homologs in different plant species [36].
We assayed the in vitro phosphorylation activity of the affinity-purified AtCPK3-GST towards the 6His-REM1.3, the 6His-REM1.3N and the 6His-REM1.3C, as well as the homologous substrate 6His-AtREM1.2. Similar to our previous results (S3B and S3C Fig), AtCPK3-GST could phosphorylate strongly both 6His-REM1.3 and 6His-REM1.3N, but not 6His-REM1.3C (Fig 5D). In accordance with the effect of AtREM1.2 in PVX-GFP propagation (S5B Fig), AtCPK3-GST can also phosphorylate 6His-AtREM1.2 (Fig 5E). Addition of Ca2+ is essential for a strong kinase activity as shown by both kinase auto-phosphorylation and trans-phosphorylation (Fig 5D and 5E). AtCPK3-GST specifically phosphorylated S74, T86 and S91 residues of REM1.3, since the phosphorylation was abolished in the phosphomimetic mutant 6His-REM1.3DDD (Fig 5F).
These results suggest that AtCPK3 is a good candidate for group 1b REM phosphorylation and further support that the S74, T86, and S91 are the phosphorylation sites of REM1.3 (Figs 3A and 5E).
CPKs harbor a variable N-terminal domain, a Ser/Thr kinase domain, an auto-inhibitory junction region and a regulatory calmodulin-like domain. The calmodulin-like domain contains four EF-hand binding motifs that determine the sensitivity of each kinase to calcium [73, 74]. To investigate the role of AtCPK3 in REM-dependent signalling, we generated AtCPK3 mutants presenting altered kinase activities. Deletion of the inhibitory junction region and the regulatory calmodulin-like domain in CPKs creates a constitutive active kinase while mutation of the aspartic acid residue in the catalytic center ‘DLK’ motif of the kinase domain to an alanine (D202A) creates a catalytically inactive or ‘dead’ kinase [67] (Fig 6A). We generated AtCPK3 full-length (AtCPK3), constitutive active (AtCPK3CA, residues 1–342) and kinase-dead (AtCPK3CAD202A) constructs for transient protein expression (Fig 6A). We evaluated their catalytic activities by expressing them transiently in Arabidopsis mesophyll protoplasts and performing immunoprecipitation coupled to kinase assays using 6His-REM1.3 and histone as a generic substrate [67]. Autoradiography confirmed that in vivo purified AtCPK3CA-HA could trans-phosphorylate both 6His-REM1.3 and histone without the addition of calcium, while the point mutation D202A drastically abolished kinase activity (S7 Fig).
We next examined the sub-cellular localization of both AtCPK3 and AtCPK3CA fused to YFP and found that both proteins disclosed a partial association with the PM, which was further confirmed by their presence, after cell fractionation, in the microsomal fraction at the expected molecular weight (Fig 6B) in good agreement with [71]. We further used AtCPK3CA to test the interaction with group 1b REMs. Bimolecular Fluorescence Complementation (BiFC) experiments showed that AtCPK3CA and REM1.3, REM1.3AAA and REM1.3DDD interact together at the level of the PM in planta. Importantly, we also confirmed the interaction of AtCPK3CA with homologous AtREM1.2 and AtREM1.3 (Fig 6C). REM1.3/REM1.3 interaction was used as a positive control, and AtCPK3CA /AtCPK3CA as a negative control.
We finally aimed to functionally characterize the AtCPK3- and REM1.3-mediated signaling in the context of PVX infection. Transient over-expression of AtCPK3-RFP alone induces a reduction of PVX-GFP infection foci suggesting that AtCPK3 is indeed important for antiviral responses in plant cells (Fig 6D). Expression of the constitutively-active AtCPK3CA-RFP had a stronger effect on PVX-GFP spreading and to a similar degree with the over-expression of REM1.3 alone (Fig 6D). AtCPK3's function towards PVX movement was observed to be mediated by its kinase activity, as the expression of the catalytically inactive mutant AtCPK3CAD202A had no effect on PVX-GFP propagation (Fig 6D).
This raised the question whether the effect of AtCPK3CA on PVX propagation was REM-dependent. To tackle this question, we stably transformed N. benthamiana plants with a hairpin construct, to induce post-transcriptional gene silencing, which resulted in lowering RNA and protein expression of group 1 endogenous NbREMs (S8A and S8B Fig). Consistent with previous studies [8], silencing of group 1 REM correlates with an increase of PVX-GFP cell-to-cell movement in inoculated leaves (S8C Fig). No difference was observed by ELISA when measuring PVX accumulation in systemic leaves (S8D Fig). Importantly, PVX assays demonstrated that AtCPK3CA ability to restrict PVX movement was impaired in two independent N. benthamiana lines underexpressing group 1 REM levels, (namely lines 1.4 and 10.2 with expression levels decreased respectively by 2 and 20 times) (Fig 6E), indicating that REMs might be the direct substrate of CPK3 in vivo.
Altogether, these data suggest that CPK3 and group 1 REMs are major regulators involved in signaling and antiviral defense at the PM level.
Protein phosphorylation is a ubiquitous and specific mechanism of cell communication [75]. The addition of a phosphate group on one or more critical residues of a given protein can induce important conformational changes that affect energetically favorable interactions and may lead to changes in its interacting network, localization, abundance and may influence the activity of protein signaling pools [76]. Although, since the initial discovery of REM1.3 in 1989, accumulating evidence suggests that the functions of REM proteins are regulated by protein phosphorylation [38–40]. The biological significance of this phosphorylation remained unclear to this date. REM proteins were among the first plant proteins described which supported the notion of PM sub-compartmentalization to functional protein-lipid nanodomains [8, 11, 77], also named membrane rafts [3, 4, 21]. In the present paper, we used REM1.3 and PVX as an experimental system to study the role of protein phosphorylation and membrane dynamics in the context of stress response.
Understanding how plants defend themselves against viruses remains a challenging field. The canonical plant immune response against viruses is mainly represented by the mechanism of RNA silencing [78, 79], while additional mechanisms of plant antiviral defense involve hormonal signaling, protein degradation, suppression of protein synthesis and metabolic regulation [51, 78, 80]. Antiviral defense presents similarities to the immune response against microbes [81–83]. Compelling evidence suggests that cell-surface as well as intracellular plant immune receptors recognize viral elicitors [55, 84–89]. An additional number of host cell components have been shown genetically to affect viral replication or cell-to-cell movement [8, 90], indicating that more sophisticated plant defense mechanisms against viruses may exist.
For instance, manipulation of REM levels in transgenic Solanaceae suggested that REM is as a positive regulator of defense against the PVX by affecting viral cell-to-cell movement [8, 14, 36]. We recently showed that REM1.3 does not interfere with the suppressor ability of PVX movement protein TGBp1, but specifically affects its gating ability [33]. Group 1 REMs could be a target for viruses (and other pathogens) to circumvent infection as illustrated by the case of Rice Stripe Virus that targets NbREM1 for degradation by 26S proteasome [91]. Nevertheless, in this study we show that REM1.3 protein levels are not altered during PVX infection (S2C and S2D Fig).
In this paper, we provide supporting mechanistic evidence that REM1.3 regulates the levels of callose accumulation at PD pit fields during PVX infection (Fig 1). Whether this function is mediated by a direct interaction with callose synthase/glucanase complexes remains however still unknown. Surprisingly, we found that REM1.3 is not dramatically recruited to PD pit fields, although its PD index is slightly increased after PVX infection (Fig 1). This suggests that association of a sub-fraction of the REM1.3 to the PD-PM region may be sufficient to increase callose accumulation, although we cannot rule out the possibility that REM1.3 may regulate PD permeability via a more indirect mechanism. The spt-PALM VAEM microscopy data supports an increase of protein mobility and redistribution to distinct domains during PVX infection (Fig 1). These findings indicate the existence of a mechanism that operates at specific REM1.3-associated PM nanodomains, capable of regulating PD permeability (Fig 1). The dynamic partitioning between PM nanodomains and PD pit fields needs to be further studied.
Since various studies have reported REM phosphorylation during plant-microbe interactions [16, 17, 39, 40], we set out to address which kinase phosphorylates REM and whether REM1.3 phosphorylation plays a role in REM-mediated anti-viral defense. Indeed, our experimental findings show that plant PVX sensing induces the activation of a membrane-bound calcium-dependent protein kinase that in turn phosphorylates REM1.3 (Fig 2, Fig 5). Importantly, we show that the kinase able to phosphorylate REM1.3 is activated specifically by the expression of two PVX proteins, namely CP and TGBp1. Deciphering the exact mechanisms allowing the molecular recognition of those PVX components will be a crucial step toward understanding REM-mediated anti-viral defense. Intriguingly, the finding that the presence of Agrobacterium also induces REM1.3 phosphorylation (Fig 2G) is in agreement with previous reports suggesting phosphorylation of REMs under bacterial infection [39, 40] and suggests that phosphorylation should be also a way to regulate -yet unknown functions- of REM1.3 in bacterial defense.
Genetic studies have established that different CPKs comprise critical plant signaling hubs by sensing and translating pathogen-induced changes of calcium concentrations [67, 68]. Biochemical characterization of the kinase phosphorylating 6His-REM1.3 showed that its strong sensitivity to calcium (Fig 5C) corresponds to homologs of phylogenetic subgroup II CPKs from Arabidopsis [67]. CPK3 is a prominent member of subgroup II, shown to function in stomatal ABA signaling [92], in salt stress response [71, 93] and in a defense response against an herbivore [94]. Interestingly, it was suggested that AtREM1.3 from taxonomical group 1 of REMs could be a substrate for AtCPK3 in response to salt stress [71]. Here we show that AtCPK3 can interact in vivo with group 1 REM (Fig 6C) and that AtCPK3 phosphorylates group 1 REM in an in vitro kinase assay (Fig 5D). Transient overexpression of AtCPK3 in N. benthamiana resulted in a reduction of PVX propagation in a REM-dependent manner, providing compelling evidence that CPK3 together with REM contribute to the plant antiviral immunity. This is the first report demonstrating the participation of CPKs in plant basal immunity against viruses.
Although [95] reports that there is no calcium signal during early recognition of PVX, the activation of CPKs by PVX supports the notion that calcium might be involved in some other late steps of plant-virus interaction like the control of intercellular connectivity. These changes in calcium concentrations in the cell are sensed by the CPKs and translated via the phosphorylation of REM and/or other unknown downstream components. In Nicotiana tabacum calmodulin isoforms are critical for the plant resistance against Tobacco Mosaic Virus and Cucumber Mosaic Virus, further illustrating the existence of virus-specific patterns of calcium signals [96, 97]. More work is needed to identify the CPK family members participating to the response and also the nature and specificity of those PVX-induced calcium changes.
AtCPK3 specifically phosphorylated REM1.3 at its N-terminal domain (residues 1–116), a domain displaying a mostly intrinsically disordered secondary structure (Figs 3A and 5). In silico analysis followed by a mutagenesis approach coupled with in vitro kinase assays revealed three major putative phosphorylation sites for REM1.3, namely S74, T86 and S91 on REM1.3. The in vitro phosphorylation of REM1.3 (Figs 3A and 5E) is almost totally lost when S74, T86 and S91 are mutated to non-phosphorylable residues, confirming these residues as major REM1.3 phosphorylation sites. Individual phospho-null mutations at those sites impaired REM1.3 ability to restrict PVX cell-to-cell movement to various extent (Fig 3D). The triple phospho-null mutant, YFP-REM1.3AAA totally obliterated REM1.3's capability to restrict PVX cell-to-cell movement (Fig 3D) and to regulate PD permeability (Fig 3E). Reciprocally, REM1.3 triple phosphomimetic mutant, RFP-REM1.3DDD appeared fully functional (Fig 3E and 3F). These results strongly support the functional involvement of single or combined phosphorylation in the N-terminal domain of S74, T86 and S91 to establish REM’s function in the context of PVX infection. This is in contrast with LjSYMREM1 from Lotus japonicus which was shown to be phosphorylated at its C-terminal domain in vitro by SYMRK [16]. Despite the fact that phosphorylation of REM proteins has been widely reported [16, 17, 39, 40, 57], this work firstly describes an associated role of REM-induced phosphorylation with its function.
Our finding that overexpression of AtREM1.2 and AtREM1.3 also restricts PVX-GFP cell-to-cell movement (S5B Fig) suggests that REM phosphorylation and its associated functions might be conserved for some REMs of taxonomic group 1b. In good agreement, AtREM1.2 and AtREM1.3 localize to the same PM nanodomains [72] and maintain conserved phosphorylation sites with REM1.3 (S5A Fig). By contrast, AtREM4.1 from subgroup 4, presenting a different N-terminal domain and different expected phosphorylation profile has an opposite effect against geminiviral propagation by promoting susceptibility to Beet curly top virus and Beet severe curly top virus [17, 57]. This further argues that REMs might be phosphorylated by diverse families of kinases in order to respond to different stimuli [57].
Overexpression of REM1.3 restricts TMV propagation (S4A Fig), and additionally modulates the movement proteins from different virus genera [33, 91]. These findings suggest that the initial hypothesis that REM1.3 causes the sequestration of the PVX virions at the PD [8] might not hold true, but rather that REM1.3 might have a more general role in plant stress and PD regulation (Figs 1 and 3). Interestingly, REM1.3 promotes susceptibility to Phytophthora infestans in N. benthamiana and localizes exclusively to the PM and the extrahaustorial membrane around non-callosic haustoria [42]. The exact role of REM1.3 as a common regulator of different signaling pathways and its role in PD permeability regulation remains to be determined.
It has been speculated that phosphorylation in intrinsically disorder regions of proteins may act as a molecular switch and confer potential protein-protein interaction plasticity [76, 98]. The intrinsically disordered REM1.3 N-terminal domain exhibits the most sequence variability in REM proteins, presumably conferring signaling specificity [32, 57]. Phosphorylation of AtREM1.3's N-terminal domain could stabilize coil-coiled-associated protein trimerization and protein-protein interactions [57]. Phosphorylated REM1.3 seems to be further targeted to PD-PM to trigger callose deposition. In good agreement, we found that the mobility in the PM of REM1.3 changed depending on its phospho-status (Fig 4). The triple phosphomimetic mutant exhibited a less confined and more mobile behavior at the PM, reminiscent of the WT protein in the context of PVX infection (Fig 4D). Similarly to the role of 14.3.3 proteins in plants [99], REM1.3 could act as a scaffolding protein, interacting with multiple members of a signaling pathway and tethering them into complexes to specific areas of the membrane. Hence, REM1.3 phosphorylation could act as a regulatory switch of protein conformations that would modulate REM1.3 specific interaction patterns and transient signalosomes at the PM. The triple phosphomimetic REM mutant might reflect a ‘functionally active’ form that constitutes REM-guided signalosomes against PVX-infection at the PM and should be exploited in future studies. The study of the phosphorylation-dependent interactions of REM1.3 (and related phosphocode) in regard to the modulation of REM1.3 PM dynamics and molecular function is the topic for future studies.
Nicotiana benthamiana plants were cultivated in controlled conditions (16 h photoperiod, 25 °C). Proteins were transiently expressed via Agrobacterium tumefaciens-mediated transformation for virus and PD functional assays as in [14, 33] or for the localization experiments as described in the appendix. For subcellular localization studies and protein extractions, plants were analyzed 2 or at 4 days after inoculation (DAI) in the phosphorylation assays. Imaging for PVX-GFP spreading assays and plasmodesmata GFP-diffusion experiments were done at 5 DAI. PVX inoculation for test ELISA was performed at 4-week-old N. benthamiana plants. Details on molecular cloning and protein work, transgenic lines generation are described in the Appendix.
All vectors constructs were generated using classical Gateway cloning strategies (www.lifetechnologies.com), pDONR211 and pDONR207 as entry vectors, and pK7WGY2, pK7YWG2, pK7WGR2, pK7RWG2, and pGWB14 and pGWB15 as destination vectors [100]. The REM1.31–116, REM1.3117–198 and REM1.3 single S74A, T86A and S91A and triple S74/T86/S91AAA and S74/T86/S91DDD mutants were synthesized in a pUC57 vector (including the AttB sites) by Genscript (http://www.genscript.com/) or GENEWIZ (http://www.genewiz.com/) and next cloned to Gateway destination vectors. AtCPK3D202A mutant was generated by site-directed mutagenesis as previously described [101] with minor modifications. For BiFC experiments, the genes of interest were cloned into pSITE-BIFC- C1nec, -C1cec, -N1nen, and–N1cen destination vectors [102]. To map the dynamics of single molecules with sptPALM, REM1.3 and phosphomutants were cloned in fusion with EOS in the gateway compatible vector pUBN-Dest::EOS [103]. EOS protein has been widely use for single molecule localization microscopy in mammals, bacteria, and plant cells. It corresponds to the name of a fluorescent protein from the stony coral Lobophyllia hemprichii which peculiarity resides in its photoconvertability. The energy of UV light can break the core polypeptidic chain of EOS fluorescent protein inducing changes in EOS spectral fluorescence properties. Due to the stochasticity of EOS photoconversion at low UV radiation (space and number of events/sec can be controlled by modulating UV laser power), single molecules can be converted, localized and tracked.
Leaf discs were cut from N. benthamiana leaves, transferred on petri plates containing culture medium (complete Murashige and Skoog medium (MS) supplemented with 30g/L saccharose, pH 5,8; phytoagar HP696 (Kalys) 5,5 g/L and the hormones: AIA 0,1 mg/L, BAP 2 mg/L) and incubated for 48 h in the growth room (16 h photoperiod, 30 μmol photons.m2.s-1, 23 °C). For the transformation, the N. benthamiana plants disk leaves were incubated with the Agrobacterium cultures (GV3101 strain) carrying the plasmid of interest for 20 min. The leaf samples were next placed on plates with the complete medium previously described. 48 hours later, the leaf fragments were washed 3 times with sterile water (with 0,1% Tween20). The leaf fragments were next washed with MS complete medium supplemented with Timentin (300 μg/mL). The leaves were next placed on plates with regeneration medium (MS culture medium, as previously described, supplemented with 300 mg/L of timentin and 150 mg/l of kanamycin). The plates were next incubated in the growth room. The explants were transferred to fresh regeneration medium with a maximum periodicity of 7 days until the development of callus. The regenerated seedlings were transferred to a rooting medium (MS, sucrose 30 g/L, phytoagar 5,5 g/L, timentin 200 mg/L, kanamycin 150 mg/L). The regenerated plants (T0) were transferred to the greenhouse for growth and self-fertilization. Homozygous T2 lines carrying a single transgene insertion were selected by segregation analysis on selective Kanamycin media and used for physiological studies and phenotypic characterization. The expression of the GFP-REM1.3 or silencing levels of endogenous NbREMs was controlled by cytological, biochemical and expression analysis. Cytological analysis of the GFP-REM1.3 expression in all leaf cells was performed to avoid chimeric expression, see S2A and S2B Fig.
Four-week-old N. benthamiana greenhouse plants grown at 22–24 °C were used for Agrobacterium tumefaciens-mediated transient expression. A. tumefaciens were pre-cultured at 28 °C overnight and used as inoculum for culture at initial OD600nm of 0.15 in pre-warmed media. Cultures were grown until OD600nm reached 0.6 to 0.8 values (3–5 h). Cultures were then centrifuged at 3,200 g for 5 min, pellet were washed twice, using water to the desired OD600nm. Bacterial suspensions at OD600nm of 0.2 and 0.1 were used for subcellular localization and Spt-PALM experiments, respectively. The bacterial suspensions were inoculated using a 1-mL syringe without a needle by gentle pressure through a <1mm-hole punched on the lower epidermal surface [104]. Transformed plants were incubated under normal growth conditions for 2 days at 22–24 °C. Transformed N. benthamiana leaves were analyzed 2–5 DAI depending on the experiment.
PVX-GFP cell-to-cell movement experiments were performed as previously described [14, 36], with minor modifications. Briefly, A. tumefaciens strain GV3101 carrying the constructs tested were infiltrated at a final optical density at 600 nm (OD600nm) = 0.2 together with the same strain carrying the plasmid pGr208, which expresses the PVX-GFP complementary DNA harboring GFP placed under the control of a Coat protein promoter, as well as the helper plasmid pSoup [105] at final OD600nm of 0.001. Viral spreading of PVX-GFP was visualized by epifluorescence microscopy (using GFP long pass filter on a Nikon Eclipse E800 with x4 objective coupled to a Coolsnap HQ2 camera) at 5 DAI and the area of at least 30 of PVX-GFP infection foci was measured using Fiji software (http://www.fiji.sc/) via a homemade macro or ImageJ. The expression levels of transiently expressed constructs were confirmed by Western blot. ELISA tests in systemic N. benthamiana leaves were performed similarly to [8] to follow the global virus accumulation. Briefly, GFP-REM1.3 or hpREM plants were mechanically inoculated with PVX, and viral accumulation in systemically invaded leaves (at 3 nodes above the inoculated leaf) was evaluated at 10 or 14 DAI with a specific anti-PVX coat protein antibody (Sediag). Five plants per line for GFP-REM1.3 and 8 for hpREM plants were tested per experiment. GFP diffusion at PD experiments was performed as previously described [33]. All the experiments were repeated at least three times.
Live imaging was performed using a Leica SP5 confocal laser scanning microscopy system (Leica, Wetzlar, Germany) equipped with Argon, DPSS and He-Ne lasers and hybrid detectors. N. benthamiana leaf samples were gently transferred between a glass slide and a cover slip in a drop of water. YFP and mCitrine (cYFP) fluorescence were observed with similar settings (i.e., excitation wavelengths of 488 nm and emission wavelengths of 490 to 550 nm). In order to obtain quantitative data, experiments were performed using strictly identical confocal acquisition parameters (e.g. laser power, gain, zoom factor, resolution, and emission wavelengths reception), with detector settings optimized for low background and no pixel saturation. Pseudo-colored images were obtained using the “Red hot” look-up-table (LUT) of Fiji software (http://www.fiji.sc/). All quantifications were performed for at least 10 cells, at least two plants by condition with at least 3 independent replicates. BiFC images were taken 2 DAI by confocal microscopy (Zeiss LSM 880). Quantification of fluorescent intensities was performed by ImageJ, as described in [36].
N. benthamiana epidermal cells were imaged at room temperature (RT). Samples of leaves of 2-week-old plants expressing EOS constructs were mounted between a glass slide and a cover slip in a drop of water to avoid dehydration. Acquisitions were done on an inverted motorized microscope Nikon Ti Eclipse (Nikon France S.A.S., Champigny-sur-Marne, France) equipped with a 100× oil-immersion PL-APO objective (NA = 1.49), a TIRF arm, a Perfect Focus System (PFS), allowing long acquisition in oblique illumination mode, and a sensitive Evolve EMCCD camera (Photometrics, Tucson, USA). Images acquisitions and processing were done as previously described [45].
Single molecule fluorescent spots were localized in each image frame and tracked over time using image processing techniques such as a combination of wavelet segmentation [106] and simulated annealing algorithms [107]. The software package used to extract quantitative data on protein localization and dynamics is custom written as a plug-in running within the MetaMorph software environment. This plugin is now property of Molecular devices company (https://www.moleculardevices.com/sites/default/files/en/assets/product-brochures/dd/img/metamorph-super-resolution-software.pdf).
Single molecule localization organization analysis, Log(δ1/δ) correspond to the ratio between the local molecule density to overall molecule density at the PM. After correction for artefacts due to multiple single-molecule localization (described in [36] and now presented in materiel and methods section), we computed potential nanodomain by applying a threshold δ1i>2δN, where δN is the average localization density at PM level and δ1i is the density in presumed protein-forming nanodomain, with a minimal area of 32 nm2 and with at least 5 localizations per nanodomain.
SR-Tesseler software was used to produce Voronoï diagrams, and subsequently quantify molecule organization parameters as previously recommended [50]. Taking in account fluorophore photophysical parameters, localization accuracy and the first rank of local density of fluorescent molecules, correction for multiple detections occurring in a vicinity of space (w) and blinking tolerance time interval (t) are identified as the same molecule, merged together and replaced by a new detection at a location corresponding to their barycentre. Because first rank of local density of fluorescent molecules was below 0.5 mol/mm2 (c.a ranking from 0.1 to 0.3 mol/mm2), we used a fixed search radius w of 48 nm as recommended [50]. To determine the correct time interval t, the photophysics of the fluorophore namely the off-time, number of blinks per molecule and on-time distributions are computed for each cell. For example, for a dataset composed of 618,502 localizations, the average number of blinks per molecule was 1.42, and the number of molecules after cleaning was 315,929. As a control, the number of emission bursts (439,331), counted with t = 0, divided by the average number of blinks per molecule (1.42) was only 2.15% different. After correction for artefacts due to multiple single-molecule localization, we computed potential cluster using a threshold d1i>2dN, where dN is the average localization density at PM level and d1i is the density in presumed protein-forming nanocluster, with a minimal area of 32 nm2 and with at least five localization by cluster.
Over the two independent experiments 54 446 single molecule trajectories have been observed (34 740 Mock / 19 706 PVX). We then computed single molecule mobility behavior (Diffusion coefficient and Mean square displacement) using trajectories of at least 8 time points (tracked for at least 0.16 s; representing 19495 trajectories in total, 12073 for Mock condition and 7422 for PVX condition).
Prediction of putative phosphorylation sites was performed by Diphos, DEPP and NETPHOS coupled with published data. Disordered domains were performed by pDONR VL XT.
6His-REM1.3 and mutant recombinant proteins were purified from bacteria using fast flow chelating sepharose resin (Amersham) according to manufacturer’s instructions and as in [14]. For the in vitro REM1.3 phosphorylation assays about 2 μg of total plant extracts were incubated with 1 μg of affinity-purified 6His:REM1.3 protein variants for 10 minutes at room temperature and in a phosphorylation buffer (Tris-HCl 30mM, EDTA 5mM, MgCl2 15mM, DTT 1mM, Na3VO4 2,5 mM, NaF 10 mM and 10 μCi/reaction ATP [γ-33P]- (3000Ci/mmol, Perkinelmer). The buffer contained also 10–5 M of free Ca2+ which allows the detection of 6His-REM1.3 phosphorylation also in mock conditions. Gradual concentrations of free Ca2+ as in [108] were added for Fig 5C. Reactions were performed for 15 minutes in a volume of 25 μl. The reactions were stopped by the addition of 15 μl of 6x loading buffer. Proteins were separated by SDS-PAGE and phosphorylation status of REM1.3 was analysed by autoradiography using a phosphor-Imager and quantified by ImageQuant TL program.
CPK3-HA was transiently expressed in mesophyll protoplasts and immunopurified with anti-HA antibodies as performed in [109] while CPK3-GST recombinant protein was purified from bacterial extracts as reported in [69]. For in vitro kinase assays, the tagged CPK was incubated with 0.5–1 μg histone or 6His-REM1.3 proteins in the following kinase reaction buffer (20 mM Tris HCl pH 7.5, 10 mM MgCl2, 1 mM DTT, 50 μM cold ATP, ATP [γ-33P] 2 μCi per reaction, 1 mM CaCl2 or 5 mM EGTA) in a volume of 15 μL for 30 min at RT. The reaction was stopped with 5 μL 4X Laemmli buffer, then samples were heated at 95 °C for 3 min. Proteins samples were separated by SDS-PAGE on 12% acrylamide gel. After migration, the gel was dried before exposing against a phosphorScreen to reveal radioactivity on a Storm Imaging system (GE Heathcare). The gel was then rehydrated for Coomassie staining.
SDS-PAGE, Western Blot analysis, protein extractions and recombinant protein purification were performed in E. coli as in [14]. Cell fractionation and extractions followed the established protocol from [59] and [62]. Anti-REM antibodies were previously described in [8].
All relevant data are within the paper and its Supporting Information files are available from Arabidopsis Genome Initiative (https://www.arabidopsis.org/index.jsp), and GenBank/EMBL (https://www.ncbi.nlm.nih.gov/genbank/) databases under the accession numbers: StREM1.3 (NP_001274989), AtREM1.2 (At3g61260), AtREM1.3 (At2g45820), AtCPK3 (At4g23650).
|
10.1371/journal.pcbi.1005261 | Functionality and Robustness of Injured Connectomic Dynamics in C. elegans: Linking Behavioral Deficits to Neural Circuit Damage | Using a model for the dynamics of the full somatic nervous system of the nematode C. elegans, we address how biological network architectures and their functionality are degraded in the presence of focal axonal swellings (FAS) arising from neurodegenerative disease and/or traumatic brain injury. Using biophysically measured FAS distributions and swelling sizes, we are able to simulate the effects of injuries on the neural dynamics of C. elegans, showing how damaging the network degrades its low-dimensional dynamical responses. We visualize these injured neural dynamics by mapping them onto the worm’s low-dimensional postures, i.e. eigenworm modes. We show that a diversity of functional deficits arise from the same level of injury on a connectomic network. Functional deficits are quantified using a statistical shape analysis, a procrustes analysis, for deformations of the limit cycles that characterize key behaviors such as forward crawling. This procrustes metric carries information on the functional outcome of injuries in the model. Furthermore, we apply classification trees to relate injury structure to the behavioral outcome. This makes testable predictions for the structure of an injury given a defined functional deficit. More critically, this study demonstrates the potential role of computational simulation studies in understanding how neuronal networks process biological signals, and how this processing is impacted by network injury.
| Neurodegenerative diseases such as Alzheimer’s disease, Creutzfeldt-Jakob’s disease, HIV dementia, Multiple Sclerosis and Parkinson’s disease are leading causes of cognitive impairment and death worldwide. Similarly, traumatic brain injury, the signature injury of the Iraq and Afghanistan wars, affects an estimated 57 million people. All of these conditions are characterized by the presence of focal axonal swellings (FAS) throughout the brain. On a network level, however, the effects of FAS remain unexplored. With the emergence of models which simulate an organism’s full neuronal network, we are poised to address how neuronal network performance is degraded by FAS-related damage. Using a model for the full-brain dynamics of the nematode Caenorhabditis elegans, we are able to explore the loss of network functionality as a function of increased neuronal swelling. The relatively small neuronal network generates a limited and tractable set of functional behaviors, and we develop metrics which characterize how these behaviors are impaired by network injuries. These metrics quantify the severity of TBI and/or neurodegenerative disease, and could potentially be used to construct diagnostic tools capable of identifying various cognitive deficits. Additionally, we apply classification trees to our results to make predictions about the structure of an injury from specific cognitive deficits.
| Understanding networked and dynamic systems is of growing importance across the engineering, physical and biological sciences. Such systems are often composed of a diverse set of dynamic elements whose connectivity are prescribed by sparse and/or dense connections that are local and/or long-range in nature. Indeed, for many systems of interest, the diversity in connectivity and dynamics make it extremely challenging to characterize dynamics on a macroscopic network level.
Of great interest in biological settings is the fact that such complex networks often produce robust and low-dimensional functional responses to dynamic inputs. Indeed, the structure of their large connectivity graph can determine how the system operates as a whole [1, 2]. Neuronal networks, in particular, may encode key behavioral responses with low-dimensional patterns of activity, or population codes, as they generate functionality [3–8].
Unfortunately, all biological networks are susceptible to pathological and/or traumatic events that might compromise their performance. In neuronal settings, this may be induced by neurodegenerative diseases [9–11], concussions, traumatic brain injuries (TBI) [12–14] or aging. In this work, we extend a computational model to investigate behavioral impairments in the nematode C. elegans when the underlying neuronal network is damaged. Specifically, we consider how the low-dimensional population codes are compromised under the impact of an injury. Characterizing the resulting cognitive and behavioral deficits is a critical step in understanding the role of network architecture in producing robust functionality.
A hallmark feature of damaged neuronal networks is the extensive presence of Focal Axonal Swellings (FAS). FAS has been implicated in cognitive deficits arising from TBI and a variety of leading neurological disorders and neurodegenerative diseases. For instance, FAS is extensively observed in Alzheimer’s disease [10, 11], Creutzfeldt-Jakob’s disease [15], HIV dementia [16], Multiple Sclerosis [17, 18] and Parkinson’s disease [19]. Most concussions and traumatic brain injuries also lead to FAS or other morphological changes in axons [20–25]. Such dramatic changes in axon geometry may disrupt axonal transport [26, 27], and can potentially hinder the information encoded in neural spike train activity [28–30]. Injured axons thus provide an important diagnostic marker for the overwhelming variety of cognitive and behavioral deficits [9, 28, 31], in animals and humans [23, 32–34].
The massive size of human neuronal networks and their complex activity patterns make it difficult to directly relate neuronal network damage to specific behavioral deficits. C. elegans, in contrast, has only 302 neurons, and its stereotyped connectivity (i.e. the worm’s “Connectome”) is known [35]. This relatively small neuronal network generates a limited and tractable set of functional behaviors (see Table 1 of [36]), with much of its locomotion/crawling behavior approximately confined to five observable motor states related to forward and backward crawling, omega turns, head sweeps and brief pause states. Furthermore, these behaviors are well described as a superposition of only a few principal component body-shape modes [37]. The combination of a fully-resolved neuronal network and a tractable low-dimensional output space makes C. elegans an ideal model organism for studying the impact of network damage on behavioral deficits. Indeed, it is the only such neuronal network model currently available allowing for such a direct translational study of network damage (injury) to behavioral responses.
More precisely, computational models of C. elegans nervous system dynamics for the full or partial connectome successfully generate motorneuron outputs that can be related to behavior [38], allowing for interpretable outputs even without accounting for muscular, mechanical or environmental factors, e.g. [39]. We consider the model in [39], which applies a single-compartment membrane model to the full somatic connectome; neurons are approximated as passive linear units connected by linear gap junctions and nonlinear chemical synapses. Synaptic activation depends sigmoidally upon pre-synaptic voltage in equilibrium, and approaches this equilibrium value linearly in time. All neurons are approximated as identical, with order-of-magnitude parameter assignments, except for their connectivity data.
Fig 1(a) demonstrates a simulation of the putative forward crawling behavior identified in [39] within this model of C. elegans neural dynamics along with its projection onto principal component body-shape modes [37]. In this perspective, we understand forward crawling as corresponding to a limit cycle (i.e. a closed periodic trajectory) in the principal component space of simulated neural recordings. Extending this framework to damaged networks as in Fig 1(c) allow us to explore how axonal pathologies lead to impaired functionality and behavioral deficits. Even in our idealized injury simulations, the network’s impaired activity displayed significant variability. This highlights one of the most challenging aspects of the field: the need for effective metrics to distinguish different types of behavioral deficits. We propose such a criterium by using techniques borrowed from statistical shape analysis to quantify distortions in the main features of dynamical activity. This metric is shown to be related to the functional outcome of an injury. We further apply classification trees to our results to relate functional deficits to specific patterns of FAS. This leads to experimentally-testable predictions about the effects of neuronal network-damage to the crawling motion of C. elegans and potentially new avenues for clinical diagnostics. Indeed, our studies show that network damage leads to a diversity of dynamical/behavioral deficits.
We investigate how network distributed FAS as illustrated in Fig 1(c) may affect its ability to generate desired responses to an input. Network features associated with behavioral outcomes are best understood in model organisms such as the C. elegans since it has a limited repertoire of functional responses that include forward and backward crawling, omega turns, head sweeps and brief pause states. Our focus in these studies will be on the behavior of forward crawling since a variety of experimental ablation studies have identified key neurons associated this functionality. For instance, stimulation of PLM neurons excites densely-connected interneurons, which in turn, activate motorneurons responsible for forward body motion [40]. Experimentally, optogenetic stimulation of the PLM neurons directly induces a forward motion response [41, 42].
Details of the underlying neurocircuitry were found by a series of ablation studies, where the functional role of a neuron is evaluated by disconnecting it from the network and observing behavioral deficits [39, 43]. The coordinated body motion of a crawling worm is well documented in videos and its postural dynamics were revealed by principal component analysis to consist of only a few dominant modes [37]. Specifically, the sinusoidal body-shape undulations which describe the worm’s forward motion is well-described by circular trajectories (limit cycles) on the phase-space of its first two principal components. An analogous mathematical form is present in the collective motorneuron dynamics following PLM stimulation [39].
This commonality suggests that observed behaviors do retain fundamental signatures of the underlying network dynamics. We show such a trajectory for (simulated) motorneuron responses to PLM excitation in Fig 1(a). This low-dimensional representation captures 99.3% of the total energy of the system, and can be artificially mapped to crawling body-shape modes. Although this mapping is still far from a mechanistic description of the worm’s coordinated body movement, we believe it captures important aspects of the crawling behavior. See the Methods section for details. Importantly, functional deficits of the C. elegans dynamics are understood as excursions/perturbations from the ideal limit cycle trajectory. Damaged networks will be shown to fail to produce the low-dimensional output codes necessary for generating the optimal forward crawling limit cycle.
The robustness of the dynamical signatures (population codes) associated with behavior are investigated in injured neuronal networks. Our injury statistics and FAS models are drawn from state-of-the-art biophysical experiments and observations of the distribution and size of FAS. Fig 2 shows prototypical FAS injuries from stretching [26] and TBI in the optic nerve of mice [25]. Fig 2(d) shows a histogram of the probability of injury and size of the FAS. These are used in our computational model [39].
In a simulated injury, we assign to each affected neuron an axonal swelling from the distribution in Fig 1(b). Values are scaled by an (overall) injury intensity parameter μ, such that
1 + μ ∝ E swollen axon area healthy axon area (1)
Fig 1(c) exemplifies different injury settings: μ = 0 reproduces the original (uninjured) network, and lower/higher values of μ correspond to mild/severe injuries. The presence of axonal swellings ultimately distorts the forward-motion limit cycle dynamics. Fig 3 shows dynamical anomalies for different connectome injuries. Notice how they induce qualitatively different changes to the closed orbit regarding location, size and shape. Fig 3(c) reproduces the specific simulated ablations from [39], leading again to different dynamical effects.
A much larger ensemble of simulations (1,447 randomly-chosen injuries, as well as the code necessary to generate more) and their corresponding effects to fundamental low-dimensional structures are included in the Supporting Materials. Increasing values of μ typically shrink and shift the limit cycles within the plane. In all simulations, there was always a sufficiently high injury level in which
μ * = {injured limit cycle collapses into a stable fixed point } (2)
This occurs for instance, in Fig 3(b) when μ = 3.80. Recent blast injury studies on C. elegans show that many of the nematodes display temporary paralysis before recovering to crawling behaviors [45]. We would suggest that during the peak of the FAS, the injury levels on many of the nematodes are above μ*, thus leading to a collapse of a limit cycle to a fixed point where no motion is possible, i.e. it is in a paralyzed state.
Despite their common statistical distribution, randomly drawn injuries induce qualitatively different changes in the shape of the limit cycle. Additional distorted sets are shown in the rows of Fig 4 (along with 1,447 random-injury simulation sets in the Supporting Materials). Thus, random injuries of equitable strength can lead to significantly different behavioral deficits. Importantly, the deformation of the two-dimensional limit cycle can be used to characterize such functional differences. To distinguish dynamical signatures of potentially different functional deficits, we evaluate the Procrustes Distance (PD) between healthy and injured limit cycles. The PD is an important tool from statistical shape analysis to measure the similarity between two shapes after discounting effects due to translation, uniform scaling, or rotation. Fig 4 depicts PD values for pairs of healthy/injured limit cycles as a function of injury level μ. All curves are plotted until the injured limit cycle collapses into a fixed point (μ = μ*), and the colored dots in the rightmost plots correspond to the same-colored limit cycles on the left plots.
Recent experimental work which induced mild TBI in C. elegans found that increasing the number of shock waves to which the worm was exposed reduced the worm’s average speed and, in many cases, led to temporary paralysis [45]. The results of our simulations can be compared to these results:
In Fig 5 we plot the location of the fixed points into which limit cycles collapse (the “endpoints”, occurring at injury level μ = μ*). We consider the following question: does the location of this endpoint (and thus the behavioral outcome of the injury) relate to the PD curve, and does it relate to the structure of the injury itself? Towards this end, we construct two simple classes of behavioral outcomes: endpoints which end in either the “upper” or “lower” part of the distribution (for which we label the endpoints as red and green, respectively).
Panel (b) of Fig 5 shows the average PD curve for the two classes. They are qualitatively different: the average PD curve of “upper” endpoints is smoothly rising, whereas the average PD curve of “lower” endpoints has an extended declining region. Shown also are the average scaling factor and translation distance of the distorted cycles. Unlike the average PD curves, these change monotonically and are not distinct between classes. This suggests that the shape of the PD curve carries information about the functional outcome of the injury. We quantify this by fitting a classification tree to predict the endpoint class from the shape of the PD curve: this was found to predict endpoint class with a cross-validation error of 22.0%. By comparison, randomly shuffling the labels leads to nearly double the cross-validation error, with an average of (44.6 ± 1.4)%.
Of even greater interest is any possible relationship between injury structure and behavioral output which could, given a specific pattern of distorted dynamics, make predictions about the class of neural injury. To this end, we fit a classification tree to predict the endpoint class from the injury. Fig 6 shows a classification tree which predicts endpoint class with a cross-validation error of only 14.6%. This is much less than the error from a random class, suggesting that we can meaningfully relate the structure of a specific injury to a specific behavioral outcome. Classification trees provide a highly interpretable and predictive method for making this connection, and make specific experimental predictions for the injuries corresponding to functional deficits.
The dynamic model for the C. elegans connectome simulates its neuronal responses to stimuli with a number of simplifications aimed at keeping the number of parameters at a minimum: we use a fairly standard and simple single-compartment membrane equation, and treat all neurons as identical save for their connectivity. Many neurons in the network are nearly isopotential [46, 47], and it is a common and reasonable simplification to model neurons via single-compartment membrane equations, with membrane voltages as the state variables for each neuron. Given this, Wicks et al. constructed a single-compartment membrane model for neuron dynamics [48], which we later extended to incorporate connection data for the full somatic connectome [39]. We assume that the membrane voltage dynamics of neuron i is governed by:
C V i ˙ = - G c ( V i - E c e l l ) - I i G a p ( V → ) - I i S y n ( V → ) + I i E x t (3)
The parameter C represents the whole-cell membrane capacitance, Gc the membrane leakage conductance and Ecell the leakage potential of neuron i. The external input current is given by I i E x t. Note that this is, essentially, a fairly standard single-compartment membrane equation [49], and its governing equations are formally identical to that used by Wicks et al. [48] except for our use of the full somatic connectome, our simplifying parameter assumptions, and minor differences in the treatment of synaptic dynamics taken from [50].
In all simulations within this paper, we set I i E x t to be constant for the PLM neuron pair and zero for all other neurons. This assures that densely connected interneurons will stimulate the motorneuron subcircuits responsible for forward crawling behavior. Neural interaction via gap junctions and synapses are modeled by the input currents I i G a p ( V → ) (gap) and I i S y n ( V → ) (synaptic). Their equations are given by:
I i G a p = ∑ j G i j g ( V i - V j ) (4) I i S y n = ∑ j G i j s s j ( V i - E j ) (5)
We treat gap junctions between neurons i and j as ohmic resistances with total conductivity G i j g. We assume that I i S y n is also modulated by a synaptic activity variable si, which represents the conductivity of synapses from neuron i as a fraction of their maximum conductivity. This is governed by:
s i ˙ = a r ϕ ( v i ; β , V t h ) ( 1 - s i ) - a d s i (6)
Here ar and ad correspond to rise and decay time, and ϕ is the sigmoid function ϕ(vi; β, Vth) = 1/(1 + exp(−β(Vi − Vth))). This form of sigmoidal activation is taken from [50]. Note that it can be shown (by setting s ˙ = 0) that, as in [48], the equilibrium value of si depends sigmoidally upon Vi.
We keep all parameter values from [39] (see Table 1. The Connectome data, consisting of the number of gap junctions N i j g and number of synaptic connections N i j s, are taken from Varshney et al. [35] (as available on WormAtlas [51]). As in that study, we consider only the 279 somatic neurons which make synaptic connections (excluding 20 pharyngeal neurons, and 3 neurons which make no synaptic connections).
Each individual synapse and gap junction is assigned an equal conductivity of g = 100pS (such that G i j g = g · N i j g and G i j s = g · N i j s). The values of cell membrane conductance and capacitance are affected by injuries, but in the uninjured case are set as equal for all neurons with values of Gc = 10pS and C = 1pF. Note that in uninjured simulations, all neurons are modeled as identical except for their connectivity and the assignment of them as excitatory or inhibitory (where Ej will have one of two values corresponding to these classes).
The model is valuable because it generates a low-dimensional neural proxy for behavioral responses. Specifically, constant stimulation of the tail-touch mechanosensory pair PLM creates a two-mode oscillatory limit cycle in the forward motion motorneurons [39]. This same dynamical signature was revealed in video analysis of the body-shape of the crawling worm [37]. Thus the model is consistent with the observed biophysics. Specifically, we calculate this plane by first simulating the forward-motion motorneuron membrane voltages (class DB,VB,DD,VD) in response to a PLM Input of IPLML, IPLMR = 2 × 104 Arb. Units for the uninjured model. We take time snapshots these membrane voltages V → M ( t ), collect them into a matrix V, and take that matrix’s singular value decomposition. That is:
V = [ V → M ( t 0 ) , V → M ( t 1 ) … ] = P · Σ · Q T (7)
where P and Q are unitary and Σ is diagonal. The columns of P are the principal orthogonal modes. As in [39], the first two of these modes (the first two columns of P) almost entirely capture the dynamics of the system within this subspace under constant PLM stimulation. Projection of the full-system dynamics onto this plane consists of projecting the system’s motorneuron dynamics onto these modes.
Note that the single-compartment model which we employ ignores the spatial extent of neurons and specific location of each connection. Our simplified injury model therefore must treat injury as a whole-cell effect. Focal Axonal Swellings (FAS) increase the volume of an axon, which in turn, should alter the cell’s capacitance and leakage conductance within our model. If we approximate a neuron by a single cable of length l and constant cross-section a, we may assume that the circuit parameters will scale with the axonal volume, i.e.,
C ∝ a · l (8a) G c ∝ a · l (8b)
When an axon swells, its healthy cross-sectional area aH will increase to some swollen value ai > aH. Thus we assume that the healthy values for capacitance C and conductance Gc will also change according to
C i = C · ( a i / a H ) = C · ( 1 + μ · m i ) (9a) G i c = G c · ( a i / a H ) = G c · ( 1 + μ · m i ) (9b)
We define the individual damage mi to neuron i as proportional to the relative excess area from swelling, i.e., mi ∝ (ai − aH)/aH. Values of mi are computed from the experimentally derived distributions in Fig 2. Specifically, we construct FAS from the axonal swelling data of Wang et al. [25], which used confocal microscopy to measure injury-induced swellings in the optic nerve of Thy1-YFP-16 mice. Taken together, these define an “injury vector” m →, which we then normalize to | | m → | | 2 = 1. After normalizing, the injury vector is then scaled by a global injury intensity defined as follows:
μ = ⟨ a i / a H ⟩ - 1 ⟨ m i ⟩ (10)
Mild traumatic brain injuries yield small values of μ indicating that the average area of swollen axons is small. Severe brain injuries yield high values of μ, indicating that large swellings are more common. We leave the PLM pair of neurons receiving input uninjured. All other neurons have their mi values assigned from the experimental statistical distributions. The governing equation for an injured neuron is now
C V i ˙ = - G c ( V i - E c e l l ) - ( I i G a p ( V → ) + I i S y n ( V → ) ) / ( 1 + μ · m i ) (11)
We can readily interpret the limiting cases: when μ ⋅ mi = 0, the original governing equation is recovered, and thus μ = 0 corresponds to the healthy case. When μ ⋅ mi is large, gap junction and synaptic currents have no effect. The neuron’s voltage decays exponentially to its leakage potential, effectively isolating it from the network.
Note that our random assignment of swelling values neglects any spatial structure of the injury. This could be easily modified by using a distribution which depends on the spatial location of the neuron. Furthermore, this is a very simple model for neuronal swelling, in keeping with our simple model for neurons. It necessarily neglects the actual geometry of swelling. The use of a multi-compartment model would enable this in future studies. Ultimately, there is currently limited biophysical evidence for making more sophisticated models. As such, we have tried to capitalize on as many biophysical observations as possible so as to make a model that is consistent with many of the key experimental observations.
We use MATLAB (version R2013a) to solve the system of neuronal dynamical equations via Euler’s method, using a timestep of 10−4s. We consider an ensemble of 1,447 different types of injury (set of targeted neurons), for which the global intensity μ may vary from 0 (uninjured) to a critical value μ*. When the intensity exceeds μ* (found by a bisection algorithm), the limit cycle collapses to a fixed point. To obtain intermediate values, we perform five simulations linearly spaced throughout (0, 0.9μ*) and ten additional simulations throughout (0.9μ*, μ*).
We classify the resulting injured trajectories as a Fixed Point or a Periodic Orbit according to the following criteria:
Note that these criteria classify very small periodic orbits as fixed points, since their behaviors are very similar. The method may also classify sufficiently slow, long-timescale oscillatory transients as periodic. These tests ignore the first 5 seconds of simulation time (50,000 timesteps), chosen heuristically as a typical timescale of transient decay. After this initial wait, we check the criteria at the end of each subsequent 5 seconds of simulation time until convergence is detected. The results were not observed to be sensitive to the length of this interval.
Stephens et al. [37] found that the forward crawling motion of C. elegans is well described by two principal component body-shape modes called eigenworm modes. When moving forward, the modes alternate within its phase space forming a limit cycle. Kunert et al. [39] also found a two-dimensional limit cycle, but for the collective motorneuron activity after PLM stimulation. They interpret this similar dynamical signature as a neuronal analog to the observed behavioral pattern.
To interpret the distorted neural activity caused by our simulated injuries, we construct a map from the neuronal activity plane onto the eigenworm plane. The body-shape modes were extracted from Figure 2(c) of [37]. We first calculate the optimal linear mapping of the healthy trajectory onto a circle (see Fig 3a). We then use this calibration for all other trajectories. This artificially translates anomalous neuronal responses to anomalous body motions. Our procedure has a number of limitations, for which we list a few:
The lack of direct neuronal analogs for injured network modes limits our ability to interpret arbitrary impaired behavioral responses. Further computational work could also find neuronal proxies for additional behavioral modes so as to enable a more complete mapping. Recent work on blast injuries of worms [45] could potentially help extend the analysis by providing injured eigenworm mode projections.
Procrustes Distance (PD) measures the dissimilarity between shapes, and in our context, we wish to compare the shape of the trajectories of the healthy neural responses (circular orbits in the phase plane) with the distorted ones produced after simulated injuries. For that, we use the function procrustes.m from MATLAB’s Statistics and Machine Learning Toolbox. We collect N data points from each trajectory and annotate their (x, y) coordinates in a (2 × N) shape matrix S. The PD between two distinct shapes SA and SB is given by
P D = min b , R , c ∥ S B - b · S A · R + c → ∥ 2 (12)
In other words, it finds the optimal (2D) rotation matrix R, scaling factor b > 0, and translation vector c → to minimize the sum of the squares of the distances between all points. Intuitively, it compares shapes discounting translation, rotation, or scaling. To calculate the PD curves as in Fig 4, we use the uninjured (μ = 0) limit cycle as our first shape SA. The second shape SB is the limit cycle calculated for each injury at the indicated value of μ.
We pre-process the trajectories to extract data points only within a single period. Since injuries usually distort the trajectory length, we use MATLAB’s spline.m function to interpolate them and collect the same number of data points. Both limit cycles must also be phase-aligned, which we achieve by finding the phase that minimizes the Procrustes Distance.
We hypothesize that both the injury itself and the PD curves contain meaningful signatures of behavioral outcomes of a given injury. For example, there is always a critical injury level μ = μ* in which the injured response collapses into a fixed point. Our artificial map suggests that this endpoint location corresponds to the shape of a paralyzed worm. We thus wish to relate endpoint location to (1) the shape of the PD curve, and to (2) the injury vector m →.
For these purposes, we classified the endpoints simply by dividing the endpoint distribution along its major axis. Specifically, we take the distribution of endpoints in Fig 5, calculate the leading principal orthogonal mode (via taking the Singular Value Decomposition, as mentioned earlier), and classify the points by the value of their projection onto this mode (where we arbitrarily classify projection values ≥ −0.01 as the “upper plane” and < −0.01 as the “lower” plane). Given this definition, 63.2% of the points lie within the upper plane, and 36.8% lie in the lower plane. Note that all of the forthcoming analysis could be equally well applied to any other output feature, and so we choose this classification for its relative simplicity.
We calculate the average PD curve within each class. Since the PD curves may have a different number of points, we first pre-process them. Specifically, we normalize the maximum μ and Procrustes Distance to 1 for all curves, and then interpolate them using MATLAB’s spline.m such that all curves have the same number of points. We then simply take the average and standard deviation to obtain the average curves shown within Fig 5. This figure also plots the average scaling and translation curves as a function of injury level, for each class. Scaling factors (i.e. the factor by which the size of the distorted limit cycle has decreased from the original cycle) are given as an output of MATLAB’s procrustes.m as used above. Translation distance is found by calculating the location of the mean of each distorted cycle, and then calculating the distance by which this mean is displaced from the origin. These curves are then normalized, interpolated and averaged, yielding the average curves in Fig 5. Note that, unlike the PD curves, translation and scaling are monotonic and not distinct between classes, and thus they do not carry the same information about the functional outcome of the injury.
We use the ClassificationTree class from MATLAB’s Statistics Toolbox (version R2013a). Fitting and cross-validation are performed using the included methods ClassificationTree.fit and kfoldLoss with default settings (10 folds). The minimum leaf size was set by calculating cross-validation error over a range of minimum leaf sizes (see Fig 6b). For both PD curves and Injuries, cross-validation errors are optimal at a minimum leaf size of around 40. We use this minimum leaf size for all fits.
The classification tree that uses normalized PD Curve Shapes to predict the endpoint class yield a cross-validation error of 22.0%. We can compare this to the random case (i.e. the case where PD Curve Shape has no relationship to the class) by repeating this analysis with randomly shuffled class labels. For 100 trials with randomly-shuffled labels, the observed cross-validation error was 43.8 ± 1.4%. Injury vectors were also used to fit classification trees for predicting endpoint classes (see Fig 6). The cross-validation error of 14.6% was significantly lower in this case, while the randomly-shuffled labels analysis returned a error of 44.6 ± 1.3% (consistent with the random error above). In both cases we observe that the cross-validation error is far below what we would expect if the data had no relation to the classes.
Thus we can predict (with cross-validated accuracy exceeding 85%) the region into which the endpoint will fall given a specific injury. Moreover, the classification tree in Fig 6 is very simple to interpret and depends on only three neurons: ALML, AVM and SDQL. As per WormAtlas [51], all three of these neurons have sensory functions (ALML and AVM are mechanosensory; SDQL is an interneuron but is oxygen-sensing).
This study introduces a tractable framework for analyzing how biophysically-inspired injuries distributed across a physical neuronal network induce behavioral deficits. The specific injuries we consider arise from FAS which has been implicated in most leading neurodegenerative diseases, aging and TBI. By identifying low-dimensional population codes within our model which correspond to a known behavior, a proxy metric for cognitive deficit can be constructed. Specifically, limit cycles in our dominant features serve as a neural proxy for actions such as forward motion in the C. elegans. Such trajectories can be artificially mapped onto experimental body-shape modes, and suggests a behavioral interpretation of the distorted limit cycles resulting from an injury. Our analysis also suggests that there is a diversity of functional deficits that arise from the same level of injury on a connectomic network.
The ability to provide a theoretical understanding of functional, cognitive and behavioral deficits due to connectomic injuries is a the forefront of TBI and neurodegenerative disease studies. Both have an enormous societal impact and implications. Specifically, TBI is annually responsible for millions of hospitalizations [52, 53], with reports estimating that 57 million people worldwide experienced some form of TBI [14]. It was also manifest in around 15% of all veterans of the Iraq and Afghanistan wars, with blast injuries being the signature wound of these conflicts [14, 53]. Numerous studies show that even mild concussions, if induced repeatedly, can lead to permanent brain damage; the issue is constantly debated in the sports media, but especially in football [54]. Neurodegeneration affects orders of magnitude more people than TBI through diseases such as Alzheimer’s disease [10, 11], Creutzfeldt-Jakob’s disease [15], HIV dementia [16], Multiple Sclerosis [17, 18] and Parkinson’s disease [19]. Thus, any study that can help understand how FAS compromises cognitive function is of great value.
Simulated injuries distort dynamical signatures in the network’s activity, such as limit cycles. Our Procrustes Distance metric quantifies how much the shape of the limit cycle is distorted, compared to the healthy cycle. Our results indicate that as different injuries evolve, this metric follows qualitatively different trends (as in Fig 4). In all trials, a sufficiently high injury level μ = μ* collapses the limit cycle into a stable fixed point. The shape of the PD curve helps inform the location of this fixed point (as in Fig 5). This suggests that the shape of the PD curve, as the injury evolves, may help predict the eventual behavioral outcome (e.g., the body shape the worm will assume during temporary paralysis). Thus we have prescribed a method to monitor the dynamics of the injured worm and the implications of the injury as it evolves. Finally, our classification trees divides neural injuries into two distinct classes of functional outcomes (i.e. endpoints in the “lower” or “upper” portions of the distribution). Its cross-validation predictive accuracy is over 85% and implicates only three specific neurons (ALML, AVM, and SDQL). This relationship between injury structure and behavioral outcome is simple, interpretable and testable. Such trees can be fit for arbitrary injured behaviors and could be used more broadly for any given model of injured full-Connectome dynamics.
The metrics and methods described in this work can potentially be used to construct diagnostic tools capable of identifying a variety of cognitive deficits. Moreover, the severity of a TBI injury and/or neurodegenerative disease can be quantified by measuring its metric distance from the normal/healthy performance. Our work gives clear mathematical tools capable of formulating such diagnostic tools for assessing injuries and functional deficits.
The present study has many limitations, many due to the lack of biophysical evidence required to build better models. For example, though we treat all neurons as identical passive, linear units, it is known experimentally that different neurons appear to exhibit different behaviors (for example, some neurons appear to be functionally bistable [55] and could be modeled as such, as in [56]). We predict the results of injuries only on the two “forward-motion” motorneuron modes, ignoring other modes potentially associated with impaired behaviors. Furthermore, the exact mapping of our motorneuron voltage modes onto these body-shape modes is ambiguous. The model lacks muscles and body features of the worm which limits our ability to make more general predictions. We also neglect external feedback mechanisms required for sustained and spontaneous forward motion, and assume that tail-touch neurons are constantly stimulated. It is uncertain how such feedback mechanisms would alter the trajectory. The order-of-magnitude parameter estimates of our model parameters also make direct quantitative comparisons difficult.
We believe the merit of this study lies not so much on the specific results presented, but on the new directions and methodologies it opens for future work. In fact, computational and experimental studies on the effects of network injury are still at their infancy for C. elegans and other models. Many limitations of this work could be overcome with a more detailed model for the C. elegans neuronal network both before and after injury. Coupling this with an external, mechanical model would allow for more general predictions. This could be accomplished with simplified mechanical models for locomotion (such as in [56]) or with more complete, future “in-silica” models such as OpenWorm [57]. The development of such models, which do not ignore the spatial extent and shape of neurons, would allow for the study of the effects of injuring individual connections, or the effect of injuring individual neurons non-homogeneously. This study suggests that such modeling work should also consider how to model neural injuries, after which our analysis techniques could be applied directly.
Experimental studies would not only test our model, but also in, in conjunction with our work, provide a new testbed for models of injured connectomic dynamics. Our Procrustes Distance metric, shown here to carry information about the eventual outcome of an injury, may also be useful in the real-time analysis of injury progression. Thus our study provides a way forward in monitoring behavioral outcomes of injured networks.
Ultimately at present, limitations in biophysical measurements and neural recordings make it extremely difficult to identify more sophisticated underlying mechanisms responsible for dysfunctions in neural networks, especially when circuits display intrinsically complex behavior and functional activity. We believe the rapid advancement of recording technologies in neuroscience will significantly help refine the model presented here.
Given that the modeling of neuronal networks is one of the most vibrant fields of computational neuroscience [49, 58, 59], our contribution provides a comprehensive study of how the effects attributed to FAS jeopardize the network functionality, opening new possibilities and objectives for the study of network architectures.
|
10.1371/journal.pntd.0001968 | Deletion of Parasite Immune Modulatory Sequences Combined with Immune Activating Signals Enhances Vaccine Mediated Protection against Filarial Nematodes | Filarial nematodes are tissue-dwelling parasites that can be killed by Th2-driven immune effectors, but that have evolved to withstand immune attack and establish chronic infections by suppressing host immunity. As a consequence, the efficacy of a vaccine against filariasis may depend on its capacity to counter parasite-driven immunomodulation.
We immunised mice with DNA plasmids expressing functionally-inactivated forms of two immunomodulatory molecules expressed by the filarial parasite Litomosoides sigmodontis: the abundant larval transcript-1 (LsALT) and cysteine protease inhibitor-2 (LsCPI). The mutant proteins enhanced antibody and cytokine responses to live parasite challenge, and led to more leukocyte recruitment to the site of infection than their native forms. The immune response was further enhanced when the antigens were targeted to dendritic cells using a single chain Fv-αDEC205 antibody and co-administered with plasmids that enhance T helper 2 immunity (IL-4) and antigen-presenting cell recruitment (Flt3L, MIP-1α). Mice immunised simultaneously against the mutated forms of LsALT and LsCPI eliminated adult parasites faster and consistently reduced peripheral microfilaraemia. A multifactorial analysis of the immune response revealed that protection was strongly correlated with the production of parasite-specific IgG1 and with the numbers of leukocytes present at the site of infection.
We have developed a successful strategy for DNA vaccination against a nematode infection that specifically targets parasite-driven immunosuppression while simultaneously enhancing Th2 immune responses and parasite antigen presentation by dendritic cells.
| Filarial infections are endemic in more that 80 countries, affecting over 120 million people and putting 1 billion more at risk. Antifilarial drugs must be administered regularly to infected people to control the disease, but they are contraindicated in under 6 year-olds and in pregnant women. Further, reports of drug resistance are now accumulating. A vaccine would therefore greatly help fight these diseases. Live attenuated L3 filariae larvae can evoke a protective immunity but their production is impractical and use in humans unacceptable while the efficacy of sub-unit vaccines has been poor. Filariae secrete proteins capable of suppressing their host's immune response, and have the potential to interfere with immunisation. We therefore decided to vaccinate hosts against secreted parasite products that modulate host immune responses rather than against structural components of the worms, and to boost the host's immune system by directly enhancing the uptake of parasite material by antigen presenting cells. This strategy generated substantial protection against both adult and offspring of a filarial parasite in mice. This provides a strong proof of principle for the anti-immunomodulatory approach we have developed.
| DNA vaccination is a promising technology that is being developed to combat diseases such as flu, HIV, and cancer [1]. Furthermore, the stability and relatively low production cost of DNA vaccines provide hope for treating individuals in developing countries. After more than 20 years of mass drug treatment programs based on a limited choice of drugs, lymphatic filariasis and onchocerciasis remain major public health problems. Control programs in some areas are now threatened by the emergence of drug-resistance [2], [3], while in communities where both onchocerciasis and loiasis are endemic, mass treatment with ivermectin is contraindicated because of the risk of severe side effects associated with death of Loa loa microfilariae [4]. These circumstances argue strongly for the development of vaccines to complement drug treatment strategies.
Filarial nematodes are tissue-dwelling parasites that can be killed by exposure to Th2-mediated effector mechanisms, with eosinophils and antibody particularly critical for protection against re-infection [5]. However, these parasites establish chronic infections in a large number of species, including ∼150 million humans in whom immunopathological reactions cause a spectrum of clinical diseases. The persistence of filarial parasites has been shown to be enabled, in part, by excretory/secretory (ES) products [6]–[10]. As a consequence, the efficacy of a vaccine against filarial infections is likely to depend on how well it disrupts parasite immune evasion mechanisms. Given that the maintenance and transmission of filarial infections requires very few adult parasites [11] and that our previous work suggests that they are able to increase their fecundity in response to host immune attack [12], any intervention strategy should be assessed by its ability to suppress the transmissible stages, the microfilariae. Not only is this critical to reducing disease transmission, but in onchocerciasis microfilariae are the main cause of pathology.
The rodent filarial nematode, Litomosoides sigmodontis, has been used to elucidate many of the regulatory pathways that allow parasite establishment and is ideally suited for assessing vaccination success against filarial nematodes as the parasites can develop patent infections in BALB/c mice [5], [13]. The infective larvae are inoculated subcutaneously, and over 4–6 days migrate to the pleural cavity, where they mature to adulthood, mate and produce microfilariae that enter the bloodstream by 60 days post-infection. Consequently, protective immunity against all stages of the parasite's lifecycle can be assessed. We tested the vaccine potential of two L. sigmodontis immunomodulatory ES products by expression of these proteins in DNA vaccine constructs designed to improve antigen presentation and enhance Th2 immune responses [14], [15].
While there have been previous attempts to immunise rodents against filarial nematodes using DNA vaccines [16]–[18], this is the first study to do so that specifically targets parasite-driven immune modulation and uses a filarial parasite that produces patent infections in laboratory mice, thus allowing manipulation and deeper examination of host immunity through methodology developed in inbred mice.
All procedures involving animals were approved by the University of Edinburgh ethical review committee, and performed under license from the UK Home Office in accordance with the Animals (Scientific Procedures) Act 1986.
All immunisations and infections were performed with female BALB/c mice, starting at ages of 6–7 weeks. Mice were housed in individually ventilated cages and infected subcutaneously with 30 or 40 L. sigmodontis infective larvae (iL3). Two experiment endpoints were chosen based on the life cycle of L. sigmodontis, D10 post-inoculation (p. i. ) when most larvae will have reached the L4 stage; and, at D60 p. i. after the onset of the patent phase.
All cloning was carried out following the recommendations of the pcDNA 3.1 Directional TOPO Expression Kit (Invitrogen). LsALT (DQ451171.1) and LsCPI (AF229173.1) were amplified from a cDNA preparation of L. sigmodontis iL3 with the primers detailed in Table S2. The desired mutation of asn66 to lys66 in LsCPI was generated with the QuikChange Site Directed Mutagenesis Kit (Stratagene) using primers detailed in Table S2. Fusion constructs containing single-chain anti-DEC205 antibody (DEC) upstream of our target antigen were produced from ready-made constructs kindly provided by Dr. Ralph Steinman [19]. Briefly, PCR products of genes of interest were digested with NotI and XbaI (Neb laboratory, UK), then ligated into an NotI and XbaI-digested anti-mouse dec-205 single chain antibody - ovalbumin construct (DEC-OVA) or antibody control Ig-OVA to replace the fragment of OVA gene, respectively. All plasmids were sequenced to confirm identity.
LsALT and LsCPI were cloned into pET21b and pET29c respectively. Both plasmids were made with PCR products derived from primers spanning the full CDS minus the stop codon and with flanking restriction enzyme recognition sites. The recombinant CPI and ALT containing a T7 tag at the N terminus and His tag at the C terminus were produced in the E. coli BL21(DE3) strain (Novagen UK) and purified on His- binding resin columns using automated AKTAprime (Amersham Pharmacia).
Plasmids were injected in the tibialis anterior muscle of the left leg with a 27G needle, immediately followed by electroporation with an ECM 830 generator+Tweezertrodes (BTX Harvard Apparatus) using as settings 8 pulses, 200 V/cm, 40 ms duration, 460 ms interval [20]. Each mouse was immunised twice separated by 2 weeks interval with 40 µg of DNA total made up by equal quantities of each plasmid species, delivered in 50 µl PBS. As a consequence, the quantity of each individual plasmid was reduced as the number of different plasmids incorporated into the inoculums increased. However, the quantity of each one remained in excess of the minimal efficient dose [21]. In vivo expression of the gene of interest was determined by the detection of mRNA with specific primers 4 weeks after immunisation (Figure S1 and Table S3). Mice were challenged 4–6 weeks after the last immunisation.
Blood was collected from individual mice after the first immunisation, then every other week to measure the increase in antibody titres. At experiment end point cells were recovered from thoracic lymph nodes for antigen recall assays of specific cytokine production and proliferation. Pleural lavage fluid was also collected for cytokine and cellular infiltrate measurement. Cytokines IL-4, IL-5, IL-10, IFN-γ and IL-13, and total IgE levels were measured by sandwich ELISA, and specific anti-L. sigmodontis IgG1 and IgG2a responses were measured by indirect ELISA against whole soluble extract coated at 10 µg/ml, and anti-LsALT or anti-LsCPI antibodies against the respective native recombinant proteins coated at 5 µg/ml using the antibody pairs described elsewhere [22] and detected with TMB-H2O2+(KPL) at 450 nm. Titres were assessed by two fold serial dilution of the serum, and determined as the highest dilution factor for which O. D. values exceeded 3 standard deviations above control wells from the same plate. All ELISAs were repeated at least once. The 1/800 O. D. was used in some analyses.
Parasite survival was determined at experiment endpoint. Adult filariae were isolated from the pleural cavity lavage fluid in10 ml cold PBS, fixed in hot 70% ethanol and counted. Protection was calculated as (mean burden in primary infected animals - mean burden of vaccinated animals)/mean burden in primary infected animals. Microfilariae were counted in 30 µl of blood after fixation in 570 µl of BD FACS lysing solution (BD Biosciences) under an inverted microscope.
Generalised linear models (GLM) were used to compare the effects of different vaccine formulations on parasitological and immune parameters as they allow more flexibility in specifying the distribution of response variables and better model fitting through Maximum Likelihood estimation. Parasite and cell counts were modelled assuming poisson or negative binomial error distributions; antibody levels were log-transformed. Homoscedasticity and normality of the residuals were assessed with the Fligner-Killeen test of homogeneity of variances and the Shapiro-Wilk normality test, respectively. The non-parametric Kruskal-Wallis rank sum test followed by pairwise comparison with the Wilcoxon rank sum test with Bonferroni correction for multiple comparisons were applied to data that didn't conform to a standard distribution. Average parasite counts per group were used to calculate pairwise measures of protection. A principal component analysis (PCA) was performed to identify major trends across the large number of immunological measures performed. Briefly, data were scaled to null mean and unit standard deviation, and the broken-stick criterion was then used to select the principle components having the best explanatory power [23]. The biological interpretation of the resulting PC was then based on the relative contribution of each immune factor to it, e.g. a preponderance of Th2-associated factors or of pro-inflammatory factors. Individuals' predicted values from the PCA were then used as explanatory variables in a GLM to assess which components were best correlated with parasite killing. Statistical analyses are reported in figures and legends only when significant effects of vaccine formulations were found. All analyses and graphs were performed in R 2.15 [24].
LsALT is the most abundant transcript produced by the infective larvae of filarial nematodes [25], [26], and is suspected of modulating the host Th2 immune responses [9]. ALT's potential as a vaccine has been tested by several laboratories using different filarial and host species and has shown variable success in inducing immunity [17], [25], [27]–[31]. As a starting point for this study the immunisation efficacy of plasmids containing L. sigmodontis antigens was enhanced by electroporation into skeletal muscle [32]. However, our attempts using plasmids expressing LsALT (ALT) to immunise mice against L. sigmodontis failed to induce Th2 responses. We thus wanted to improve immunisation efficacy against LsALT by specifically enhancing Th2 responses with IL-4, and by enhancing dendritic cell recruitment and activation. Levels of IgG2a, IgG1and IgE were measured to determine the relative balance between Th1 and Th2 responses while IL5 production and eosinophil numbers where used as measures of vaccine-induced immunity against the filariae [33], [34].
BALB/c mice immunised with ALT and challenged with infective L. sigmodontis larvae, failed to generate LsALT-specific IgG responses or local eosinophil recruitment ten days after challenge. Although not reaching significance, the co-administration of plasmids encoding IL-4 (pIL4) increased LsALT-specific IgG1 but not IgG2a levels, while antigen presenting cell activators MIP-1α (pMIP) and Flt3L (pFlt3L) [35] marginally increased both IgG1 and IgG2a (Figure 1A–B).
Coinjecting ALT with pMIP+pFlt3L significantly increased the production of total IgE production compared to ALT alone (Figure 1C). There was an indication that eosinophil recruitment to the site of infection was also increased by the addition of pIL-4 and pMIP+pFlt3L compared to ALT alone (Figure 1D). The vaccine containing ALT alone resulted in non-significantly higher parasite survival than in pEmpty controls (Figure 1E), while the addition of pIL4 or pMIP+pFLt3L seemed to counter this facilitation. Overall, no protection was detected at day 10 post-challenge.
These data showed that ALT could generate specific antibodies, but that it was poorly immunogenic, and possibly immunomodulatory. Previous data have suggested that the variable acidic domain of Brugia malayi ALT-2 confers the protein its immune modulatory properties [9]. We thus decided to generate a mutated LsALT lacking the acidic domain (ALTm) in the hope that this would enhance LsALT-specific immune responses to the remaining epitopes. Immunising with ALTm resulted in a significant increase in LsALT-specific IgG1 production above naïve controls and a marginal increase above the native form ALT (Figure 2A), consistent with the hypothesis that the deletion of the acidic domain would overcome LsALT-driven immunosuppression. The ALTm vaccine induced a strong increase in LsALT-specific IgG2a concentrations above naïve controls (Figure 2B), but no further increase above the native form ALT. Although total IgE production and eosinophil recruitment were higher in the ALTm group than in the ALT group, this was not statistically significant (Figure 2C–D). Despite relatively high antibody production in several repeats of this experiment, significant protection against a challenge infection was never achieved (Figure 2E).
In an attempt to enhance the protective potential of ALTm, a plasmid was constructed to encode a fusion protein comprising ALTm and an anti-DEC205 single chain Fv antibody (decALTm). The expressed fusion protein directly binds dendritic cells through the DEC205 surface receptor [19]. A second plasmid encoding a fusion protein comprising an antibody with irrelevant specificity and ALTm (isoALTm) was constructed for use as a control. For these experiments we chose to include pIL-4, pMIP and pFlt3L in all vaccine formulations to maximise their protective potential. Immunising with decALTm resulted in increased LsALT-specific IgG1 and IgG2a concentrations in 2 out of 5 mice, but did not increase total IgE or eosinophilia above the non-DC-targeted form, ALTm (Figure 3A–C). Despite demonstrable enhancement of IgG1 antibody levels by both ALTm and decALTm compared to empty vector, no statistically significant reduction in parasite numbers was detected at day 60 post-challenge (Figure 3E). This suggests that immunising against LsALT alone was not sufficient to evoke protective immunity.
Because removal of immune modulatory sequences from LsALT had a significant impact on its ability to induce an immune response, we decided to apply a similar approach to another immune modulatory filarial protein. Filarial cystatins are potent downregulators of inflammation [36] and antigen processing by host cells, as shown for CPI-2 from B. malayi which inhibits asparaginyl endopeptidase activity [37]. We isolated the L. sigmodontis homologue, LsCPI, and cloned it into the same pcDNA3.1 vector as LsALT, either in its native sequence (CPI) or with a mutation that disrupts its AEP inhibitory activity (CPIm).
To verify our previous findings with immune enhancing plasmids the native and modified constructs of LsCPI were administered with or without pIL4+pFlt3L+pMIP (indicated as +adj in figure 4). LsCPI-specific IgG1 production was significantly enhanced by the co-administration of pIL4+pFlt3L+pMIP (Figure 4A). The combination of the mutation of CPI and pIL4+pFlt3L+pMIP significantly enhanced the production of IgE compared to all the other groups (Figure 4B). However, no protection from challenge infection was achieved by the vaccine regimens above (Figure 4C). We therefore followed the DC-targeting strategy used with ALT: CPI and CPIm were fused with a scFv-DEC205 sequence (decCPI and decCPIm, or isoCPI for scFv isotype control). All groups received pIL4+pFlt3L+pMIP, as this had also improved immunisation against LsALT (Figure 3). In this experiment, both the mutation of CPI and the targeting of dendritic cells enhanced immune responses to LsCPI. The resulting CPIm and decCPIm constructs induced strong increases in LsCPI-specific IgG1 compared to isoOVA-, CPI- and CPIm-immunised animals (Figure 5A) and in IgG2a compared to CPI (Figure S2A) as well as total IgE (Figure 5B). Nonetheless, there was no significant protection as neither adult worm recoveries (Figure 5C) nor microfilariae densities (Figure 5D) differed statistically between control mice and mice immunised with CPIm or decCPIm. However, the mutation of CPI showed a trend towards reduced microfilariae in the peripheral circulation (Figure 5D).
On their own neither ALT or CPI induced a substantial protective effect. However, recombinant vaccines can work more effectively in combination [16], [27], [38]–[40] and indeed in a cattle model of onchocerciasis both ALT and CPI were part of a cocktail of recombinant proteins that generated protection against natural challenge [41]. We thus chose to use a combination of both parasite antigens along with the full complement of ‘adjuvant’ plasmids in an effort to increase the level of protection. Significant protection was achieved when mice were immunised with dual modified parasite antigens and the full combination of cytokine-expressing plasmids (Figure 6). Adult parasite numbers were reduced by 71% in the mice that received the full modified vaccine relative to those that received unmodified antigens ALT+CPI, by 65% compared to those that received ALTm+CPIm+adjuvants and by 68% compared to those that received only empty plasmids (Figure 6A). Average microfilaraemia in both the full vaccine and the ALTm+CPIm groups was reduced by over 85% compared to both the ALT+CPI group and empty plasmid controls (Figure 6B). The lack of protection in the ALT+CPI group suggests that combining vaccine targets is not sufficient on its own. Further, the benefit of targeting the antigens to dendritic cells via DEC205 was to induce more rapid killing of adult parasites.
The importance of the ‘adjuvants’ was confirmed in a separate experiment. Significant protection was achieved by D60 post-challenge in the groups that received the dual vaccine decALTm+decCPIm, reaching an average 82% reduction of adult parasite burden relative to non-vaccinated animals (Figure 6C), and a 90% reduction of circulating microfilariae (Figure 6D). The dramatic effect on microfilariae numbers without full removal of the adult population suggested that at least some of the vaccine impact was occurring late in infection and was specifically targeting larval stages or worm fecundity.
To determine the relative contribution of the ‘adjuvant’ plasmids to protection, mice were immunised with decALTm+decCPIm with either pIL4, pMIP+pFlt3L, or both as in the full vaccine. While the reduction in adult worm counts by the full vaccine formulation reached 57%, the formulation with pIL4 alone showed no protective effect relative to pEmpty controls, and the formulation containing MIP+pFlt3L achieved 64% protection (Figure 6E). Effects of these formulations on the microfilariae were interesting (Figure 6F): in mice that became positive for microfilariae, vaccination with pMIP+pFlt3L reduced microfilaraemia by 90%; with pIL4+pMIP+pFlt3L, microfilariae counts were reduced 70%; however pIL4 caused a 1.5× increase in average microfilariae numbers in patent mice, which is consistent with our findings that IL-4 is associated with eosinophil-induced increase in worm fecundity in vaccinated animals [12], [42]. Despite considerable variability in worm recoveries we have always found (5 experimental repeats) that the comparison between pEmpty and the full vaccine has exceeded 65% reduction in microfilaraemia.
These data suggest that targeting immune modulatory proteins had protective effects against multiple parasite stages with substantial IL4-independent disruption of microfilariae production and gradual killing of adults, consistent with the finding that early disruption of Treg function during the onset of L. sigmodontis infection is sufficient to enhance microfilarial killing 60 days later [43].
Many of the readouts assessed for this study did not reach statistical significance, perhaps due to high variability in both immune parameters and parasite numbers. We took statistical advantage of the substantial variation in both immune and parasitological read-outs to analyse the most prominent immune correlates of protection observed in Figure 6C. For example, pleural recruitment of eosinophils and parasite-specific IgG1 production were increased in the dual-vaccinated group (Figure S2B–C) although no significant effect on total IgE production was detected (Figure S2D). In total, thirty-one variables for each mouse (Table S1) were measured. To facilitate their analysis, we reduced them to principal components (PC) that summarise major patterns in the immune response. Three PC tested as significant and interpretable that captured 50% of the variation present in the full dataset (Table 1, Figure S3). Subsequent components were rejected for lack of explanatory power. The first component (PC1) included mainly lymph-node cell production of Th2 cytokines IL-5 and IL-13. PC2 included whole worm-specific IgG1, pleural IL-5, and pleural eosinophils, neutrophils, macrophages and less prominently, pleural lymphocytes. PC3 included mainly LsALT- and LsCPI2-specific IgG1, unstimulated IL-4 production by lymph node cells in vitro, anti-CD3 induced production of IFN-γ and pleural lymphocyte numbers (see Table S1 for individual rotation values). The explanatory power of the resulting components for parasite survival was assessed by GLM. Parasite numbers were affected by only the second principal component PC2, revealing a strong negative correlation between PC2 and parasite survival (Figure 7). An analysis of variance confirmed that the decALTm+decCPIm vaccine formulation drove most of the variation in PC2 (Figure 7).
Because IgE was weakly represented in those PCs while IgG1 was strongly represented in PC2, we wanted to determine the relative contribution of IgE, IgG1 and IgG2a concentrations to parasite killing. This confirmed that only IgG1 had a significant effect on parasite numbers (P = 0.015). Likewise, the analysis of respective roles for pleural cell types in protection revealed that parasite killing was significantly affected by pleural lymphocytes and neutrophils (P = 0.002 and P = 0.03 respectively), but only marginally by eosinophil numbers (P = 0.07) despite a significant negative correlation between eosinophil and parasite numbers (r = −0.5, P = 0.02). Taken together, these data are strongly suggestive of IgG1 and pleural leukocytes as being the main effectors in the decALTm+decCPIm vaccine-induced parasite killing.
Like many parasitic helminths, filarial nematodes establish long-lasting infections that are facilitated by immunomodulatory products secreted by the parasite [7]. Host-driven immune down-regulation can contribute to parasite survival [13], while in other patients, anti-filarial responses to adult and juvenile parasites are associated with immunopathology. Indeed, since immunity to helminths is mainly mediated by Th2-type responses, the risk of a vaccine generating excessive eosinophilia, IgE-mediated mast cell degranulation and related pathologies must be considered carefully [44]. Finally, we have previously demonstrated that filarial nematodes are able to adapt their developmental schedule to the hosts's eosinophilic response, thereby shortening their time to transmission [12]. We thus inferred that a successful vaccine against filarial infections would need to: (a) evoke a Th2 response, possibly through a path different from that driven by natural challenge; (b) target the parasite molecules that suppress protective immunity; (c), avoid inducing immune hyper-responsiveness. This suggested that the best vaccine candidates would be found among excreted/secreted molecules rather than structural components.
We selected parasite antigens based on their abundance in gene expression profiles [25], [26], potential to induce protective immunity [31], [45], and on their role in immune modulation as supported by in vitro studies [8], [9], and decided to take advantage of the flexibility and ease of production of DNA vaccination. Administering plasmids encoding the native sequences of LsALT and LsCPI failed to generate strong specific immune responses in mice, perhaps because of the immunomodulatory properties of these proteins when directly expressed in eukaryotic cells. This differs from other DNA vaccination studies in which ALT from the human parasite B. malayi induced a good response in rodents following challenge [17], [27], [28]. It may be that the immune modulatory properties of ALT are host-specific and are more readily manifested in the permissive host-parasite combination used in our study. Indeed our finding that when LsALT and LsCPI were genetically modified to remove immune modulatory residues, specific antibody responses increased, provides in vivo support to the in vitro evidence that these domains are immunosuppressive [8], [9]. It further indicates that the immunosuppressive function of these proteins can be successfully removed in vivo, thereby allowing vaccines that contain them to generate significant protection against multiple stages of the parasite.
Simultaneously with the plasmids expressing parasite antigens, we administered plasmids expressing host cytokines IL-4 to enhance Th2 responses required for the elimination of filarial infections [5], as well as MIP-1α and Flt3L to enhance activation and recruitment of dendritic cells [19], [46], [47]. In addition, a sequence encoding a single-chain Fv antibody directed against the dendritic cell surface marker DEC-205 [19] was added to the parasite sequence in order to maximise the uptake of antigen by the DCs recruited by MIP-1α, Flt3L and the injections. We hypothesise that it is the combination of these steps that lead to successful protection against infection. Further, our results demonstrate the feasibility of combining different adjuvant systems in a single vaccine to increase its efficacy. However, our parasitological data display substantial variability, especially in the rate of adult worm killing. This variability is inherent to the study system, and is consistent with that observed in natural filarial infections [41], [48]–[50].
A recurring problem with DNA immunisation has been the lack of protective immunity despite the ability to generate specific immune responses [1]. The gold standard of vaccination against filarial nematodes is immunisation with irradiated larvae, and requires the presence of functional eosinophils and antibody [51], [52]. A multivariate analysis of the immune factors that lead to protection in our present study revealed a negative correlation between whole parasite-specific IgG1 (measured in the blood) and the numbers of leukocytes at the site of infection, implying that immune effectors were correctly potentiated during the immunisation phase. Intriguingly, serum concentrations of IgE did not correlate with parasite numbers, despite being enhanced by the most protective vaccine regimens, and mediating protection in another rodent model of vaccine-mediated L3 killing [52]. A further consequence of this finding is that IgG1 and antigen-specific T cell responses may be sufficient to assess immunisation efficacy and the generation of protective immunity, but that IgE may not be an accurate marker for protection in mice, which may be due to the lack of Fc epsilon receptors on murine eosinophils [53]. The more prominent association between neutrophilia and protection is in accordance with other studies showing their role in secondary immunity to Strongyloides stercoralis in mice [54]. Further, our data suggest that in addition to the importance of targeting the parasite directly, the choice of adjuvants is crucial to generating immune responses that are protective. The coinjection of IL-4, MIP-1α and Flt3L lead to substantial reductions in parasite numbers, which was maintained or even improved when only MIP-1α and Flt3L were administered. Further refinements of this vaccine formulation are underway, but it currently appears that both DEC205 and MIP-1α+Flt3L are needed but the IL-4 may be dispensable.
In conclusion, strategic use of DNA vaccine technology has allowed us to test a large number of parameters and combinations of immune modulators that would not have been logistically possible if we needed to produce all the individual recombinant proteins in active form. This study provides a proof-of-principle that targeting parasite products that suppress the immune responses of their hosts while enhancing antigen presentation can lead to significant protection. Anti-evasion immunisation is garnering an increasing amount of attention in a wide range of pathogenic systems [55]–[58]. Our study shows that it is crucial and feasible to ablate the immunomodulatory function of such candidates for them to generate protective immune responses and we expect that this strategy may be applied to a wide range of diseases. Whether or not successful immunisation against filarial and other pathogens includes DNA vaccination, the work described here provides an approach to define the antigens and modulators most likely to generate very high levels of protection against all stages of infection, along with the ability to define the best correlates of protection.
|
10.1371/journal.pgen.1001318 | A Population Genetic Approach to Mapping Neurological Disorder Genes Using Deep Resequencing | Deep resequencing of functional regions in human genomes is key to identifying potentially causal rare variants for complex disorders. Here, we present the results from a large-sample resequencing (n = 285 patients) study of candidate genes coupled with population genetics and statistical methods to identify rare variants associated with Autism Spectrum Disorder and Schizophrenia. Three genes, MAP1A, GRIN2B, and CACNA1F, were consistently identified by different methods as having significant excess of rare missense mutations in either one or both disease cohorts. In a broader context, we also found that the overall site frequency spectrum of variation in these cases is best explained by population models of both selection and complex demography rather than neutral models or models accounting for complex demography alone. Mutations in the three disease-associated genes explained much of the difference in the overall site frequency spectrum among the cases versus controls. This study demonstrates that genes associated with complex disorders can be mapped using resequencing and analytical methods with sample sizes far smaller than those required by genome-wide association studies. Additionally, our findings support the hypothesis that rare mutations account for a proportion of the phenotypic variance of these complex disorders.
| It is widely accepted that genetic factors play important roles in the etiology of neurological diseases. However, the nature of the underlying genetic variation remains unclear. Critical questions in the field of human genetics relate to the frequency and size effects of genetic variants associated with disease. For instance, the common disease–common variant model is based on the idea that sets of common variants explain a significant fraction of the variance found in common disease phenotypes. On the other hand, rare variants may have strong effects and therefore largely contribute to disease phenotypes. Due to their high penetrance and reduced fitness, such variants are maintained in the population at low frequencies, thus limiting their detection in genome-wide association studies. Here, we use a resequencing approach on a cohort of 285 Autism Spectrum Disorder and Schizophrenia patients and preformed several analyses, enhanced with population genetic approaches, to identify variants associated with both diseases. Our results demonstrate an excess of rare variants in these disease cohorts and identify genes with negative (deleterious) selection coefficients, suggesting an accumulation of variants of detrimental effects. Our results present further evidence for rare variants explaining a component of the genetic etiology of autism and schizophrenia.
| Genome-wide interrogation approaches for mapping genes often are designed to detect the common variants associated with common phenotypes or disease, generally leaving rare variants undetected or untested [1]–[5]. The Rare Allele–Major Effects (RAME) model postulates that rare (minor allele frequency <0.01 to 0.05) penetrant variants are key to the genetic etiology of common disease. Functional mutations that lead to an altered amino acid are often deleterious and potentially disease causing. Such mutations are subjected to natural selection, which either removes these alleles from the population or maintains them at low frequencies relative to the neutral expectations [3], [4], [6]–[9]. Partial genome or candidate gene resequencing of a large number of individuals holds the promise of finding both common and rare variants associated with clinically relevant phenotypes [3], [4]. While a number of studies have mapped common mutations or structural variants associated with neurological disorders like Autism Spectrum Disorder (ASD) (e.g. [10]) and Schizophrenia (e.g. [11]), the RAME model is better suited for modelling diseases that occur sporadically in families, and for testing whether rare variants contribute to these disorders.
ASD is a neurodevelopmental disorder characterized by stereotyped and repetitive behaviours and impairments in social interactions. Schizophrenia is a chronic psychiatric syndrome characterized by a profound disruption in cognition, behaviour and emotion, which begins in adolescence or early adulthood. The incidence of both ASD and schizophrenia is higher in males than in females [12], [13], which points to an important role of X-chromosome genes in the two diseases. There is significant clinical variability among ASD and schizophrenia patients, suggesting that they are etiologically and genetically heterogeneous. For ASD, genetics clearly plays an important role in the etiology, as revealed by twin and familial studies [14]–[16]. Some susceptibility regions have been identified through whole genome linkage analyses [17], [18], although they rarely coincide among the different studies [19]. Additionally, de novo mutations have been observed [20], [21] and are candidate variants in sporadic cases of either ASD or schizophrenia, but they explain a small percentage of the phenotypic variation. Together, these observations suggest that disruption of numerous genes by rare yet penetrant mutations could represent a major cause of ASD. Schizophrenia has an estimated heritability of 80% [22], and it has been recently associated with common variants at the MHC locus [11], [23], [24]. There are also several observations suggesting a causal link between rare mutations and schizophrenia, such as the fact that patients mostly have no affected close relatives, or the associations of both paternal age and decreased fertility with the disease [2]. Recent studies also reported de novo copy number variants (CNVs) in schizophrenia, providing further support for the rare variant hypothesis in non-familial cases [25]–[28], as well as an excess of rare inherited CNVs in familial cases [29].
We hypothesized that capturing genetic variation at low frequencies (minor allele frequency ≥0.005) in a large set of genes expressed in the brain will significantly contribute to our understanding of the genetic basis of ASD and schizophrenia. If the RAME model is relevant to these two diseases, the expectation is an enrichment of rare deleterious mutations among individuals diagnosed with ASD and schizophrenia. Here, we use population genetic and other statistical methods to analyze a resequencing dataset to map genes associated to these two disorders. Using this approach, we identified candidate genes by testing for genes harboring an excess of rare missense variants among individuals affected with ASD and schizophrenia, testing for selection at the gene level in each disease cohort, and assessing the impact of the variants found in candidate genes on the study-wide distribution of missense allele frequencies for each disease cohort.
We resequenced 408 selected brain-expressed genes (Table S1) in 142 ASD and 143 schizophrenia-affected individuals [21]. The ASD and schizophrenia cohorts had global ethnicity representation yet were predominantly European with a large French Canadian sub-group. It is crucial to exclude ethnic and genetic outliers when analyzing rare variants because such samples contain private alleles from other populations. While the results from the software structure [30] revealed no population structure, we identified and removed potential ethnic outliers from the analysis using self reported ethnicity and a principle component analysis using eignesoft [31] (see Materials and Methods, Figure S1). The result was a sample of individuals of European ancestry: ASD (n = 102) and schizophrenia (n = 138). A total of 285 samples from the Québec Newborn Twin Study (QNTS) [32] were screened for self reported European ancestry (n = 240) and were used as controls. Thirty-eight (19 autosomal, 19 X-linked) of the 408 brain expressed genes were sequenced in the QNTS controls, including any gene with de novo mutations previously described in one of the disease cohorts [21] or with potential protein disrupting mutations.
We identified a total of 5,396 segregating sites in the disease cohorts, including 1,111 missense and 11 nonsense variants (Table 1 and Table 2), from lymphoblastoid cell DNA. As expected, there was a reduction in nucleotide diversity (π) on the X chromosome relative to autosomes by a ratio of 0.76 (Table 1), consistent with neutral expectations for the reduced effective population size of the X chromosome [33]. The ratio of the male to female population mutation rate [34] was estimated to be 6.37, slightly higher but similar to previous estimates of the male mutation rate being four times the female mutation rate [35]. The population mutation rate (θW per base pair) for all variants (including nonsense mutations) was 3.471×10−4 and 3.47×10−4 for Schizophrenia and ASD, respectively (Table 1).
The fact that both ASD and Schizophrenia are highly heritability and that the common variant associations have been difficult to replicate support rare variants as an important component of the genetic etiology of these diseases. If rare variants are contributing to these diseases, we expect the site frequency spectrum (SFS) of missense variants to show an excess of deleterious, low frequency variants in our disease cohorts relative to either neutral expectations or controls. We analyzed the proposed detrimental effects of missense variants by estimating the potential functional effect of the missense variants observed using the software mapp [36]. Mapp was used to predict the severity of a missense mutation based on conservation in a multispecies protein alignment and the physiochemical properties of the amino acids. Severity scores indicated that 19% of the missense variants, in 47% of the genes with one or more missense variants, were likely to adversely affect protein function, considering a threshold for mapp scores of ten (approximately P = 0.01). Mapp scores are significantly higher for rare versus common variants (average 7.59 vs. 5.91, Mann-Whitney test P = 1.45×10−4, Figure S2), and the proportion of variants with high mapp scores (>10) is also significantly higher in the rare variants (21%) than in the common variants (11%) (χ2 = 8.3, P = 0.0039). These results are consistent with the presence of deleterious alleles being maintained at low frequency by selection. However, we did not observe an excess of high mapp scores in the cases relative to those in the QNTS controls indicating that rare mutations are likely subject to selection across all populations and not just cognitive related genes screened among our ASD and SCZ patients.
We applied three different methods to identify genes harboring an excess of rare missense variants in the disease cohorts. First, we identified genes with an excess of missense variants relative to the number of silent variants observed. For each gene, we used Fisher's Exact Test to test the ratio of missense to silent variants observed within the gene relative to the ratio of missense to silent variants found in the remaining pooled genes within each cohort, and separately for X-linked and autosomal loci. The ratio of missense to silent variants allows us to control for demographic effects on variation in each cohort and detect genes subject to natural selection. In both cohorts, the autosomal gene MAP1A exhibited a significant excess of missense compared to silent variants (Figure S3, Table S2) (ASD P = 0.04, schizophrenia P = 0.03 after Bonferroni correction) and 23 of the 29 missense variants described in this gene are rare (minor allele frequency <0.03). MAP1A also exhibits an excess of missense variants when compared to the total counts of missense and silent variants in the control cohort (ASD-Control P = 0.009, Schizophrenia-Control P = 0.008). To further test the total predicted effect of the missense variants; for each gene, we summed the mapp scores for all missense variants within a gene, calculated the ratio of summed mapp scores to the number of silent variants, and compared these relative the same ratio in all other genes combined, separately for autosomal and X-linked loci. While the ratio for MAP1A was not significant after multiple testing correction, GRIN3A remained significant, relative to all other genes (P = 0.011) for the ASD cohort.
Second, we tested for an excess of individuals bearing rare missense variants, using Li and Leal's collapsing method [37]. We 1) contrasted the ASD and schizophrenia cohorts to each other (n = 277 genes with one or more missense variant in at least one cohort), 2) compared ASD to the QNTS controls (n = 26 genes with one or more missense variant), and 3) compared the schizophrenia cohort to the QNTS controls (n = 26 genes with one or more missense variant), and corrected for multiple testing. For every gene, the number of individuals carrying at least one rare missense variant, and the number of individuals without rare missense variants, is compared between two cohorts (see Materials and Methods). Although the ASD–Schizophrenia comparison revealed no significant results for any gene, we observed an excess of individuals with rare missense variants for two genes in the ASD cohort relative to the controls (GRIN2B and CACNA1F, Bonferroni adjusted P = 0.026, and P = 0.031 respectively, Table S3). The schizophrenia cohort had a significant excess of individuals hosting rare missense variants at GRIN2B (P = 0.041). We repeated this analysis considering only missense variants predicted to have detrimental effects on protein function (mapp scores >10). We found GRIN3A to exhibit an excess of individuals hosting rare detrimental missense variants in the ASD cohort relative to QNTS controls (P = 0.034).
Finally, we used an extension of the McDonald Krietman test, implemented in mkprf [38], to obtain estimates of γ per gene, based on our observed polymorphisms within humans and substitutions between humans and an outgroup (Pan troglodytes, see Materials and Methods). Estimates of the population selection parameter (γmkprf = 2Nes, where Ne is the effective population size and s is the selection coefficient) will be negative when amino acid replacements are deleterious [39]. Here, we compared the per gene selection coefficients between cohorts to detect genes enriched for missense variants in one cohort compared to another. The overall distribution of γmkprf values among genes was similar between ASD and schizophrenia (Figure 1). When we contrasted ASD and schizophrenia γmkprf estimates to those estimated from a Western-European population [39], we found both the ASD and schizophrenia cohorts have significantly more negative γmkprf distributions (Wilcoxon paired test, ASD P = 0.0026, schizophrenia P = 0.0005), while the QNTS controls do not show significant differences relative to the Western-European dataset (P = 0.18). We also found individual genes that differed significantly among our cohorts. For example, we observed a significant difference in γmkprf between the QNTS and ASD cohorts for CACNA1F (Figure 1B), with the disease cohort having the lower γmkprf estimate.
Having identified a number of individual genes with an excess of deleterious rare alleles, we examined the contribution of these individual genes to estimates of the population selection parameters, γprfreq, across all variants within each cohort. Estimates of γprfreq will be negative when an excess of low frequency variants are observed, suggesting an accumulation of deleterious variants [39] through negative selection. We used the Poisson Random Fields method implemented in prfreq [40] to infer the demographic and selection parameters from cohort specific site frequency spectrum (SFS) and ask 1) whether the SFS of the missense variants showed evidence of selection relative to the SFS of silent and intronic variants; 2) if these selection estimates significantly differed between the disease cohorts and the control cohort; and 3) how removing the disease associated genes mapped with other approaches (see above) contribute to the SFS γprfreq estimates.
First, we estimated the demographic parameters for the different cohorts using the silent and intronic variants discovered among all autosomal genes. Only genes having at least one or more variant (265 for disease cohorts, 19 for controls) were included. Next, while fixing the estimated demographic parameters, the population selection parameters (γprfreq) were then estimated from the missense variants SFS (198 genes with one or more missense variant in the disease cohorts, 15 such genes for the controls; Table S4 and Figure S4). In each case, we estimated the likelihood for that model including the likelihood for a Wright-Fisher neutral model. The demographic-only model was a significant improvement relative to the neutral model (two times the difference in log likelihood, P<0.001), and the demographic-with-selection model was a significant improvement relative to the demographic-only model (P<0.001)(Table S4). The estimates of γprfreq were more negative in the two disease cohorts (γprfreqASD = −920, and γprfreqSCZ = −1,100, Table S4 and Figure S4) than for the control samples (γprfreqcontrol = −740), Specifically, the γprfreqSCZ estimate was significantly more negative than the γprfreqcontrol estimate (P = 0.01), although this pattern was not observed for ASD (P = 0.12) (see Materials and Methods). These observations indicate an excess of low frequency missense variants in our disease cohorts.
To assess each gene's contribution to the overall SFS, we developed an empirical distribution of γprfreq estimates by removing each gene from the SFS and re-estimating γprfreq. For ASD, when excluding MAP1A, GRIN3B and RPGRIP1, either individually or combined, the γprfreq values became more positive (implying less deleterious effects; γprfreqASD = −880, P[γ≥−880| empirical distribution] = 0.015 and −800 respectively) than when estimated for all variants (γprfreqASD = −920) (Figure 2). These values are also closer to the γprfreq value estimated in controls (γprfreqcontrol = −740, see above) and are at the most positive end of the empirical distribution of γprfreq values (Figure 2A). In the schizophrenia cohort, we estimated γprfreq values when removing MAP1A and GRIN3B were γprfreqSCZ = −1,020 (P[γ≥−1020 | empirical distribution] = 0.01) and γprfreqSCZ = −980 (P[γ≥−920 | empirical distribution] = 0.005), respectively (Figure 2B). When variants in both the MAP1A and GRIN3B loci were excluded from the SFS, the estimated γprfreqSCZ increased to −880, again a value similar to that observed in controls. These results indicate that the presence of a few genes, enriched with rare missense variants in the disease cohorts, is enough to alter the global SFS among all of our candidate genes. In our case, the overall excess of deleterious missense variants and the more negative γprfreq estimate in the two disease cohorts as compared to a neutral cohort is mainly caused by a very few candidate genes.
Using approaches that specifically analyze both the function and accumulation of rare variants in disease cohorts, we have found a number of candidate genes associated with ASD and schizophrenia in two cohorts with sample sizes substantially smaller than those required for GWA studies. Population genetic approaches are designed to deal with data where sample sizes are limited compared to the sizes required in GWA studies. In both the ASD and schizophrenia cohorts, GRIN2B and MAP1A contain a statistically significant excess of rare missense variants, while the excess of rare missense variants in CACNAF1 was restricted to the ASD cohort. The involvement of a particular gene (e.g. GRIN2B or MAP1A) in the etiology of both disorders may reflect a critical role for neurodevelopment processes and suggests a pleiotropic effect [41]. Population genetic models incorporating demographic and selection processes, implemented in prfreq, corroborated this result for MAP1A and CACNA1F, and identified three new candidate genes: GRIN3B, NOS1, RPGRIP1 (Table 3).
Most of our mapped genes (Table 3, Table S5) have been previously implicated in neurological disorders or in neurodevelopment. MAP1A, a member of the microtubule-associated MAP1 proteins family, is predominantly expressed in adult neurons and is involved in axon and dendrite development. Among other interactions, MAP1A participates in the linking of DISC1 to microtubules [42]. DISC1 is a protein that was described as related to the pathogenesis of schizophrenia by linkage analysis of a Scottish family [43], [44] and confirmed among Finnish cohorts [45], [46]. The association between schizophrenia and eleven genes interacting with DISC1 (including MAP1A) was explored in Finish families, with significant results in three of the candidates but not for MAP1A [47]. Among calcium channel classes of genes, such as CACNA1F, we found a significant excess of rare variants and two segregating inframe indels at CDS position 807 falling in a glutamic acid-rich coiled domain. CACNA1F has been previously associated to schizophrenia [48] and mutations have also been described for other neurological disorders [49]. Finally, two independent meta-analyses corroborate our findings for a role of GRIN2B in the etiology of schizophrenia [50], [51]. GRIN2B codes a subunit of the glutamate and N-methyl-D-aspartate (NMDA) receptor. There exist several NMDA receptors that are constructed with one or more isoforms of the NR1 subunit (GRIN1) in different combinations with NR2 (GRIN2A, GRIN2B, GRIN2C and GRIN2D genes) and NR3 subunits (GRIN3A and GRIN3B) [52]. Some studies have suggested a relationship of decreased expression and abnormalities in NMDA receptors with schizophrenia [53], [54]. Additionally, we observed key results for other NMDA related genes. GRIN3B had an excess of detrimental variants, two different analyses revealed significantly higher mapp scores in GRIN3A in the ASD cohort, while we observed a nonsense mutation in both GRIN2C and GRIN3A in our cohorts. In the case of GRIN2B, we also observed a coding indel which results in an amino acid insertion at cds position 1,353. This collection of rare and functional changes in the GRIN gene family points to an important role of NMDA receptors in these neurological disorders.
For some cases in our study, different methods identified different loci associated with ASD and Schizophrenia. Observing signficant disease associations for the same genes using all the methods employed would be ideal, however it is not expected as each statistical test evaluates a different component of variation within the data. Furthermore, the ability to use each test varies by chromosome (X vs. autosomal) and cohort. The Li and Leal collapsing method [37] evaluates the accumulation of rare missense variants in genes in cases relative to controls, and is independent of silent variants. The ratio of missense to silent mutations is cohort specific and is used to identify genes that have an excess of missense mutations conditional on the number of silent mutations, relative to the cohort-wide average. As an example, a gene can have a minimum of two rare missense variants and generate a significant result with the collapsing method, but might not have sufficient power to test the ratio of rare missense to silent variants.
In contrast, we do expect genes with high missense to silent variant ratios, to have negative γmkprf estimates, and to contribute to more negative γprfreq estimates overall. We observed these patterns for the GRIN2B gene (Table 3) in both diseases which was supported by a significant collapsing method test result. Like GRIN2B, CACNA1F also shows concordance among methods within the ASD cohort; including a nominally significant missense to silent variant ratio, a significant collapsing method result, and a negative shift in the γmkprf estimate relative to the control cohort. Finally, MAP1A has a significant ratio of missense to silent variants relative to the rest of the genes in both cohorts, and appears to have a signficant impact on the overall γprfreq estimates for both the ASD and SCZ cohorts. Due to limited resequencing data in controls, we were unable to apply the collapsing method to MAP1A. In conclusion, for GRIN2B, MAP1A, and CACNA1F, we see concordant results among multiple methods when data availability allows testing by the different methods.
Previous studies demonstrate that genes associated with Mendelian disease or cancer have more negative population selection parameters compared to genes implicated in complex diseases [55]. This pattern may be explained by a late-onset effect, or by a potential enrichment of positively selected genes, among the loci involved in complex disease [55], [56]. Another explanation is that genes which accumulate either rare inherited or de novo mutations are also more likely to accumulate rare mutations which individually have either a high impact, as in the case of Mendelian disorders, or have an intermediate impact in aggregate.
In this paper we hypothesized that rare variants in neurologically expressed genes are enriched in ASD and schizophrenia cohorts. Our findings support a rare allele-major effect model as we have uncovered significant excess of rare variants in our disease cohorts. It remains an open question if ASD and schizophrenia are caused by variants found in a reduced set of genes such as DISC1 or NMDA receptor related genes, or in a larger number of genes associated to the same functional class or pathway/network.
In total, 408 genes were selected for sequencing: 122 in the X chromosome and 286 autosomal genes [21] (Table S1) from a comprehensive list of potential synaptic genes (n = 5,000) based on published studies and databases [57]–[62]. X-chromosome synaptic genes were chosen due to the excess of affected males as compared to females in individuals affected with schizophrenia [12] and ASD [13], and since many genes affecting neurodevelopmental brain diseases happen to be on the X-chromosome [63]. Autosomal genes implicated in synapse function, including those encoding glutamate receptors and their interactors were also chosen, because glutamate signalling is strongly implicated in synapse function [64]. A total of 38 genes (19 autosomal and 19 X-linked) with de novo mutations [21] or potentially disrupting protein mutations in the disease cohorts were chosen for sequencing in the controls.
DNA samples were available for all affected and unaffected individuals and parents. For certain individuals where blood DNA was limited, we used DNA isolated from an Epstein-Barr Virus transformed lymphoblastoid cell line derived from the individual for the screen. The ASD cell lines samples have been frozen or regrown a maximum of two times. Genomic DNA was extracted from peripheral blood lymphocytes for each individual using Puregene extraction kits (Gentra System, USA). In all cases, rare variants were confirmed by sequencing both parents and using blood-derived DNA to rule out variations having arisen during production or growth of the lymphoblastoid cell line. Parentage was tested using 17 microsatellite markers. Primers were designed using the Exon Primer program from the UCSC genome browser. PCR products were sequenced at the Genome Quebec Innovation Centre in Montreal, Canada (www.genomequebecplatforms.com/mcgill/) on a 3730XL DNA Analyzer System. PolyPhred (v6.0) and Mutation Surveyor (v3.10, Soft Genetics Inc.) were used for mutation detection analysis. Initial screens were done on cell-line DNA from samples to conserve blood sample DNA. PolyPhred (v6.0) scores of 40 or higher were used as the threshold cut-off for all sequencing reads. When reads did not meet this criterion they were resequenced. Chromatograms for all rare variants (singletons or homozygous doubletons) were manually checked. The sensitivity of singleton detection was previously assessed by contrasting the heterozygote calls from the sequencing to heterozygote calls made from Affy SNPChip genotyping, obtaining an estimated false negative rate of heterozygote calls of ∼2% [21]. Variants passing all validation steps were retained for analysis, resulting in a high confidence dataset.
Structure [30] was used to analyze the degree of population structure. Analysis of all samples showed no significant population structure as a large proportion of samples were of the same ethnicity. Principal Component Analysis (PCA) is more susceptible to samples with excess of private alleles and revealed genetic variability between individuals. We removed samples of non-European ancestry (self-reported ethnicity, ethnic outliers) and used eigensoft [31] to identify and remove remaining genetic outliers (defined below). All autosomal variants excluding those with calls in less than 20 per cohort were used for PCA analysis (4,645 SNPs). We used the linkage disequilibrium correction and calculated the top 10 principal components (PCs) and removed individuals with PC projections greater than two standard deviations or more from the mean, for all significant principal components, using 10 iterations. Individuals exhibiting excess of rare variants genome-wide are likely to be genetic outliers and readily identifiable with PCA and removed, while individuals with an excess of rare variants in specific genes are retained. To ensure the PCA outliers were true outliers, we used a one-sided t-test to assess if the proportion of missense singletons in each PCA outlier was higher than in PCA retained samples across all autosomal genes (see Text S1). Structure (admixture model) was used to reassess levels of population structure within the final sample set (Figure S1C).
Mapp was used to predict severity and assign scores to each missense variant. Orthologous protein sequences were obtained using the Galaxy Browser (http://main.g2.bx.psu.edu/) and the UCSC Human Genome Browser (hg18, http://genome.ucsc.edu/cgi-bin/hgGateway) to generate columns of aligned orthologous amino acids. Mapp scores and P values were calculated as shown in Stone and Sidow [36]. Mapp assesses variation observed at each amino acid position with respect to six physicochemical properties (hydropathy, polarity, charge, side-chain volume, free energy in alpha-helical conformation, and free energy in beta-sheet conformation) after weighting each protein sequence to mitigate the influence of phylogeny.
Fisher's exact test was used to detect deviations in the missense to silent variant ratio within genes. For each gene, the ratio of missense (or rare missense) to silent variants was contrasted to the same ratio in all remaining genes. This analysis was conducted in the ASD and schizophrenia cohorts testing autosomal and X-linked genes separately. We used the Bonferroni correction for multiple tests (n = 277). Excess of predicted deleterious load within genes was evaluated by summing the mapp scores for all missense variants within a gene and testing the ratio of summed mapp scores to silent variants within the gene relative to all other genes. This was done separately for autosomes and X-linked genes, and separately for each disease cohort.
To identify genes with an excess of individuals bearing rare missense variants, we used Li and Leal's collapsing method [37]. For ASD vs schizophrenia, ASD vs QNTS, and schizophrenia vs QNTS controls, the number of individuals with at least one rare missense mutation and the number of individuals with no rare missense variants was determined for each cohort, and these counts made up the cells of the two by two table. We assessed statistical significance using Fisher's Exact test and used Bonferroni's correction for multiple tests (number of genes nASD-SCZ = 277, nASD – QNTS Controls = 26, nSCZ- QNTS Controls = 26). This analysis was repeated, considering only missense variants with a mapp score >10.
|
10.1371/journal.pntd.0006217 | Serum metabolome changes in adult patients with severe dengue in the critical and recovery phases of dengue infection | Dengue virus (DENV) is the most prevalent arbovirus leading to an estimated 100 million symptomatic dengue infections every year. DENV can cause a spectrum of clinical manifestations, ranging from mild dengue fever (DF) to more life threatening forms such as dengue hemorrhagic fever (DHF). The clinical symptoms of DHF become evident typically at the critical phase of infection (5–7 days after onset of fever), yet the mechanisms that trigger transition from DF to DHF are not well understood. We performed a mass spectrometry-based metabolomic profiling of sera from adult DF and DHF patients at the critical and recovery phases of infection. There were 29 differentially expressed metabolites identified between DF and DHF at the critical phase. These include bile acids, purines, acylcarnitines, phospholipids, and amino acids. Bile acids were observed up to 5 fold higher levels among DHF compared to DF patients and were significantly correlated to the higher levels of aspartate transaminase (AST) and alanine transaminase (ALT), suggestive of liver injury among DHF. Uric acid, the most abundant antioxidant in the blood, was observed to be 1.5 fold lower among DHF compared to DF patients. This could result in decreased capacity of endogenous antioxidant defense and elevated oxidative stress among DHF patients. In the recovery phase, the levels of eight metabolites were still significantly higher or lower among DHF patients, including chenodeoxyglycocholic acid, one of the bile acids observed at the critical phase. This indicates potential prolonged adverse impact on the liver due to DENV infection in DHF patients. Our study identified altered metabolic pathways linked to DHF in the critical and recovery phases of dengue infection and provided insights into the different host and DENV interactions between DF and DHF. The results advance our understanding on the mechanisms of DHF pathogenesis, alluding to possible novel therapeutic targets to dengue management.
| Dengue is a re-emerging disease caused by dengue viruses (DENV) and half of the world’s population is now at risk of dengue infection. DENV can cause a spectrum of clinical manifestations ranging from mild dengue fever (DF) to the potentially lethal dengue hemorrhagic fever (DHF). The molecular mechanisms that trigger transition from DF to DHF are not well understood. To gain insights on the mechanisms of DHF pathogenesis, we performed metabolomic profiling of sera from adult DF and DHF patients at the critical phase of infection (5–7 days after onset of fever), the time point when the clinical symptoms of DHF become more evident. 29 differentially expressed metabolites between DF and DHF were identified, allowing us to discover a variety of metabolite pathways linked to DHF. They include bile acid biosynthesis, fatty acid β-oxidation, purine and pyrimidine metabolism, phospholipid catabolism, tryptophan and phenylalanine metabolism, and so on. The results advance our understanding on the mechanisms of DHF pathogenesis, and certain altered pathways being harnessed as therapeutic targets to alleviate DHF. These differentially expressed metabolites have the potential as biomarkers for disease monitoring and evaluation of therapeutic interventions.
| Dengue is a re-emerging disease caused by four closely related serotypes of dengue viruses (DENV) and it is endemic in the tropical and sub-tropical regions of the world. An estimated 100 million symptomatic dengue infections occur annually, with over half of the world’s population living at the risk of infection [1]. Infection with DENV can cause a spectrum of clinical manifestations ranging from mild dengue fever (DF) to the potentially lethal dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS), which are characterized by abnormal hemostasis, vascular leakage and liver damage [2]. As yet, there is no specific treatment for dengue and the management of DHF patients is primarily supportive. Furthermore, the lack of an appropriate animal model poses great challenges in the study of dengue pathogenesis.
Dengue is a dynamic disease. After the incubation period, the illness begins abruptly and is followed by the three phases—febrile (day 0–4 post onset of fever), critical (day 5–7 post onset of fever) and recovery (day 21–28 post onset of fever). DF patients would recover uneventfully after 5–7 days of acute illness, but for DHF patients, the initial febrile period is followed by a rapid onset of vascular leakage, thrombocytopenia and hemorrhage at the critical phase. The continual loss of intravascular volume from plasma leakage can rapidly lead to hypotension and circulatory failure in DHF patients, which can result in death. The pathological difference between DF and DHF suggest differential virus-host interaction in the susceptibility to the disease, and both viral and host immune factors are likely involved. Although numerous efforts have been made in the last decade [3–6], the mechanisms that trigger transition from mild DF to more life threatening DHF at the critical phase are not fully understood, hampering the design of effective treatments for DHF.
Metabolomics is a rapidly emerging field in systems biology that refers to the global investigation of metabolite pool in biological systems in response to biological stimuli or perturbations [7]. Metabolites are the end products of cellular regulatory processes and form a direct link between molecular changes and phenotypes, and by providing a snapshot of an organism’s metabolic status, metabolomics holds the promise of finding metabolites or metabolic pathways related to disease processes [8,9]. It has been applied to infectious diseases to study host-pathogen interactions, including DENV infections [10–13]. In our previous study, we investigated serum metabolome difference between adult DF and DHF patients enrolled from a prospective dengue study at early febrile phase of DENV infection and found potential marker metabolite to predict DHF at early phase of the infection [14]. In the present study, metabolomics investigation was conducted on the same patient cohort at both the critical phase and recovery phases, with an aim to study the potential mechanisms that may contribute to the transition from DF to DHF at the critical phase of DENV infection. Interestingly, the majority of identified differentially expressed metabolites between DF and DHF in the febrile phase were different from those identified in the critical phase, showing that metabolomics study could capture the distinct molecular changes in DF and DHF patients at different phases of DENV infection. In the recovery phase, the levels of a few metabolites were still significantly altered in DHF patients, indicating prolonged effects of DENV infection on DHF patients. Our results identified metabolic pathways linked to dengue progression and provided insights on the mechanisms of DHF pathogenesis.
Enrollment of all eligible individuals was based on written informed consent and the collected samples were anonymized. The protocols were approved by the Domain Specific Review Board of the National Healthcare Group, Singapore (DSRB/E/2009/432).
The details of patient recruitment, sample collection and the study protocols of the study have been described earlier [14]. Briefly, the study cohort of dengue patients were recruited from the Prospective Adult Dengue Study (PADS), which is a cohort study of acute febrile adults at a tertiary care center, Communicable Diseases Center, Tan Tock Seng Hospital, Singapore. Adult patients (≥ 18 years) presenting with acute onset of fever (≥ 37.5°C) without rhinitis or other clinical alternatives were included in the study (Febrile stage, < 96 hours post onset of fever; Critical, Day 5–7, Convalescence, Day 21–28). Venous blood samples were collected, aliquoted and frozen at -80°C for hematological, virological and serological analysis. The study design, hypotheses, patient characteristics, assay methods, statistical methods and modeling methods were reported as per REMARK which is important for generalizability [15]. DF and DHF patients were classified according to the WHO 1997 and 2009 dengue guidelines [2,16]. To fulfill the case definition of DHF, all four of the following criteria must be present, namely: fever or history of fever, hemorrhagic tendencies, thrombocytopenia and evidence of plasma leakage [16]. Hematoconcentration was determined by the hematological analyzer and expressed as % of the volume of whole blood that was made up of red blood cells. Hematocrit increase of over 20% of the values at convalescence phase is considered a common clinical index of plasma leakage and DHF diagnosis. The PADS cohort comprised of both DF and DHF patients recruited at different phases of dengue infection and DENV2 is the predominant DENV type. Based on the anti-DENV IgG and IgM seropositivity or seronegativitiy, immune status of the patients was determined. Anti-DENV IgM seropositivity and IgG seronegativity indicated that 40–52% of DF and 22–41% of DHF are primary cases, and the remaining are secondary cases. 12–16% DF patients and 59–76% DHF patients had platelets <50×103/μL at any point as determined during their daily routine total blood count.
A detailed hematological and virological analysis was performed and included white blood cell count (WBC), red blood cell count (RBC), blood hemoglobin (HGB), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), platelet count (PLT), lymphocyte percentage (LYMPH%), lymphocyte count (LYMPH), mixed cell count (MXD), neutrophil percentage (NEUT%), neutrophil count (NEUT), red blood cell distribution width-coefficient of variation (RDW-CV), and quantitation of peripheral viral titers using reverse transcriptase-polymerase chain reaction (RT-PCR) crossover values (Ct). Dengue viral infection was confirmed by RT-PCR [17], or NS1 detection by Dengue NS1 Ag Strip (Bio-Rad, Marnes-la-Coquette, France) at the Environmental Health Institute, Singapore, or typing by virus isolation and immunofluorescence using DENV type-specific monoclonal antibodies (ATCC: HB46-49). Dengue-immune status (primary or secondary DENV infection) was based on Dengue IgG levels in the acute sera, using a commercially obtained ELISA (PanBio, Brisbane, Australia) according to the manufacturer’s protocol.
A volume of 50 μL from each serum sample was thawed at 4°C and serum proteins were precipitated with 200 mL ice-cold methanol, which contained 10 mg/mL 9-fluorenylmethoxycarbonyl-glycine as an internal standard. After vortexing, the mixture was centrifuged at 16,000 rpm for 10 minutes at 4°C and the supernatant was collected and evaporated to dryness in a vacuum evaporator. The dry extracts were then redissolved in 200 μL of 98:2 water/methanol for liquid chromatography-mass spectrometry (LC-MS) analysis. Quality control (QC) samples were prepared by mixing equal amounts of serum samples from all the samples and processed as per other samples. The QC sample was run after each 8 samples to monitor the stability of the system and all samples were randomized.
Metabolomics analysis was performed as previously described [14]. Briefly, the supernatant fraction from sample preparation step was analyzed using Agilent 1290 ultrahigh pressure liquid chromatography system (Waldbronn, Germany) equipped with a 6550 QTOF mass detector managed by a MassHunter workstation. The column used for the separation was an Agilent rapid resolution HT Zorbax SB-C18 (2.1×100 mm, 1.8 mm; Agilent Technologies, Santa Clara, CA, USA). The oven temperature was set at 45°C. The gradient elution involved a mobile phase consisting of (A) 0.1% formic acid in water and (B) 0.1% formic acid in methanol. The initial condition was set at 5% B. A 7 min linear gradient to 70% B was applied, followed by a 12 min gradient to 100% B which was held for 3 min, then returned to starting conditions over 0.1 min. Flow rate was set at 0.4 ml/min, and 5 μL of samples was injected. The electrospray ionization mass spectra were acquired in both positive and negative ion mode. Mass data were collected between m/z 100 and 1000 at a rate of two scans per second. The ion spray voltage was set at 4,000 V, and the heated capillary temperature was maintained at 350°C. The drying gas and nebulizer nitrogen gas flow rates were 12.0 L/min and 50 psi, respectively. Two reference masses were continuously infused to the system to allow constant mass correction during the run: m/z 121.0509 (C5H4N4) and m/z 922.0098 (C18H18O6N3P3F24).
Raw spectrometric data in untargeted metabolomics were analyzed by MassHunter Qualitative Analysis software (Agilent Technologies, US) and the molecular features characterized by retention time (RT), chromatographic peak intensity and accurate mass, were obtained by using the Molecular Feature Extractor algorithm. The features were then analyzed by MassHunter Mass Profiler Professional software (Agilent Technologies, US). Only features with an intensity ≥ 20,000 counts (approximately three times the limit of detection of our LC-MS instrument), and found in at least 80% of the samples at the same sampling time point signal were kept for further processing. Next, a tolerance window of 0.15 min and 2 mDa was used for alignment of RT and m/z values, and the data normalized to spiked 9-fluorenylmethoxycarbonyl-glycine internal standard. Raw spectrometric data in targeted metabolomics were processed using MassHunter Workstation Quantitative Analysis software (Agilent Technologies, US).
For statistical analysis, nonparametric Mann–Whitney Test with Benjamini-Hochberg Multiple Testing Correction was employed, and statistical significance was set at p<0.05. For multivariate data analysis using principle component analysis (PCA) or Orthogonal projections to latent structures discriminant analysis (OPLS-DA), data were normalized by median-centering and dividing by standard deviation. PCA and OPLS-DA were performed using the software package SIMCA-P 13.0 version (Umetrics AB, Umea, Sweden). Metabolites with Variable Importance in the Projection (VIP) values>1 were considered to be influential for the separation of samples in OPLS-DA analysis. In addition, the fold change (FC) analysis was also performed to further filter the features and only those features with FC > 1.5 were selected as potential significantly altered metabolites. The hierarchical cluster analysis (HCA), a cluster analysis method which seeks to build a hierarchy of clusters, was performed using MeV4.0.
The structure identification of the differentially expressed metabolites was based on our published work [18]. Briefly, the elemental compositions of the metabolites were first calculated based on the exact mass, the nitrogen rule and the isotope pattern by Masshunter software from Agilent. Then, the elemental composition and exact mass were used for open source database searching with a cutoff value of 10 parts per million (ppm), including LIPIDMAPS (http://www.lipidmaps.org/), HMDB (http://www.hmdb.ca/), METLIN (http://metlin.scripps.edu/) and MassBank (http://www.massbank.jp/). Next, MS/MS experiments were performed to obtain structural information via the interpretation of the fragmentation pattern of the metabolite. The MS/MS spectra of possible metabolite candidates in the databases were also searched and compared. Finally, the metabolites were confirmed by comparison with the standards where commercially available. The metabolites are listed according to the minimum reporting standards for chemical analysis in metabolomics recommended by Metabolomics Standard Initiative (MSI) [19,20]. Briefly, a four-level system ranging from Level 1 (identified metabolites, e.g. based upon the co-characterization with reference standards) via Levels 2 (putatively annotated compounds, e.g. without chemical reference standards, based upon physicochemical properties and/or spectral similarity with public/commercial spectral libraries) and 3 (putatively characterized compound classes, e.g. based upon characteristic physicochemical properties of a chemical class of compounds, or be spectral similarity to known compounds of a chemical class) to Level 4 (unidentified or unclassified metabolites which can still be differentiated based on spectrum data). For metabolic pathway analysis, MetaboAnalyst was used to identify relevant pathways [21].
We characterized serum metabolome changes between DF (n = 25) and DHF (n = 27) patients at both the critical phase and recovery phase of infection using LC-MS. We first evaluated the stability and reproducibility of the LC-MS method by performing PCA on all the samples including the 6 QC samples [22]. As shown in S1 and S2 Figs, the QC samples are clustered in PCA scores plots of sera (S1 and S2 Figs), indicating good stability and reproducibility of the chromatographic separation of the metabolomics analysis.
In the critical phase, a total of 29 MSI Levels 1 and 2 metabolites, were significantly expressed between DHF and DF patients, were identified. HCA could successively merge similar groups of objects and the objects are joined together in a hierarchical fashion from the closest, that is most similar, to the furthest apart, that is the most different. HCA based on the metabolome profile could segregate DHF and DF patients in the critical phase (Fig 1), where all but two DHF patients were classified together with DF. These differentially expressed metabolites belonged to classes such as bile acid, acylcarnitine, phosphatidylcholine (PC), lysophosphatidylcholine (LysoPC), amino acid and derivative, dipeptide, purine and pyrimidine (Table 1). Compared to DF patients, 10 of the 29 differentially expressed metabolites were increased in DHF patients. Among them, the levels of the bile acids showed about 5 times higher in DHF patients than in DF patients (Fig 2). Furthermore, these bile acids demonstrated positive correlations to AST and ALT levels, indicating they may be associated with liver dysfunction in DHF (Fig 3). The other 19 differentially expressed metabolites, including purines and pyrimidines, most of the acylcarnitines and lipids, were decreased in DHF patients. In the recovery phase, 8 altered MSI Levels 1 and 2 metabolites between DF and DHF patients were structurally identified including bile acid, phospholipids, amino acid, dipeptide, and fatty acid (Table 2). Among these differentially expressed metabolites, the level of the bile acid, chenodeoxyglycocholic acid, was still about 3 times higher in DHF patients than in DF patients, indicating prolonged effects of DENV infection on DHF patients (Fig 4).
To assess their potential in differentiating DF and DHF, Receiver Operating Curve analyses were performed for the differentially expressed metabolites in the critical phase. Several metabolites showed good performing Area Under Curve (AUC), including serotonin (AUC = 0.85, 95% C.I. 0.75–0.95, p<0.0001), uridine (AUC = 0.81, 95% C.I. 0.68–0.93, p = 0.0002), glycoursodeoxycholic acid (AUC = 0.77, 95% C.I. 0.65–0.90, p = 0.0007), uric acid (AUC = 0.76, 95% C.I. 0.63–0.89, p = 0.001) (S3 Fig).
We have previously identified more than 20 differentially expressed metabolites between DF and DHF in the early febrile phase of dengue. Surprisingly, a distinct set of differentially expressed metabolites between DF and DHF were found in critical phase as compared to the early febrile phase. There are only two common metabolites between these two phases and they are serotonin and a dipeptide.—We have shown previously through a stable-isotope dilution mass spectrometry method that serotonin was significantly changed among DHF patients as compared to DF patients in the febrile phase, and serotonin remained significantly altered among the DHF patients in the critical phase. This finding demonstrated the high reliability and confidence of our methodologies and results in our previous and current untargeted analysis approach. In addition, for the three bile acids which were significantly increased among DHF patients as compared to DF patients in the critical phase, their levels was also higher among DHF patients in the febrile phase, but it was not statistically significant. On the other hand, a few fatty acid amides were significantly lower among DHF patients as compared to DF patients in the febrile phase, and their levels became similar between DF and DHF patients in the critical phase.
We used MetaboAnalyst, a pathway analysis tool, to determine the underlying biochemical pathways revealed by the identified metabolites. The perturbed metabolic pathways in DHF in the critical phase included bile acid biosynthesis, fatty acid β-oxidation, purine and pyrimidine metabolism, phospholipid catabolism, tryptophan and phenylalanine metabolism. (S4 Fig) In the recovery phase, the bile acid biosynthesis and phospholipid catabolism pathways were still altered, while other metabolic pathways were no longer disturbed in DHF patients (S5 Fig).
Dengue is a very dynamic disease, and the symptoms of DHF usually do not occur until the critical phase of infection. The critical phase is known as the danger period of dengue because a small proportion of patients may undergo sudden deterioration over 24–48 h. If the patient recovers, there are no sequelae in uncomplicated dengue, symptoms resolve and clinical parameters normalize. The mechanisms that contribute to the transition from DF to DHF are not well understood, although the pathological differences between DF and DHF suggest differential virus-host interactions in the susceptibility to the disease.
In our previous study a systematic characterization of serum metabolome at early febrile phase (< 96 hours post onset of fever) was reported from adult DF/DHF patients enrolled from a prospective dengue study [14]. In the present study, metabolomics investigation was conducted on the same patient cohort at the critical phase (day 5–7 post onset of fever) and recovery phase (day 21–28 post onset of fever) to map the serum metabolome and identify metabolic pathways linked to dengue progression. Interestingly, the majority of identified differentially expressed metabolites between DF and DHF in the critical phase were different from the ones identified in the early febrile phase, showing that metabolomics study could capture the specific molecular changes in DF/DHF patients at different phases of infection. In the recovery phase, the levels of a number of lipids were still significantly altered in DHF patients, indicating prolonged effects of DENV infection on DHF patients. Our results could help to identify metabolic pathways linked to dengue progression and understand the mechanisms of DHF pathogenesis.
We have previously found that the levels of bile acids are higher in DF patients than in healthy controls at febrile and critical phases of infection [13]. In this study, we further showed that DHF patients had even higher levels of bile acids than DF patients at both critical and recovery phases of infection. Bile acids, which are formed in the liver as the end products of cholesterol metabolism, can not only facilitate hepatobiliary secretion of endogenous metabolites and nutrient absorption in the intestine, but also play important roles in regulating glucose and lipid metabolism through bile acid receptors [23]. When present in abnormally high levels, bile acids are highly toxic to cells, and impaired bile acid homeostasis could contribute to the pathogenesis of liver and intestinal diseases [24]. Many studies have implicated that DENV has adverse effect on liver functions in dengue patients and dengue infection is believed to be an important cause of acute viral hepatitis in endemic countries [25,26]. The hepatocytes and Kupffer cells in the liver could be the targets for DENV replication and a number of liver hepatic histological changes have been observed in dengue patients [26–28]. Accumulated bile acids in the hepatocytes could result in mitochondrial damage and may eventually lead to apoptosis or necrosis [29]. It has been shown that DHF patients had a higher risk of developing acute liver failure (p < 0.001) and liver pathology has been viewed as one of the hallmarks of DHF [2, 30]. In the clinic, liver enlargement and elevated transaminases have been used as the clinical features of hepatic histological changes in dengue patients [31], and we indeed found in this study that AST and ALT levels were significantly higher in DHF than in DF. Put together, elevated bile acids among DHF patients and their positive correlations to AST and ALT levels indicated that these bile acids may play an important role in liver pathology in DHF patients. Furthermore, a recent phase II clinical trial showed that INT-747, a potent agonist of bile acid farnesoid X receptor, could improve histological activity and reduce fibrosis in adult patients with non-alcoholic steatohepatitis [32], indicating bile acid receptors may be a potential therapeutic target for liver disease management in DHF.
In this study, decreased uric acid was observed in DHF patients compared to DF patients at the critical phase. Uric acid is an end product of purine catabolism, and numerous studies have shown that uric acid is a major antioxidant in the blood and can help to protect against free-radical oxidative damage [33,34]. It is reported that the total antioxidant capacity in the blood was positively correlated with uric acid concentration, and low circulating levels of uric acid have been associated with disease conditions with a significant redox imbalance due to increased production of oxidant species and decreased capacity of endogenous antioxidant defenses [35,36]. Indeed, increased ROS and NOS have been reported in DHF patients [37]. Furthermore, we have previously found higher levels of phenylalaninein DF patients than in healthy controls. In this study, when compared to DF patients, a further increase of phenylalanine in DHF patients was observed. We have proposed that the accumulation of phenylalanine was due to elevated oxidative stress status in dengue patients where superoxide and peroxynitrite increase oxidation of tetrahydrobiopterin (BH4), thereby depleting the available BH4 pool [38]. BH4 is a co-factor for phenylalanine (4)-hydroxylase (PHA), an enzyme required for converting phenylalanine to tyrosine. Thus, higher level of phenylalanine also indicates increased oxidative stress in DHF patients. Together, decreased uric acid could result in impaired ability to scavenge the ROS and NOS in DHF patients. It had been shown that systemic uric acid administration could increase serum antioxidant capacity in healthy volunteers [39], which might be a potential intervention strategy of the condition of DHF.
In our previous studies, we have shown that anti-inflammatory responses played an important role in modulating pro-inflammatory processes to prevent the development of pathologies by excessive or prolonged inflammation in DF patients, and higher levels of Docosahexaenoic acid (DHA) were found in DF patients than in healthy controls [13]. In current study, a decrease of DHA in DHF patients was observed when compared to DF patients. DHA, an omega-3 polyunsaturated fatty acid, is known for the anti-inflammatory activity by suppressing the production of pro-inflammatory cytokines [40], and the lower level of DHA could indicate impaired anti-inflammatory functions in DHF patients compared to DF patients. We have also found earlier that the levels of many acylcarnitines, essential intermediates of fatty acid β-oxidation, were higher in DF patients than in healthy controls, suggesting disturbed energy metabolism in DF patients [13]. Interestingly, when compared to DF patients, decreased levels of acylcarnitine were found in DHF patients in this study, indicating DHF actually induced less disturbance in fatty acid β-oxidation than DF.
Voge et al had also conducted mass spectrometry-based serum metabolomics studies on non-dengue (ND) controls, DF and DHF patients and identified 13 metabolites that statistically differentiate DHF, DF and ND groups, including amino acids, lipids and vitamins [12]. Interestingly, most of the 13 differentially expressed metabolites were not found in our current study. This is likely due to the different analytical platforms used in the two studies, hydrophilic interaction liquid chromatography versus RP liquid chromatography, leading to different metabolite coverage. Nevertheless, both studies showed that metabolomics approach could identify metabolites associated with distinct disease outcomes of dengue infection.
In summary, our study identified differentially expressed metabolites and the underlying metabolic pathways linked to DHF in both the critical and recovery phases of dengue infection. These results provide insights into the mechanisms of DHF pathogenesis, and certain altered pathways might serve as therapeutic targets to alleviate DHF.
|
10.1371/journal.pntd.0006924 | Laboratory challenges of Plasmodium species identification in Aceh Province, Indonesia, a malaria elimination setting with newly discovered P. knowlesi | The discovery of the life-threatening zoonotic infection Plasmodium knowlesi has added to the challenges of prompt and accurate malaria diagnosis and surveillance. In this study from Aceh Province, Indonesia, a malaria elimination setting where P. knowlesi endemicity was not previously known, we report the laboratory investigation and difficulties encountered when using molecular detection methods for quality assurance of microscopically identified clinical cases. From 2014 to 2015, 20 (49%) P. falciparum, 16 (39%) P. vivax, 3 (7%) P. malariae, and 2 (5%) indeterminate species were identified by microscopy from four sentinel health facilities. At a provincial-level reference laboratory, loop-mediated isothermal amplification (LAMP), a field-friendly molecular method, was performed and confirmed Plasmodium in all samples though further species-identification was limited by the unavailability of non-falciparum species-specific testing with the platform used. At a national reference laboratory, several molecular methods including nested PCR (nPCR) targeting the 18 small sub-unit (18S) ribosomal RNA, nPCR targeting the cytochrome-b (cytb) gene, a P. knowlesi-specific nPCR, and finally sequencing, were necessary to ultimately classify the samples as: 19 (46%) P. knowlesi, 8 (20%) P. falciparum, 14 (34%) P. vivax. Microscopy was unable to identify or mis-classified up to 56% of confirmed cases, including all cases of P. knowlesi. With the nPCR methods targeting the four human-only species, P. knowlesi was missed (18S rRNA method) or showed cross-reactivity for P. vivax (cytb method). To facilitate diagnosis and management of potentially fatal P. knowlesi infection and surveillance for elimination of human-only malaria in Indonesia and other affected settings, new detection methods are needed for testing at the point-of-care and in local reference laboratories.
| In Southeast Asia, Plasmodium knowlesi, a malaria parasite of macaques, was recently discovered to infect humans. This emerging disease is important because it has potential for causing severe disease and death, and it is a threat to malaria elimination efforts in the region. In this report from Aceh Province, Indonesia, where P. knowlesi was only recently discovered, the authors report on the laboratory challenges of distinguishing this species from other human species. Using several different molecular methods, they investigated 41 malaria cases which by microscopy, were initially reported as: P. falciparum (49%), P. vivax (39%), P. malariae (7%), and indeterminate (5%). Only after using a P. knowlesi-specific nPCR method and sequencing, did they find that nearly half were P. knowlesi. Consistent with a sparse literature, a field-friendly molecular method (genus-specific LAMP) reliably detected P. knowlesi, while use of a more standard reference laboratory molecular method (18S rRNA nPCR targeting the four human-only species) missed the infections. Also another reference laboratory molecular method (cytb nPCR) mis-classified P. knowlesi infections as P. vivax due to cross-reactivity. To address the emerging threat of P. knowlesi, new detection methods are needed for point-of-care and reference testing.
| Plasmodium knowlesi is a newly emergent zoonotic human malaria species previously thought to only infect macaques. Since the first report of a human case from Peninsular Malaysia in 1965 [1] and the large cluster of human knowlesi malaria in Sarawak in 2004 [2], endemic cases have been reported from other Asian countries including Brunei, Cambodia, India, Malaysia, Myanmar, Philippines, Singapore, Thailand, Vietnam, Indonesian Borneo [3–5], and more recently Sumatra Island [6, 7].
The identification of P. knowlesi infection is important for clinical and public health reasons. Infection in humans is most often uncomplicated, but 6–9% of symptomatic patients develop severe malaria and 0.3–1.8% of cases die [8–10]. Fatal outcomes have been associated with misdiagnosis of parasite species by microscopy, resulting in delays in appropriate management [11, 12]. From a public health perspective, malaria control programs aim to decrease morbidity and mortality from all Plasmodium species affecting humans. As P. knowlesi infection is associated with a number of different risk factors than infections caused by other Plasmodium species [6, 13] (e.g. forest-related exposures), it may require different interventions. For subnational and national areas aiming to achieve and maintain malaria elimination, or the interruption of local transmission of human-only species, as is the goal in Indonesia, accurate species identification is critical.
In most of Asia, microscopy is the standard for malaria diagnosis and surveillance. However microscopy has recognized limitations in diagnostic accuracy and species identification [14]. For P. knowlesi specifically, different asexual blood stages can resemble P. falciparum and P. malariae, and in routine practice it is misidentified as all human-only species [15]. Therefore, a variety of PCR methods have been utilized to distinguish P. knowlesi from other Plasmodium species [16, 17]. With its simpler requirements and faster turnaround time, loop mediated isothermal amplification (LAMP), another nucleic acid-based detection method, may be a more practical alternative in resource-limited field settings [18–20]. However, the relative benefits and limitations of LAMP and the various other PCR methods are not clear, particularly for field settings.
To support malaria elimination efforts in Aceh Province, Indonesia, a pre-elimination area with known endemicity of P. vivax and P. falciparum, we introduced the use of molecular detection for quality assurance of microscopy-identified cases from health facilities by establishing LAMP testing at the provincial level reference laboratory. As previously reported, the finding of indeterminate species triggered further molecular testing that led to the first reported finding of P. knowlesi in Indonesia outside of Borneo [6]. Epidemiological investigation revealed that P. knowlesi infection was associated with forest exposures, particularly overnight stays due to work [6].
In this study, we present the laboratory details of this real-world investigation whereby the use of serial molecular detection methods including LAMP, two nPCR methods, P. knowlesi-specific nPCR, and sequencing led to the identification and confirmation of P. knowlesi infection. Challenges encountered in this experience have relevance to malaria diagnosis and surveillance in other settings where P. knowlesi may be present and can inform research and development of improved P. knowlesi detection methods.
The study was conducted in Aceh Besar District, Aceh Province, Sumatra island, Indonesia, a low-transmission setting that aims to eliminate malaria by 2020. The 2013 incidence of malaria was 0.4/1000, and 68 (39%) of cases were reported as P. vivax, 71 (41%) as P. falciparum, and the remaining 34 unspecified or mixed P. falciparum/P. vivax [6]. The sentinel sites included five primary health centers that reported 78% of all cases reported in Aceh Besar in 2013. During the study period June 2014 to December 2015, 41 patients were diagnosed with microscopy-confirmed malaria and recruited for enrolment. This number of cases was a convenience sample from an umbrella study where health facility-identified cases triggered active case finding in villages [6].
After written consent was obtained and prior to treatment, venous blood was collected and partly used to prepare dried blood spots (DBS) using Whatman 3MM paper. DBS along with remaining whole blood were initially stored at 4°C, transferred to -20°C within a week of collection, and then stored at -80°C. Antimalarial treatment was based on microscopy results and according to Indonesian government’s national policy.
Ethical approval for the study was obtained from the National Institute of Health, Research and Development of the Indonesian Ministry of Health (number LB.02.01/5.2/KE.111/2014 and LB.02.01/5.2/KE.211/2015) and IRB Committee of the University of California, San Francisco. Written informed consent was obtained from all adults or a parent or guardian for participants less than 18 years of age.
For quality assurance of microscopy performed at health centers, blood smears were re-read by certified microscopists at the provincial laboratory according to national guidelines. For further quality assurance at the provincial-level, LAMP was selected due to its field-friendly platform. Initial extraction of DNA and LAMP testing were performed at the provincial laboratory. DNA was extracted from DBS using the Saponin/Chelex method [21]. Pan-LAMP testing followed by Pf-LAMP specific testing for Pan-LAMP positive samples was also performed using the commercially available Loopamp MALARIA Pan/Pf detection kit in accordance to manufacturer’s instructions (EIKEN Chemical, Co., Ltd., Japan). Species identification for non-falciparum species was not available with this LAMP platform, but this was not anticipated to be a problem because Aceh was considered to be endemic for only P. falciparum and P. vivax malaria before the study was launched [22]. As such, Pan-LAMP positive, Pf-LAMP negative samples were expected to be P. vivax.
Further molecular testing was performed at the Malaria Pathogenesis laboratory at the Eijkman Institute in Jakarta, using chelex-extracted DNA from a second DBS. Genus-specific PCR targeting the mitochondrial cytb gene followed by AluI enzyme digestion for species identification of the four main human species was used initially, as previously described [23]. After a report of indeterminate species and suspicion of P. knowlesi by a field microscopist, as well as limited data on the performance of the cytb nPCR method for detection of P. knowlesi, additional methods were employed including nPCR testing targeting the 18S rRNA gene for the four human-only species [24], and P. knowlesi-specific nPCR [16] for all samples. For a proportion of samples testing positive by P. knowlesi specific nPCR, DNA was extracted from whole blood using the QIAamp DNA Mini kit (Qiagen, CA) and Sanger targeted genome sequencing [25] was performed (Eijkman Institute Sequencing Facility). To prevent DNA contamination, all extractions were performed in rooms separate from where amplification was conducted. Extracted DNA was stored at -20°C.
Results from microscopy and each molecular method were compared to a gold standard established through serial molecular testing: P. falciparum and P. vivax classification were based on species-specific positivity by both cytb and 18S rRNA nPCR, and P. knowlesi classification was based on genus-specific PCR positivity by both cytb and 18S rRNA nPCR and P. knowlesi-specific nPCR positivity. With regards to diagnostic performance for species identification, we were not able to calculate sensitivity, specificity, or negative predictive value (NPV) due to having not included a representative sample of microscopy-negative infections. However, positive predictive values (PPV) were calculated.
From June 2014 to December 2015, 41 malaria cases were included in the study analysis. Forty-two were initially identified from the sentinel health facilities by microscopy and confirmed by cross-checking at the provincial laboratory, but one case (P. vivax by microscopy) was excluded as the DBS had insufficient blood for subsequent molecular testing. The 41 cases included: 20 P. falciparum (49%), 16 P. vivax (39%), 3 P. malariae (7%), and 2 with indeterminate morphology (5%) (Table 1). Parasite density ranged from 66 to 355,400 parasite/μL blood. The median and range of parasite density (in brackets) for microscopy-diagnosed P. falciparum, P. vivax and P. malariae were 5,447 (66 to 54,970), 32,157 (703 to 355,400) and 3,842 (1,760 to 7,133). The parasite densities of the indeterminate samples were 803 and 1,473, respectively. Microphotography of the indeterminate samples showed resemblance to other species (Fig 1).
Genus-specific Pan-LAMP testing at the provincial laboratory was positive in all 41 isolates (examples in Fig 2), and 8 tested positive by Pf-LAMP testing (Table 1). By cytb PCR genus-specific testing and using the AluI restriction digest reaction for species identification, 8 (19.5%) were classified as P. falciparum, 33 (80.5%) as P. vivax. By 18S rRNA nPCR, there were 8 P. falciparum (19.5%), 14 P. vivax (34.1%), and 19 (46.3%) did not amplify. P. knowlesi-specific nPCR was positive in 19/41 (46.3%), of which 11 underwent sequencing and showed 100% identity to a published P. knowlesi 18S rRNA gene sequence (P. knowlesi strain H1 chromosome 3, GenBank accession number AM910985).
Microscopy was unable to classify or mis-classified 23 of 41 (56%) malaria cases confirmed by the gold standard of serial molecular testing (Table 1). These included all 19 P. knowlesi cases, of which 17 were mis-classified as P. falciparum (n = 8), P. vivax (n = 6), or P. malariae (n = 3), and 2 were unable to be classified. There were also 4 P. vivax cases that were mis-classified as P. falciparum by microscopy. Sixty percent (12/20) of cases identified by microscopy as P. falciparum were either P. vivax or P. knowlesi; 37.5% (6/16) of cases identified by microscopy as P. vivax were P. knowlesi. All P. malariae and indeterminate species by microscopy were P. knowlesi.
Genus-specific testing by LAMP identified all infections, though species identification was limited by the unavailability of non-falciparum species-specific testing with the platform used. Pf-LAMP testing mis-classified one P. knowlesi mono-infection as P. falciparum but otherwise correctly identified all the P. falciparum cases.
Of cases classified as P. vivax by cytb PCR, 58% (19/33) were later confirmed as P. knowlesi and showed a similar banding pattern to P. vivax (Table 1 and Fig 3A). Using 18S rRNA species-specific nPCR for the four main human species, P. falciparum and P. vivax were correctly identified but all P. knowlesi infections were missed (Fig 3B). There was no cross-reactivity with P. vivax using P. knowlesi-specific nPCR (Fig 3C).
The positive predictive values (PPV) for species identification by different diagnostic methods using the gold standard of serial molecular testing are shown in Table 2. PPV was low for P. falciparum, P. vivax, and P. malariae identification by microscopy and for P. vivax identification by cytb nPCR. Where samples were available, PPV was high for all other methods.
To support malaria diagnosis and surveillance in Aceh Province, a low transmission setting in Indonesia that is aiming for malaria elimination, we utilized molecular testing for quality assurance of microscopy-confirmed cases from health facilities. As previously published, this work resulted in the first report of P. knowlesi in Indonesia outside Borneo, and an epidemiological investigation revealed that forest exposures are a key risk factor for this zoonotic infection [6]. In this study, we report the details and difficulties of species identification using microscopy at the point of care and a variety of molecular methods at reference laboratories. Microscopy mis-classified P. knowlesi cases as P. malariae or P. falciparum, as commonly reported elsewhere, but also as P. vivax, which has been less commonly reported [15]. The PPVs for the identification of other species (Pf, Pv, and Pm) were also poor. At the provincial reference laboratory, LAMP, a field-friendly molecular method, was useful in confirming all Plasmodium infections, though further species identification was limited by the unavailability of non-falciparum species-specific testing with the platform used. Use of less field-friendly nPCR methods at a national reference laboratory to identify P. knowlesi infection was not straightforward. All P. knowlesi cases did not amplify with a standard nPCR method (18S rRNA) targeting the four human-only species. With the cytb method, there was cross-reactivity with P. vivax for all P. knowlesi cases. We highlight the difficulties of P. knowlesi diagnosis at the point-of-care and reference laboratory levels in a setting where endemicity was not previously known and bring attention to an emerging challenge for malaria elimination.
The recent discovery and emergence of P. knowlesi, a fifth human species previously thought to only infect macaques, has created an additional challenge for species identification. Microscopy is difficult because the morphology at different stages resembles other malaria species [26]. The diagnostic sensitivity and specificity of available immunochromatographic rapid diagnostic tests (RDTs) for P. knowlesi detection is poor, leaving no other useful point-of-care diagnostic test [27–29]. Despite some global knowledge on the potential geographical distribution and extent of transmission of P. knowlesi [4], this information may lack resolution at local levels, and health workers and microscopists on the front-lines may have limited knowledge and/or a low index of suspicion for P. knowlesi. In our study, the investigation into P. knowlesi was prompted by the observation by an astute microscopist of unusual morphology in two malaria cases, as well as the known local presence of pig-tailed and long-tailed macaques and Anopheles leucosphyrus, a known vector on Sumatra island [30].
For quality control in reference laboratories, none of the nucleic acid-based methods for both genus and species-specific identification were found to be suitable. With LAMP, a molecular detection method that has been promoted for use in resource-limited settings due to the rapid turnaround time and simple methods, genus-specific testing was reliable, as has been reported from Malaysia [20]. However a P. knowlesi-specific commercial kit was not available for use in our study, and evaluations of other P. knowlesi-specific LAMP assays have reported cross-reactivity with P. vivax [18]. The P. knowlesi-specific PCR method utilized in this study did not cross-react with P. vivax infections, with excellent specificity as observed previously [16]. The nPCR methods used have problems with missed infections and/or species mis-classification. With commonly used 18S rRNA nPCR targeting the four human-only species, a commonly used reference standard, P. knowlesi either does not amplify (as occurred in this study) or is mis-classified as P. vivax due to high sequence homology at the target sequences [31, 32]. With the cytb nPCR method that we used, our finding of cross-reactivity between P. knowlesi and P. vivax has not been previously reported, but can also be explained by high sequence homology at the target mitochondrial sequences. Others have reported P. knowlesi amplification using P. vivax-specific PCR targeting the mitochondrial gene cox1 [33]. Other more sensitive and specific molecular methods for P. knowlesi detection in mixed species settings have recently been developed [7, 34, 35] and could be considered for future surveillance in our study setting.
The challenge of accurate P. knowlesi detection is of both clinical and public health significance. In Malaysia, where the clinical disease has been well studied, P. knowlesi is associated with at least as high a risk of severe disease compared with P. falciparum [36] and in early series, a high proportion had fatal outcomes [8, 37]. Following a number of interventions in Sabah state, case-fatality rates have fallen 6-fold [9]. These have included improved and now routine statewide molecular surveillance, more recent laboratory microscopy reporting of “P. malariae” as “P. knowlesi”, and enhanced implementation of standardized referral and clinical protocols, including first-line use of artemisinin-based combination therapy and early intravenous artesunate [9, 36]. Progression to severe disease is due not only to missed diagnoses, but also its ability to cause severe malaria at relatively low parasite densities [36]. Mis-classification of P. knowlesi as P. vivax, as occurred at the point of care in our study, also results in unnecessary treatment with primaquine, an antimalarial not indicated for P. knowlesi, but necessary for radical cure of the latent liver stages with P. vivax. In our study, we did not experience any severe adverse events from the unnecessary use of primaquine, but use in subjects with underlying severe glucose-6-phosphate dehydrogenase deficiency is known to be associated with life-threatening hemolysis.
While only recently recognized in areas of Aceh and North Sumatra, there has been little molecular surveillance of P. knowlesi distribution and incidence elsewhere in Indonesia, particularly across Kalimantan, Sulawesi and other regions of Sumatra, where modelling predicts a high risk of human infection [38]. From a public health perspective, accurate identification of P. knowlesi is critical to the design and implementation of effective malaria interventions. In a related study in Aceh Province and also in Malaysia, adult males with forest-related and agricultural occupational exposure are at significantly higher risk of being infected with P. knowlesi [6, 13]. Interventions would therefore need to be targeted to this population. As well as continued promotion of conventional malaria prevention activities to reduce peridomestic transmission [13], other interventions would need to be targeted to P. knowlesi-transmitting mosquitos, the interface between humans and macaques, and to individual risk factors for infection identified in different settings. Further investigation into the epidemiology and transmission of P. knowlesi in Aceh Besar is needed.
Limitations of microscopy to identify P. knowlesi are well established. Our challenges using LAMP and PCR for species identification in a setting with previously unknown P. knowlesi endemicity add to a growing literature on the limitations of molecular methods as well. For settings approaching malaria elimination and/or where epidemiological conditions are predicted to support P. knowlesi transmission to humans, quality assurance of malaria diagnosis and species identification is essential, but at present, practical and accurate methods are not available for local and peripheral reference laboratories. Development, evaluation and implementation of improved P. knowlesi detection methods for use at both the point-of-care and in local reference laboratories are needed.
|
10.1371/journal.pcbi.1000907 | Evolution under Fluctuating Environments Explains Observed Robustness in Metabolic Networks | A high level of robustness against gene deletion is observed in many organisms. However, it is still not clear which biochemical features underline this robustness and how these are acquired during evolution. One hypothesis, specific to metabolic networks, is that robustness emerges as a byproduct of selection for biomass production in different environments. To test this hypothesis we performed evolutionary simulations of metabolic networks under stable and fluctuating environments. We find that networks evolved under the latter scenario can better tolerate single gene deletion in specific environments. Such robustness is underlined by an increased number of independent fluxes and multifunctional enzymes in the evolved networks. Observed robustness in networks evolved under fluctuating environments was “apparent,” in the sense that it decreased significantly as we tested effects of gene deletions under all environments experienced during evolution. Furthermore, when we continued evolution of these networks under a stable environment, we found that any robustness they had acquired was completely lost. These findings provide evidence that evolution under fluctuating environments can account for the observed robustness in metabolic networks. Further, they suggest that organisms living under stable environments should display lower robustness in their metabolic networks, and that robustness should decrease upon switching to more stable environments.
| One of the most surprising recent biological findings is the high level of tolerance organisms show towards loss of single genes. This observation suggests that there are certain features of biological systems that give them a high tolerance (i.e. robustness) towards gene loss. We still lack an exact understanding of what these features might be and how they could have been acquired during evolution. Here, we offer a possible answer for these questions in the context of metabolic networks. Using mathematical models capturing the structure and dynamics of metabolic networks, we simulate their evolution under stable and fluctuating environments (i.e., available metabolites). We find that the latter scenario leads to evolution of metabolic networks that display high robustness against gene loss. This robustness of in silico evolved networks is underlined by an increased number of multifunctional enzymes and independent paths leading from initial metabolites to biomass. These findings provide evidence that fluctuating environments can be a major evolutionary force leading to the emergence of robustness as a side effect. A direct prediction resulting from this study is that organisms living in stable and fluctuating environments should display differing levels of robustness against gene loss.
| High-throughput single gene deletion studies in several organisms revealed that a large fraction of genes have little or no detectable fitness effects when compromised [1]–[5]. These observations raise the question of how biological systems can acquire and maintain such robustness against gene loss. As for any biological trait, robustness could be adaptive, resulting from direct selection for it, or non-adaptive, resulting as a byproduct of other selective pressures [6]. Understanding which of these modes apply is important both to distill the design principles of biological systems and to understand how amenable robustness is to manipulation [7].
Direct selection for robustness against gene loss is expected to be weak [8], becoming relevant only under high mutation rates [9], [10]. In line with these theoretical findings, empirical analyses find only limited contribution of gene duplications to the observed robustness [11]–[15]. On the other hand, different forms of robustness are shown to evolve in non-adaptive fashion under certain conditions. For example, in near-neutral fitness landscapes mutational robustness can emerge easily [16]. In metabolic networks, it is argued that properties of enzyme kinetics can render the systems robust against partial loss-of-function mutations [17], [18]. Moreover, robustness against small mutations is shown to evolve in gene regulatory networks selected for dynamic stability [19], [20] and robustness against gene deletions is shown to evolve in signaling networks under parasite interference [21].
It is possible that biomass production and adaptation to multiple environments act as similarly realistic selective pressures on metabolic networks that could lead to the emergence of robustness as a byproduct. The former can drive the emergence of isoenzymes for increased dosage [22], resulting in a clear case of functional redundancy mediated robustness. The latter could lead to multiple pathways, each specializing in processing metabolites present in one of the multiple environments. These multiple pathways could compensate for each other, particularly, in rich media [7]. This scenario is in line with the observation that the estimated fraction of dispensable genes at both metabolic [23]–[25] and genome scale [26] reduces dramatically when multiple environments are considered. The most recent computational analysis of metabolic networks from Escherichia coli and Saccharomyces cerevisiae finds that, when the effect of deletion is tested in all possible environments, only half of all reactions determined to be dispensable under rich media could be considered dispensable for “real” [25]. Further, almost all of the remaining cases can be explained by recent duplications, horizontal gene transfer events or pleitropic effects (i.e. compensation by multifunctional enzymes) [25]. It is important to note that these studies typically judge dispensability based on stoichiometric approaches such as flux balance analysis (FBA). By focusing only on lethal knockouts, and ignoring the fitness effect of non-lethal ones, these approaches therefore overestimate robustness.
Taken together, the above described studies suggest that observed robustness against gene deletion in metabolic networks is a byproduct of their evolutionary dynamics under changing environmental conditions. Early studies on the effects of changing environments in evolution have shown that it can facilitate polygenic variation [27], [28] and can lead to modularity at network level [29]–[32]. In addition, abrupt changes in selective pressure are shown to lead to significant changes in metabolic networks [33]. Here, we specifically study the effects of fluctuating selection on the emergence of robustness in metabolic networks. Using a well-accepted scenario of duplication and specialization [34]–[37], we simulated evolution of metabolic networks under selection for converting environmentally available metabolites into biomass. These simulations started from initial networks composed of unspecific enzymes, which duplicate and specialize as evolution progresses, resulting in metabolic networks with high biomass production rate. To test the effect of the environment on the properties of evolved networks we performed simulations under stable and fluctuating environments (Figure 1). Networks that evolved in a stable environment were selected for biomass production in either one of two different minimal media or in a rich medium; network fitness was a function of biomass production rate given the metabolites in the media. Networks that evolved in a fluctuating environment faced changes between these three media and were selected for biomass production in all of them; network fitness was defined as the geometric mean of fitness values in each of the individual media. The resulting networks were tested for their robustness against gene loss. For networks that evolved in a fluctuating environment, robustness was determined separately in each medium and over all media allowing us to investigate whether any resulting robustness in these networks is apparent or real. A detailed schematic of the simulations and analysis is given in Figure 1.
To study the effect of selection under fluctuating environments on metabolic network properties, we relied on a proposed evolutionary scenario [35]. According to this scenario, metabolic networks characterized by large numbers of enzymes with high specificity have evolved from ancestral networks consisting of few enzymes with broad specificity [34], [37]. Such evolution could be driven by selection for increased growth rate (i.e. biomass production rate), and mutations affecting kinetic properties of enzymes and resulting in gene duplications. Although a number of alternative scenarios for the evolution of novel enzymes and metabolic pathways have been proposed [38], this scenario is plausible for the early evolution of metabolic networks.
Here, we implement this scenario using a computational model of metabolic networks. In brief, the model consists of metabolites, enzymes that catalyze the transfer of biochemical groups between metabolites, and transporters that can allow intake and release of metabolites (see Methods). We start evolutionary simulations with enzymes that can catalyze all group transfer reactions. In the course of evolution enzymes can subsequently specialize through duplications and mutations. This process is driven by the assumption that there is a trade-off between catalytic activity and specificity. This assumption is well supported by the existence of specialized enzymes in nature and by several directed evolution experiments that exploit such trade-off for protein engineering [37]–[39]. The model structure allows us to capture both subfunctionalization [40], [41] and neofunctionalization [42]; two processes that are believed to be at the core of evolution of gene duplicates [43]–[46]. Running evolutionary simulations that mimic natural evolution as a deterministic process we evolve networks towards a local optimum and analyze the aspects in which these optima differ for different fitness landscape. The deterministic approach to simulating evolution corresponds to a scenario with a large population and low mutation rate (also referred to as strong selection - weak mutation scenario [47]). In summary, the presented model captures the dynamics and stoichiometry of metabolic networks and the evolution of these properties. Previously, we have shown that it can result in the evolution of complex metabolic networks that have very similar global properties to their natural counterparts [36].
Using this model we have run evolutionary simulations under different environmental scenarios (see Methods). In particular, we have evolved metabolic networks under three stable environments and a fluctuating one (Figure 1). In all these environments fitness was related to the ability of the network to convert available metabolites into biomass (see Methods). The three stable environments respectively contained either one of two randomly chosen pairs of metabolites (minimal media; M1 and M2) or both of them (rich media; R = M1 + M2). The fluctuating environment was assumed to vary between these three media. In all these simulations network fitness increased quickly as evolution progressed, and enzymes became more specialized (Figures 2 and 3). To understand how evolution under these different scenarios affected network robustness, we have analyzed the effect of single gene knockouts on fitness. As shown in Figure 4, we found that in networks evolved under fluctuating media, single gene deletions had significantly lower fitness effects compared to networks that evolved in stable media. Interestingly, the difference in robustness against gene deletion was most prominent when fitness was measured under rich media and was completely lost when it was measured over all media seen during evolution (see also Figure 2). Hence, fluctuating evolution resulted in the emergence of an “apparent” robustness against gene deletion that became most detectable in rich media.
To test that these results are robust against the main assumptions of the model and the simulation scheme, we have analyzed an alternative model. In this model, enzymes were allowed to maintain broader activity by introducing a small background rate for all reactions an enzyme can catalyze (see Methods). This approach is inline with the idea of “underground metabolites” [38] and allows us to start or continue an evolutionary simulation from any starting network. We use this ability to change the simulation scheme so that we start simulations under fluctuating environments from networks that already have evolved under stable environments. This alternative approach is potentially more inline with conditions in nature where networks can experience sudden changes in environmental conditions [33]. We find qualitatively the same results as with the previous analysis; robustness of networks evolved under a stable environment increase when they further evolve under fluctuating environments (Figure S2). As before, this higher robustness is apparent, however it is not completely lost when considering all environments a network experiences during evolution (Figure S2).
To understand the basis of such robustness we have analyzed the structure of networks resulting from evolution under stable and fluctuating media. As mentioned above, all evolutionary simulations resulted in enzymes that are specialized and in networks with faster biomass production compared to the ancestral ones. However, networks evolved under fluctuating media displayed two important features that distinguished them from networks evolved under stable media. Firstly, fluctuating environments resulted in networks that contain more redundant paths. The average number of independent fluxes (see Methods) that can be channeled through the network ranged from 1.2 to 1.4 for networks that evolved in stable environments, while it was 4.1 for networks evolving in fluctuating environment (Table 1). The extent of redundant paths in the latter networks is clearly seen in sample networks shown in Figure 3. Secondly and related, networks evolved under fluctuating environment contained significantly more multifunctional enzymes, i.e. enzymes that catalyzed more than one group transfer reaction (see Methods). Of the 100 independent simulations for each scenario, 72% of networks that evolved under fluctuating environments contained at least one multifunctional enzyme compared to 36%, 40% and 28% of networks evolved under stable environments M1, M2, and R respectively. Further, in networks evolved under fluctuating environment 24% of all enzymes were multifunctional, while only 6–9% were multifunctional in networks evolved under stable environment (Table 1).
These clear differences in the global properties of networks evolved under fluctuating and stable media suggest that both the number of independent fluxes and the number of multifunctional enzymes in a network contribute to its robustness. To better understand the relation between these properties and robustness, we performed a detailed analysis of the fitness effects of single gene deletions (Table 1). In networks evolved under stable environments, the deletion of monofunctional enzymes had, on average, 7 to 11 fold larger effect compared with the deletion of multifunctional enzymes. Consequently, networks that contained multifunctional enzymes were on average 3 to 6 fold more robust compared to networks without any such enzymes.
Deletion of multifunctional enzymes might result in lower fitness effects either because these enzymes behave as isoenzymes (i.e. back up the function of another enzyme) or catalyze reactions that are non-essential but beneficial. We find evidence for both of these possibilities. Firstly, almost all multifunctional enzymes have low dosage, suggesting that the reactions they catalyze are non-essential but beneficial when they occur at low rate. Because these reactions are beneficial when occurring at low rate, there is no selective pressure for the multifunctional enzymes to duplicate (so to increase dosage) and potentially specialize. Secondly, most multifunctional enzymes have functions that overlap with other enzymes in the network. This finding results from a structural analysis of all networks evolved under one of the stable environments (M1-networks) and under fluctuating environments (V-networks) and that contain a multifunctional enzyme: We find that among the 36 M1-networks with multifunctional enzymes, 19 contain at least one multifunctional enzyme that behaves as an isoenzyme. Among the 72 V-networks with multifunctional enzymes, 42 contain at least one multifunctional enzyme that behaves as an isoenzyme. In both M1- and V-networks, multifunctional enzymes in the remaining networks catalyze at least one reaction that is not directly involved in biomass production, further supporting their non-essential role.
Interestingly, the difference in fitness effects of deleting multi- vs. mono-functional enzymes were significantly reduced in networks evolved under fluctuating environments. Considering fitness effects in different media, deleting monofunctional enzymes in these networks had, on average, only 1–2 fold larger effect compared with deleting multifunctional enzymes. Similarly, networks containing multifunctional enzymes were only 1 to 3 fold more robust than those without any such enzymes. These analyses suggest that while multifunctional enzymes can contribute significantly to robustness against gene deletion, evolution under fluctuating media increased robustness of networks mostly through generation of independent fluxes.
If evolution under fluctuating environments is the driving force behind emergence of robustness against gene deletion, would it be possible that robustness is lost as the environment stabilizes? Indeed, redundant paths (i.e. independent fluxes) might infer a cost on the organism due to increased number of enzymes that are not always required, or because some of the paths are in fact disadvantageous in some environments. This is the case in our simulations as we find certain gene deletions to increase fitness above wild type levels in networks evolved in fluctuating media (Figure 4). To further analyze the possibility of loosing robustness in stable environments we take networks evolved under fluctuating media and continue their evolution under any of the three stable media. As shown on Figure 4, we find that such subsequent evolution results in complete loss of any gained robustness against gene deletion; the distribution of fitness effects of gene deletions for these networks is the same as for those which have evolved in stable media originally. This reduction in robustness is accompanied by a reduction in both the number of multifunctional enzymes and independent fluxes (Table 1).
The finding that a switch in the environment towards stability leads to reduction in robustness fits nicely with the observation that prokaryotes specializing on one mode of energy generation has much reduced fraction of dispensable genes compared to generalists [48], however, it should be taken with care as we model switch to a stable environment to be perfect while in reality it is possible that environments are never entirely stable. It can be shown that even very rare fluctuations could maintain functional redundancy mediated robustness; for example, a gene providing a fitness advantage of s in a given environment could be maintained even if that environment is seen only once every s/u generations, where u is the mutation rate [25].
Here we have provided evidence that fluctuating environments can lead to emergence of robustness against gene loss in metabolic networks. Using computer simulations that embed a plausible scenario of metabolic network evolution, we found that selection for biomass production rate in a fluctuating media leads to emergence of networks, which can tolerate single gene deletions more readily. This robustness against gene loss is highest when fitness is measured under rich media, where all metabolites seen during evolution are considered to be available, and diminishes as fitness is measured separately under each media. We find that the molecular basis of such robustness in evolved networks is an increase in the number of independent fluxes and multifunctional enzymes.
These findings are perfectly in line with observations made in natural, current-day metabolic networks. Computational analysis of metabolic networks from E. coli and S. cerevisiae finds that most of the observed robustness in rich media is apparent, strongly diminishing as different environments are considered separately [23]–[25]. While these works have suggested that such robustness could be due to compensating pathways and to enzymes that have differential efficiencies under different environmental conditions [23], [25], the presented study provides a clear evolutionary route to these features. Further, it indicates that considering dynamic response of metabolic networks might reveal more severe fitness effects of gene deletions when considering multiple environments.
Interestingly, we find that robustness and its underlying features would be lost entirely as the environment stabilizes and network evolution continues. This leads to the prediction that robustness of the metabolic network in an organism should be directly correlated with environmental conditions it experiences; organisms whose metabolism depends on stable resources should display lower robustness. This prediction is supported both from a specific gene deletion study in Mycoplasma genitalium [3], which has a minimal metabolism, and from a larger comparison of gene dispensability in specialist and generalist prokaryotes [48].
Any attempt to fully distill design principles of biological systems has to consider evolutionary dynamics [49]. This can be achieved with in silico evolution as presented here or alternatively by considering the space of possible metabolic networks and how evolution could move in this space. The two approaches are complementary; recent modeling studies using the latter approach are providing us with important insight on common design principles that can result in evolution [50], [51], while approaches like the one presented here show how different selective pressures can shape the global properties of metabolic networks. As with any modeling study, the presented analysis has limitations and potential caveats. In particular, our analysis was limited in network size due to computational costs associated with evolutionary simulations and the generic model and the measure of robustness had to be based on several simplifications and assumptions about metabolism. While we find that our main findings are robust against such limitations and the main modeling choices, a full confirmation of our results can only be achieved with experimental studies. In this regard, we note that long-term evolution experiments under stable lab conditions provide a direct test bed to confirm the ideas presented here. These studies have already shown that evolution under stable environments reduce the metabolic breadth of E. coli [52]. We would expect that it has also reduced its “apparent” robustness against gene loss.
Methods have been described in detail previously [36]. In brief, we implement a well-accepted scenario of metabolic network evolution [35], where an ancestral network composed of few unspecific enzymes evolves through mutations altering kinetic rates and duplications. At the core of this scenario is the argument that new enzyme activities result from specialization of enzymes with broad activity [34]. There is now empirical evidence that such specialization have led to the evolution of most, if not all, enzyme superfamilies [37]. In addition, laboratory evolution has been successfully employed to select or de-select for promiscuous functions, thereby altering enzyme function(e.g. [38], [39]).
The details of different modeling choices we made are as follows.
Metabolites are assumed to consist of five different biochemical groups. Each biochemical group is either present once or is absent, resulting in a total of 32 possible metabolites. Each metabolite can be represented by a binary string of length 5, where “1” at position g denotes the presence of group g, whereas “0” denotes the absence of that group. Metabolites are associated with a random free energy that is taken from a uniform distribution between zero and one, and that is required to specify thermodynamic properties of the biochemical reactions. For the production of biomass, it is necessary to have at least one donor and acceptor of each group as external metabolites. Thus, a minimal medium contains two randomly chosen metabolites as a donor-acceptor pair. Rich medium consists of two different random donor-acceptor pairs. Four randomly chosen metabolites are involved in biomass formation. All of these random choices are made independently for each evolutionary simulation.
Enzymes catalyze the transfer of a specific biochemical group. We assume that groups are transferred by a “ping-pong mechanism”: A donor of a group transfers the group to the appropriate enzyme and is thereby transformed into its corresponding acceptor. The enzyme then transfers the group to an acceptor, thereby transforming it into its corresponding donor. Thus an enzyme can be in two possible states, Ei(1) and Ei(0), with Ei(1) + Ei(0) = Ei. Here, Ei is the total dosage of enzyme i, Ei(0) is the concentration of the enzyme without its group being bound to it, and Ei(1) is the concentration of enzyme with its group being bound to it. The free energy difference between both states of an enzyme is assumed to be a random value taken from a uniform distribution ranging from zero to one. We further assume that in principle all metabolites that contain a specific biochemical group (i.e., half of the 32 metabolites) can serve as a donor for the transfer reaction involving that group, whereas all metabolites that do not contain that group can in principle serve as an acceptor. We assume linear kinetics for the transfer of a group to an enzyme, given by v = kij(Ei(0)X(ij)(1)−Ei(1)X(ij)(0)/qij), where kij is the rate constant of the reaction j of an enzyme i, X(ij)(1) is the concentration of the donor of the reaction, X(ij)(0) is the concentration of the corresponding acceptor, and qij is the equilibrium constant of the reaction resulting from the free energies of the reactants. For the transfer of a group from a donor to an acceptor, two half-reactions need to be coupled. This results in Michaelis-Menten–like kinetics and implies that functional enzymes need to maintain nonzero rate constants for at least two reactions that are coupled. Enzymes that maintain nonzero rate constants for more than two reactions are defined as multifunctional. We assume that in the initial network there are 5 enzymes that are unspecific and transform groups from each donor to each acceptor with the same rate constant. The initial dosage of enzymes is Ei = 1.
We assume that transporters transport metabolites passively across the cell membrane. The rate of transport is given by v = Titij(Xj−Xjext), where Ti is the dosage of the transporter i, tij is the rate constant for the transport of metabolite j, Xj is the metabolite concentration in the cell, and Xjext is the metabolite concentration in the environment. We assume that in the initial network there is a single transporter that transports all metabolites with the same rate constant. The initial dosage of the transporter is T0 = 1.
Biomass is formed by the condensation of specific metabolites. The rate of biomass formation follows linear kinetics given by the product vBM = kBMΠiXi over all metabolites Xi that are involved in biomass formation. The rate constant is set to kBM = 1 in all simulations. We assume that the formation of biomass leads to growth. The growth rate is given by W = 1/V*dV/dt = vBM/(C0+CEE+CTT), where C0 is the amount of biomass that is required for structural compounds (i.e., those compounds that are not directly involved in cellular metabolism), CE is the amount of biomass per enzyme, CT is the amount of biomass per transporter, E is the total dosage of enzymes, and T is the total dosage of transporters. The parameters C0, CE, and CT are set to 10, 1, and 1, respectively. Note that due to cell growth, metabolites are constantly diluted at a rate equal to the growth rate. The fitness of a network in a given medium is assumed to be proportional to the steady-state growth rate. The fitness in an environment that fluctuates between different media is given by the geometric mean over the fitness values a network has on each of the media.
We assume that enzymes can either catalyze a large number of reactions with low activity, or a lower number or reactions with improved catalytic activities. Specifically, we assume that the sum Σj kij1/α and Σj tij1/α over all rate constants kij or tij of an enzyme or transporter, i, respectively, is constant. For values of α>1, increasing the rate constant for a single reaction has an over-proportional effect on all other rate constants. In our simulations we use Σj kij1/α = 1, Σj tij1/α = 1, and α = 2. This implies, for example, that a transporter catalyzing the transport of a single metabolite has a four times higher rate constant for this reaction than a transporter that is specialized on the transport of two metabolites.
The resulting trade-off between enzyme specificity and activity in the model is inline with the general findings from protein engineering and directed evolution experiments [37], [53]. In particular, the tradeoff in our model allows specialized enzymes to retain some (minor) catalytic activity for other reactions. This resembles a situation described as weak negative tradeoff [37]. However, because two specialized enzymes will be better than a single multifunctional enzyme present at double dosage, there is also selection for specialization. The presence of such selection in the model seems justified by the fact that most enzymes in natural metabolic networks are specialized.
In an alternative model, we further relax the assumption of a strong tradeoff between specificity and catalytic activity and allow enzymes to specifically maintain a background activity for all possible reactions. This alternative model allows us to use any network for the starting point of evolutionary simulations. Using this model, we have analyzed whether forcing specialization of enzymes towards specific reactions (which in some extent decreases complexity in the system) has any effect on our conclusions. As shown in Figure S2, we find that this alternative model to produce qualitatively the same results as with the main model.
We assume that there are two types of mutations: (1) mutations that change the kinetic properties of an enzyme and (2) mutations that change the number of copies of an enzyme, i.e., gene deletions and duplications. For the first type of mutation we assume that the value of kij1/α or tij1/α, respectively, for a single reaction is either increased or decreased by a small value of m = 0.05, while the rate constants of the other reactions are decreased, or increased appropriately. Gene deletions and duplications decrease and increase the dosage of an enzyme respectively. To simulate evolution, we first calculate the effect of all possible mutations in the current network on the steady-state growth rate to obtain the mutant with maximal increase in fitness. This mutation is then assumed to become fixed and the resulting network is used to search for the next mutations. Details on the calculation of the steady states are as described in [36]. Gene duplications and deletions are assumed to be rare compared to mutations affecting the catalytic properties of enzymes and transporters and are considered only if none of the mutations affecting kinetic properties are beneficial. We find that relaxing this assumption and considering duplications as frequently as other mutations does not alter the conclusions given in the main text (Figure S1). An evolutionary simulation ends if there are no beneficial kinetic mutations, gene duplications or deletions.
For each set of independent simulations, we randomly chose nutrients, metabolites involved in biomass formation, and the free energies of the metabolites. Changing free energies of the metabolites alters the energetic landscape of the initial network and might favor different pathways even if the topology of the network remains the same.
Note that, while changing the many parameters of the model could easily alter the properties of individual evolved networks, the qualitative nature of the results presented here would be maintained. This is because our analysis is a comparative one among networks that evolve under different evolutionary scenarios. While model details would change the outcome of evolution in specific simulation, they would do so in similar ways under the different scenarios considered.
Specific parameter choices in this study differ from a previous study using the same model [36], as here we run a large number of simulations of smaller networks. To be able to manage the computational cost of these simulations, we adjusted the parameters for the costs of biomass formation (from 50 to 10), the number of metabolites involved in biomass formation (from 8 to 4), and the mutation size (from 0.01 to 0.05).
The evolved networks contain many enzymes in multiple copies. These multiple copies of a single enzyme could be seen as isoenzymes. As one would expect for isoenzymes, the knockout of a single copy would have a relatively small effect. We here determine robustness by knocking out all enzymes of the same type. This gives a measure for how essential the reaction catalyzed by an enzyme is.
More specifically, we calculate robustness as network's rate of biomass production (vBM, see above) in the face of enzyme knockouts. To measure it, we delete each enzyme existing in a given network one by one and calculate the rate of biomass formation for each mutant. This allows us to characterize the full dynamical effect of a gene deletion on biomass formation, rather than just viability (i.e. non-zero vs. zero biomass production rate) and effects on yields.
To understand the global structure of evolved networks, we measure metabolic flux from metabolites to biomass. In particular, we calculate, for each network, the number of independent fluxes by using the kernel of the stoichiometric matrix derived from that network. The details of this technique is discussed in detail elsewhere [54].
|
10.1371/journal.pcbi.1006892 | How Dendrites Affect Online Recognition Memory | In order to record the stream of autobiographical information that defines our unique personal history, our brains must form durable memories from single brief exposures to the patterned stimuli that impinge on them continuously throughout life. However, little is known about the computational strategies or neural mechanisms that underlie the brain's ability to perform this type of "online" learning. Based on increasing evidence that dendrites act as both signaling and learning units in the brain, we developed an analytical model that relates online recognition memory capacity to roughly a dozen dendritic, network, pattern, and task-related parameters. We used the model to determine what dendrite size maximizes storage capacity under varying assumptions about pattern density and noise level. We show that over a several-fold range of both of these parameters, and over multiple orders-of-magnitude of memory size, capacity is maximized when dendrites contain a few hundred synapses—roughly the natural number found in memory-related areas of the brain. Thus, in comparison to entire neurons, dendrites increase storage capacity by providing a larger number of better-sized learning units. Our model provides the first normative theory that explains how dendrites increase the brain’s capacity for online learning; predicts which combinations of parameter settings we should expect to find in the brain under normal operating conditions; leads to novel interpretations of an array of existing experimental results; and provides a tool for understanding which changes associated with neurological disorders, aging, or stress are most likely to produce memory deficits—knowledge that could eventually help in the design of improved clinical treatments for memory loss.
| Humans can effortlessly recognize a pattern as familiar even after a single presentation and a long delay, and our capacity to do so even with complex stimuli such as images has been called "almost limitless". How is the information needed to support familiarity judgements stored so rapidly and held so reliably for such a long time? Most theoretical work aimed at understanding the brain's one-shot learning mechanisms has been based on drastically simplified neuron models which omit any representation of the most visually prominent features of neurons—their extensive dendritic arbors. Given recent evidence that individual dendritic branches generate local spikes, and function as separately thresholded learning/responding units inside neurons, we set out to capture mathematically how the numerous parameters needed to describe a dendrite-based neural learning system interact to determine the memory's storage capacity. Using the model, we show that having dendrite-sized learning units provides a large capacity boost compared to a memory based on simplified (dendriteless) neurons, attesting to the importance of dendrites for optimal memory function. Our mathematical model may also prove useful in future efforts to understand how disruptions to dendritic structure and function lead to reduced memory capacity in aging and disease.
| To function well in a complex world, our brains must somehow stream our everyday experiences into memory as they occur in real time. An “online” memory of this kind, once termed a “Palimpsest” [1], must be capable of forming durable memory traces from a single brief exposure to each incoming pattern, while preserving previously stored memories as long and faithfully as possible (Fig 1). This combined need for rapid imprinting and large capacity requires that the memory system carefully manage both its learning and forgetting processes, but we currently know little about how these processes are implemented and coordinated in the brain.
A number of quantitative models have been proposed for palimpsest-style online memories, and have addressed a variety of different issues, including: how memory capacity scales with network size, how metaplastic learning rules can increase memory capacity, and the tradeoff between initial trace strength and memory lifetimes [1–8]. A few studies with a more empirical focus have addressed the biological mechanisms underlying recency vs. familiarity memory [9]; the coordination of online learning with long-term memory processes; and the details of memory-related neuronal response properties during online learning tasks [10–12].
Nearly all previous models of online learning have assumed that the neurons involved in memory storage are classical "point neurons”, that is, simple integrative units lacking any representation of a cell’s dendritic tree. This simplification is notable, given the now substantial evidence from both modeling and experimental studies that dendritic trees are powerful, functionally compartmentalized information processors that can augment the computing capabilities of individual neurons in numerous ways [7,13–59].
Beyond their contributions to the computing functions of neurons, it is also increasingly apparent that dendrites help to organize and spatially compartmentalize synaptic plasticity processes [7,40,60–86].
Thus, given that dendrites can act as both signaling and learning units within a neuron, it is important to understand how having dendrites could affect the brain’s online learning and memory processes. In this paper, we focus on the role that dendrites may play in familiarity-based recognition, a function most closely associated with the perirhinal cortex [87,88].
Here, we introduce a mathematical model that allows us to calculate online storage capacity from the underlying parameter values of a previously proposed dendrite-based memory circuit [7]. The model includes biophysical parameters (dendritic learning and firing thresholds, network recognition threshold), wiring-related parameters (number of axons, number of dendrites, number of synapses per dendrite), and input pattern statistics (pattern density, noise level) (see Table 1). As an example of the model’s use, we study the interactions between memory capacity, dendrite size, and pattern statistics, and cross-check the results using full network simulations. We found that dendrites containing a few hundred synapses (as opposed to a few tens or a few thousand) maximize storage capacity, providing the first normative theory that accounts for the actual sizes of dendrites found in online memory areas of the brain.
We modeled the memory network depicted in Fig 2a, consisting of a set of axons that form sparse random connections with the dendrites of a population of target neurons. An “infinite” sequence of random binary patterns is presented by the axons to the dendrites, each one causing one-shot changes to certain synapses within the network, where the goal of the network is to respond weakly to any pattern on its first presentation, and strongly for as long as possible to patterns that have been previously experienced. We define capacity as the number of consecutive training patterns stretching from “now” back into the past that can be classified as familiar with a low false negative (i.e. “miss”) rate, while maintaining a low false-positive (i.e. “false alarm”) rate to randomly drawn distractors (Fig 1).
The network structure and plasticity rules have been previously described in [7], but are repeated here for clarity. A population of neurons with a total of M separately thresholded dendrites receives inputs from NA input axons (Fig 2b). Each dendrite receives K synaptic contacts randomly sampled from the NA axons, for a total number of synapses NS=M∙K. The connectivity matrix is assumed to be fixed.
Input patterns are binary-valued vectors x = {x1,…,xNA} for which component xi is 1 if the ith axon is “firing” and 0 otherwise. We quantify density/sparsity of the patterns by the fraction of axons fA firing in each pattern; the value of fA ranged from 0.008 to 0.18 in this study, as we found empirically in previous work that sparse patterns maximize capacity in this type of memory [7]. To model a biologically realistic form of input variability, we assumed that each active axon (xi=1) produces a burst of spikes, where the number of spikes in the burst is drawn from a binomial distribution with mean μburst=Nburst·Pburst=4 spikes/burst. Pburst ranged from 1 (no noise) to 0.4 (high noise), with Nburst varying inversely. Inactive axons (xi=0) were assumed to produce no spikes. We denote the noisy spike count version of an input component x~i~xi∙Binom(Nburst,Pburst).
Synapses are characterized by both a weight wij, where the subscript indicates a connection between axon i and dendrite j, and an additional scalar parameter αij, representing the synapse’s “age”. The weight of each synapse is binary-valued, and can change between weak (w = 0) and strong (w = 1) states when the dendrite containing the synapse undergoes a learning event; the conditions that trigger a learning event are discussed below. The age variable at each synapse tracks the number of learning events that have occurred in the parent dendrite since the synapse last participated in learning.
Two different measures of a dendrite’s activation level determine how the dendrite responds to an input, and whether it undergoes a learning event. The “presynaptic” activation measure is based on the activity levels of the set of axons Dj that make contact with the jth dendrite
apre(j)=∑iϵDjx˜i.
In words, apre(j) is the total number of presynaptic spikes arriving at all the synapses impinging on the jth dendrite, regardless of their postsynaptic weights, and is thus a measure of the maximum response the dendrite could muster to that input pattern assuming all of the activated synapses were strong (w=1).
The more conventional “postsynaptic” activation level takes account of the synaptic weights in the usual way:
apost(j)=∑iϵDjwij·x˜i.
When the postsynaptic activation level exceeds the “firing” threshold θF, the dendrite is said to fire, that is, generates a response rj = 1. The responses of all dendrites within a neuron sum linearly to produce the neuron’s response (Fig 2b), and the responses of all neurons in the network sum linearly to produce the overall network response r. The overall response of the network can therefore be written directly as a sum over all the M dendritic responses:
r=∑jϵ[1,M]rj
so that the network can be viewed as a single “super neuron” with M dendrites.
Finally, an input pattern is classified as “familiar” if r≥θR, and “novel” if r<θR, where θR is the recognition threshold (Fig 2b).
The goal of learning is to ensure that learned patterns going back as far as possible in time produce suprathreshold network responses (r≥θR), while randomly drawn patterns do not. Learning of any given pattern occurs in only the small fraction of dendrites that cross both the presynaptic and postsynaptic learning thresholds (apre(j)>θLpre and apost(j)>θLpost). When this occurs, a “learning event” is triggered in the dendrite, and all active synapses belonging to that dendrite “learn”, as follows. If an active synapse is currently in the weak state, it is “potentiated” (i.e. both strengthened and “juvenated”: wij→1,αij→0), or if it is already in the strong state, then it remains strong but is juvenated (wij=1, αij→0). All strong synapses in the dendrite that are not active during the learning event remain strong but grow older (wij=1,αij→αij+1). Thus αijcounts the number of learning events that have occurred in the dendrite since the synapse last learned, and thus represents the age of the most recent information that that synapse is involved in storing. Note that a synapse’s age variable counts learning events within its parent dendrite only, and any given dendrite learns only rarely, so the counter need have only a small number of distinct values, on the order of ~12 under the simulation conditions explored in this paper. To maintain a constant fraction of strong synapses (we used fs=0.5), and thereby to prevent saturation of the memory, in each dendrite undergoing learning, a number of strong synapses are depressed (wij→0) equal to the number of weak synapses potentiated during that learning event. A key feature of the learning rule is that the synapses targeted for depression are those that learned least recently (i.e. having the largest values of αij), so that the information erased during depression is the “oldest” stored information. This “age-ordered depression” strategy substantially increases online storage capacity [5], especially in a 2-layer dendrite-based memory where the very sparse use of synapses during pattern storage gives each strong synapse, and the information it represents, the opportunity to grow old [7].
One of the key quantities involved in calculating storage capacity is L, the length of the age queue within a dendrite (see Fig 3). An approximate expression for L is given here; the derivation can be found in the Methods.
L is a measure of the time a pattern feature persists in a dendrite, and given that age queues progress at roughly equal rates in all the dendrites involved in storing a pattern, it also effectively measures a pattern’s lifetime in memory–counted in units of dendritic learning events. L can be understood intuitively through an oversimplified example: If 10 synapses are strengthened on a dendrite during a learning event, and there are 120 strong synapses on the dendrite, then L would be ~12. That is, after ~12 learning events have elapsed since a pattern was first stored, the 10 synapses involved in storing the pattern are now the oldest on the dendrite and must be depressed, and the memory is lost. The actual expression for L is more complex as it takes into account the fact that strong synapses do not inexorably progress to the ends of their age queues–they can be rejuvenated one or more times during the course of their lifetimes, in which case the same strong synapse participates in the representation of more than one pattern.
To convert from L to a number of training patterns, we must multiply L by the approximate number of patterns per dendritic learning event, or “learning interval” 1PL, where PL is the probability that an arbitrary dendrite learns a particular pattern. This gives an expression for capacity:
C≈LPL=1PL[log(1−fS)log(1−θLpreK⋅μburst)−1]
(2)
Although PL is conceptually simple, its expression is complicated since it depends on pattern density, noise level, two learning thresholds, dendrite size, and fS, and so it is omitted here for clarity (see in the Methods section for the full expression and some discussion).
The expression for C measures how long patterns persist in memory, but a different calculation is needed in order to predict the memory’s recognition performance, that is, the false positive and false negative error rates ϵ+ and ϵ- that we can expect to obtain during a pattern’s lifetime. These error rates depend on the separation of the distributions of responses to trained vs. untrained patterns (Fig 1). These two distributions can be computed from the network parameters to determine whether the allowable error rate tolerances θ+ and θ- will be met during the lifetime calculated in Eq 2 (see Methods).
How can the expression for online storage capacity (Eq 2) be exploited? Given that one of the unique features of our model is that dendrites are the learning units, we used the model to determine how capacity varies with dendrite size, which in turn allows us to determine the optimal dendrite size. In particular, we asked: for a fixed total number of synapses in the memory network (NS=M∙K), if the goal is to maximize online storage capacity, is it better to have many short dendrites (i.e. large M, small K), a few long dendrites (small M, large K), or something in between? Furthermore, how does the optimal dendrite size vary with properties of the input patterns, such as pattern density and input noise level? To address these questions, we fixed network parameters Nsand fs and then for varying combinations of the pattern-related parameters (fA,Nburst,Pburst), we computed C as a function of dendrite size K, using values of the learning, firing, and recognition thresholds (θLpost,θLpre,θF,θR) optimized for each value of K through a semi-automated grid search. The “optimal” dendrite size under a particular set of input conditions was the value of K that maximized capacity, subject to the constraint that immediately after training, responses to trained patterns were strong enough, and responses to random patterns were weak enough, that both the false positive (ϵ+) and false negative (ϵ-) error rates fell below specified tolerances (we used 1% for both). Note that though K appears explicitly only once in Eq 2, as a result of the capacity optimization process, all of the thresholds, and consequently θLpre and PL in Eq 2 depend implicitly on K. The net effect of these dependencies is analyzed in detail in the sections below on penalties for long and short dendrites.
Capacity is plotted in Fig 4a as a function of K for pattern density values ranging from 0.8% to 18%. In the case with fA=1.5%, capacity peaked at ~30,000 patterns when dendrites each contained 256 synapses, and declined substantially for both short (K<100) and long (K>1000) dendrites. As the pattern density increased (to 18%) or decreased (to 0.8%), peak capacity varied nearly 5-fold, favoring sparser patterns, but over the more than 20-fold range of pattern densities tested, peak capacity always occurred for dendrites ranging from 100–500 synapses (grey shaded area). Focusing on the high-capacity (sparse) end of the range with fA<3%, peak capacity was confined to the narrower range of 200–500 (i.e. “a few hundred”) synapses. We also observed that sparser patterns led to a preference for longer dendrites, an effect we unpack below using full network simulations. It is important to clarify that the higher recognition capacity seen for sparser patterns does not result from the fact that sparser patterns contain less information, thereby reducing storage costs per pattern (see S1 Text). We also note that in the more realistic conditions modeled in the full network simulations (see below and Fig 5), peak capacity saturates at slightly higher pattern activation densities (around 1.5%) than is predicted by the analytical model, and the optimal pattern density may be higher still under conditions of increased background noise (S1 Fig shows strong susceptibility to background noise even at 3% pattern density).
To test the effect of pattern noise on capacity, we varied the input noise level by choosing combinations of Nburst and Pburst whose product was always μburst=4 spikes, but that yielded narrow or broad spike count distributions for each active pattern component (Fig 4b, see histogram insets). In this way, we varied the degree to which a trained pattern resembled itself upon repeated presentations. The variation in event counts arising from the above scheme could be viewed as representing either variation in the number of action potentials arriving at the presynaptic terminal from trial to trial, or variation in the number of synaptic release events caused by a given number of action potentials, or a combination of both effects. As expected, higher noise levels reduced peak capacity (Fig 4b), except in the long dendrite range (K>1000) where central limit effects rendered dendrites insensitive to this type of noise. In keeping with this effect, optimal dendrite size increased slightly as the noise level increased, but again, peak capacity was consistently seen for dendrites in the “few hundred” synapse range. Even higher levels of noise were not considered because a simple, biologically available saturation strategy that maps multiple release events into a relatively constant post-synaptic response can largely mitigate the effects of this type of noise. (We did not include a multi-input saturation mechanism in our model to avoid the added complexity).
To verify that the preference for dendrites in the few hundred synapse range was not an artifact of “small” network size, we generated capacity curves from Eq 2 for networks scaled up 256-fold from a base size of N = 5.12 million synapses to ~1.3 billion synapses. The results are shown on a log plot in Fig 4c. As shown in Fig 4d, the scaling power for dendrite sizes K = 64, 256, and 1024 were, respectively, 0.98, 0.97, and 0.97, confirming earlier observations that storage capacity in an optimized dendrite-based memory grows essentially linearly with network size [7]. All the while, the preference for dendrites containing a few hundred synapses remained essentially invariant.
To cross-check the results of the analytical model, we simulated a full memory network, and measured capacity empirically as a function of K. Unlike the analytical case, in which capacity was assumed to be proportional to the calculated length of dendritic age queues, in the network simulations we performed explicit old-new recognition memory tests, and optimized system parameters to achieve false positive and false negative error rates of 1%. In the interests of greater biological realism, we replaced the hard dendritic firing threshold and binary input-output function with a continuous sigmoidal input-output function given by 11+e-sx-θF, and optimized over the slope parameter s along with the 4 threshold parameters. In addition, we relaxed the strict assumption of the analytical model that every input to the network was statistically independent of every other, and instead arranged for each input axon to form ρ synaptic contacts within the memory area, rather than just one. This “redundancy” factor, ρ, set by default to 200, introduced some degree of correlation in the input patterns, and lowered peak capacity somewhat, but had no effect on our main conclusions.
Fig 5a depicts one such simulation with 5.12 million synapses. In the top panel, blue dots show responses to trained patterns, red dots show responses to randomly drawn (untrained) patterns that establish the baseline trace strength (green dashed line) above which stored pattern traces must rise to be recognized. Consistent with the analytical model, responses to trained patterns remain essentially constant during an extended post-training period, in this example spanning ~10,000 patterns. After the flat post-training phase, in contrast to the relatively abrupt fall in trace strength envisioned by the analytical model, a more gradual decline is seen, reflecting the variable times at which the synapses encoding each pattern reach the end of their age queues in different dendrites. Note that the false negative error rate begins to climb during this trace decay period, as the lower fringe of the trained response distribution (blue) progressively merges with the untrained background distribution (red). In this simulation, capacity was reached at ~21,000 patterns, which by our specification is the point where both false positive and false negative error rates equaled 1%. Mirroring the approach taken with the analytical model, multiple simulations were run with varying firing, learning, and recognition thresholds to find the combination of parameters that maximized capacity for each value of K, subject to the same error rate constraints as before. As an additional check of the analytical model, we histogrammed synapse ages within a dendrite (for many dendrites) (Fig 5b), and found that they conformed to a geometric distribution as predicted (red line shows a fitted exponential decay), up to the “cliff” at the end of the age queue (blue dashed line).
Capacity was measured for dendrite sizes between 32 and 4,096 synapses, and the results are shown in Fig 5c and 5d, which are the analogues of Fig 4a and 4b, respectively. When compared to the curves produced by the analytical model, the capacity curves produced by full network simulations had similarly placed capacity peaks and similar qualitative dependence on pattern density and noise levels. In one minor difference, we noted that under the more realistic conditions modeled in the full network simulations, peak capacity saturated at slightly higher pattern activation densities (around 1.5%) than was predicted by the analytical model (Fig 4a).
To determine whether the predictions regarding optimal dendrite size would survive under even more challenging “real world” operating conditions, we added increasing amounts of background noise (spurious spikes added to nominally inactive pattern components), on top of the pre-existing burst noise and pattern correlations. As in the case of burst noise, the background noise level varied between 2 extremes: zero noise, which maximized capacity, and a “high noise” level that reduced storage capacity by roughly a factor of 2 compared to the no-noise case. As in the case of burst noise, we did not consider very high noise levels on the grounds that the deleterious effects of background noise can be compensated by a relatively simple mechanism, for which there is evidence: pre-synaptic terminals with low release probability for “singleton” spikes, along with paired pulse facilitation [89], would allow the effects of sporadic background spikes to be suppressed while maintaining strong responses to signal-carrying bursts. Even at background noise levels capable of causing a significant reduction in peak capacity, the effect of background noise on optimal dendrite size was negligible (S1 Fig). Only at very high levels of background noise, where capacity was reduced more than twofold, did optimal dendrite size change significantly, moving outside of the of the “few hundred” synapses per dendrite range (S1 Fig).
Next we examined the effect of increasing correlations in the input patterns. Given that a single axon can in fact form many thousands of synaptic contacts, corresponding to a much higher redundancy factor than we used in our base simulation, we ran simulations using redundancy factors ρ=5,000 and ρ=10,000 (Fig 5f), which meant that groups of 5,000 or 10,000 synapses scattered across the memory were activated identically. Given previous reports that input correlations can be very deleterious to capacity [10], we speculated that these drastic reductions in the effective dimensionality of the input patterns would severely challenge a memory architecture that was designed to perform optimally with random inputs, or at least significantly alter its behavior. As shown in Fig 5f, however, even in the high-redundancy case (with a 10,000-fold reduction in input space dimensionality), peak capacity dropped by only a factor of ~2 compared to the case with ρ=200, with little to no change in optimal dendrite size.
We next took advantage of the full network simulations to probe the mechanisms that lead to the capacity costs associated with both short and long dendrites. Fig 5e shows two important quantities: the average number of dendrites (μLD) and synapses (μLS) used to store a single pattern in the simulations from Fig 5c. The significance of these quantities is discussed below as we work through the distinct capacity penalties for long and short dendrites.
As shown in Fig 5e, as dendrites grow longer, dendrite usage per stored pattern drops from a value around 10 (at peak capacity) to a “floor” of roughly ~7 dendrites at the long-dendrite end of the range, whereas synapse usage climbs steeply from a baseline of around 150 synapses. To understand the source of the lower bound of ~7 on the average number of dendrites used to store each pattern, it is useful to consider the situation that holds when, in the interests of resource efficiency, we attempt to store each pattern with the minimum possible trace strength: one dendrite. One dendrite firing in response to a familiar pattern is in principle sufficient for recognition, if it is reliable (i.e. occurs > 99% of the time), and if the network’s response to untrained patterns is reliably zero (i.e. > 99% of the time). In a large network, given that each dendrite participates in learning with equal (small) probability, the distribution of the number of dendrites that undergoes a learning event is approximately Poisson with mean μLD=PL·M. Given that a Poisson distribution is characterized fully by its mean, setting μLD=1 by adjusting the learning thresholds, which control PL, means that one dendrite will undergo a learning event for each presented pattern–on average–which is the goal. However, with a mean of 1, the probability that zero dendrites learn is surprisingly high: ~37% (Fig 6a, top plot). Thus, in aiming to use a single dendrite to encode a pattern on average, more than a third of all patterns presented to the network would produce no memory trace at all, leading to a false negative error rate far above the 1% acceptable threshold. To avoid this pitfall, it is critical to reduce the probability to below 1% that zero dendrites learn, which according to the Poisson distribution requires a mean μLD=5 dendrites. This requires a remarkable 5-fold increase in PL relative to the theoretical minimum, with a corresponding 5x increase in synapse resource consumption (Fig 6a, middle plot). Worse, given increased variability in the number of learning dendrites as well as increased readout failures due to input noise and correlations, storage capacity turns out to be maximized when an even higher value of PLis used, achieved by further loosening the learning thresholds, which for our combination of system parameters leads to the empirically obtained optimal value of μLD=~7 dendrites at the long-dendrite end of the range. Given this floor of ~7 dendrites, it becomes clear why synapse usage increases as dendrites grow longer: the number of synapses used in a dendrite that undergoes a learning event is roughly proportional to the dendrite length K, since the number of synapses that learn is roughly proportional to the number of synapses activated in the dendrite, which is proportional to dendrite size. Tied to this increase in synapse usage per pattern, as the total number of dendrites M in the system decreases (because each one contains a larger fraction of the synapses), the frequency with which each dendrite must participate in learning increases, which speeds the per-pattern rate at which synapses move along their age queues. Thus, from a capacity standpoint, it is ideal to choose system parameters such that the minimum encoding bound of 7 dendrites is actually used (or whatever minimum number of dendrites is needed, given the settings of the error rate thresholds and noise level), but having met this lower bound, dendrites should be kept as short as possible.
The reasons capacity declines as dendrites grow shorter are complex, and are discussed only briefly here (see the S1 Text and S3 and S4 Figs for more details). We first consider why dendrite usage increases for short dendrites, rather than remaining at the minimal encoding bound. Short dendrites are intrinsically more susceptible to variability in crossing their learning and firing thresholds, since fewer active synapses are involved. As dendrites become very short, this requires the network to increase dendrite usage far above the nominal lower bound of μLD=5. For example, under sparse activation (fA=1%), medium noise conditions (Pburst=47,Nburst=7) with dendrites containing ~200 synapses, when the system is optimized for capacity, μLD≈15 (blue solid curve in Fig 5e), substantially more than the number of dendrites used under maximum capacity conditions. While this increase in dendrite usage is more than offset by the reduced dendrite size, which tends to reduce synapse usage, the total number of synapses altered during learning in fact remains approximately constant, implying that a larger fraction of synapses is modified within each short dendrite that engages in learning. This higher synapse burn rate in short dendrites leads to shorter age queues, and in the end lowers capacity.
The memory architecture we have studied is ordinary, in the sense that it consists of axons making contacts directly onto the neurons whose firing represents the memory trace, but is out-of-the-ordinary among online learning models in that it includes a layer of thresholded dendritic units interposed between the input axons and the final common output of the network.
The main contributions of this paper are (1) Eq 2, which captures the interactions between key variables that influence storage capacity in a dendrite-based online recognition memory, and (2) our showing that over a wide range of input pattern statistics and network sizes, memory capacity is maximized when dendrites contain a few hundred synapses, which corresponds to the typical dendrite size found in medial temporal lobe memory areas [90]. To our knowledge, ours is the first theory that accounts for dendrite size in terms of its role in optimizing online learning capacity.
Beyond the uses we have shown here, our model could potentially be used (1) to help explain why different combinations of parameter settings co-occur in different recognition memory-related brain areas, for example in different animal species whose brains may be larger or smaller, whose sensory codes may be sparser or denser, or whose error tolerances may be tighter or looser; (2) to help distinguish brain areas involved in online familiarity-based recognition memory, the task we study here, from areas such as the hippocampus that (also) contribute to explicit recall [87,88]; and (3) to help identify which changes (e.g., spine loss, dendrite retraction, hyperexcitability, etc.) that occur in neurological disorders, aging and stress, are most directly responsible for producing memory deficits–knowledge that may eventually aid in the design of clinical interventions for those suffering from memory loss.
Why are dendrites of “medium” size optimal for storage capacity in the context of an online familiarity-based recognition memory? The simplest explanation is that short dendrites suffer from one set of disadvantages, and long dendrites suffer from another, leaving the optimal dendrite size somewhere in the middle. Short dendrites have relatively noisier post-synaptic response distributions because fewer synapses contribute to the response. As a result, a larger fraction of the synapses on a short dendrite must be modified during learning to ensure that the dendrite's response to previously trained patterns remains comfortably at the upper tail of the untrained pattern response distribution. Increasing the fraction of synapses used within a dendrite during each learning event shortens the dendrite's age queue, which comes at a capacity cost. This effect leads to a preference for longer dendrites.
But long dendrites also have their disadvantages. An online recognition memory should aim to store the weakest possible trace of each learned pattern, which in our framework corresponds to learning in a small number of dendrites near the "minimum encoding bound" (corresponding to ~7 dendrites under the conditions used in our study; see Fig 5e). This means that the longer the dendrites become, the more synaptic resources are consumed by each dendrite that learns, since the number of synapses used per dendrite during a learning event is roughly proportional to dendrite size. Clearly from this perspective, it's best to keep dendrites as short as possible.
The compromise between the need to keep dendrites long enough to avoid noise and age queue problems, and short enough to avoid excessive synapse use per learning dendrite, puts the optimal size around a few hundred synapses for biologically reasonable values of pattern activation density and noise. Of course, our assumptions regarding "biologically reasonable" pattern activation densities and noise levels are informed guesses rather than certain knowledge, and are not likely to be universal constants across brain areas, species and operating conditions. It is therefore possible that the natural dendrite sizes found in medial temporal lobe memory areas are determined in part by factors other than capacity optimization according to Eq 2. For example, developmental constraints, energy constraints, space constraints, and combinations thereof, may have been responsible for pushing the actual dendrite size in one direction or another, away from the optimal length as determined by capacity considerations alone. Nonetheless, it is useful to capture basic relationships between biophysical parameters, wiring parameters, input pattern statistics, and capacity, as a starting point for a more complete online memory model.
That mid-sized dendrites optimize capacity can be understood from another perspective. Eq 2 shows capacity is given by the ratio of L, the length of a dendrite's age queue, to PL, the probability that a dendrite learns. PL, in the denominator, grows larger as dendrites grow in size because the same average number of dendrites is always used to learn, but when dendrites are long, there are fewer of them to choose from. L, in the numerator, grows smaller as dendrites shrink in size because of the higher value of fpot needed to compensate for noise effects. Balancing these two effects, capacity is maximized for dendrites of intermediate size, for which L is not too small, and PL is not too large.
Thus, among the many roles that dendrites may play in the brain, in the context of an online familiarity-based recognition memory, separately thresholded dendrites play the critical role that they downsize the learning units from neuron-sized units (~20,000 synapses) to units containing a few hundred synapses, which are much more numerous, while still containing enough synapses to avoid the capacity costs associated with noise effects and shortened age queues. Simply put, having separately thresholded dendrites provides the memory system with more learning units of a better size. If dendrite-sized learning units were not available, so that it was necessary to construct an online recognition memory from neuron-sized units, storage capacity would be cut by an order of magnitude or more (see Fig 5c).
A general theme that emerges from our study is the importance of variability control for a recognition memory. The goal of a neural-style online recognition memory is to store a trace of each learned pattern that consumes as few synaptic resources as possible, but that nonetheless allows the network to produce a reliable recognition response on future encounters with a stored pattern. Variability in the magnitude of network responses to either learned or unlearned patterns, such as that produced by burst noise, or low pattern density, complicates this goal in at least two ways. First, increased variability in the responses to unlearned patterns raises the level of background noise, and thus the required minimum encoded signal strength that learned patterns must obtain. This in turn increases the number of synapses that must be devoted to storing each new memory. Second, increased variability in signal strength for learned patterns increases the rate of readout failures (for fixed firing and recognition thresholds). This increase in false negative errors must again be compensated for by increasing memory trace strength for all patterns, which wastefully strengthens patterns whose traces were already well above the recognition threshold.
These effects imply that a brain system devoted to recognition memory is under intense pressure to include response normalization mechanisms, presumably involving local inhibitory circuits [91–99].
It is intriguing to note that if network behavior could be perfectly normalized, so that every pattern is stored by learning in the exact same number of dendrites, e.g. 1 dendrite, then this would represent a 7-fold resource savings, presumably leading to a corresponding boost in capacity compared to the peak capacity conditions shown in Fig 5 (where an optimized high capacity network chooses to learn using 7 dendrites).
Several of the mechanisms and processes in our dendrite-based learning scheme are consistent with known biological mechanisms, including that:
The main speculative/predictive features of our model pertain to the specific conditions for LTP and LTD. First, following [7] we assumed here that the triggering of a learning event in a dendrite, which induces both LTP and LTD, depends on a compound threshold: in order to learn, a dendrite must both (1) receive an unusually strong presynaptic input, that is, unusually many axons impinging on the dendrite must be firing and releasing glutamate; and (2) experience an unusually strong post-synaptic response, that is, unusually many of the firing axons must be driving synapses that are already in a strong state. Note that a traditional Hebbian learning rule would tie learning to the post-synaptic response alone (∑wixi), placing no explicit condition on the number of axons participating (∑xi). The pre-synaptic condition was incorporated into our model opportunistically, when we observed that doing so doubled the memory's storage capacity [7]. We call the existence of a compound learning threshold a "prediction" of our model on the grounds that the brain would have been under evolutionary pressure to discover any small functional modifications that significantly boost storage capacity, and so the brain might have “discovered” this optimization–as we did. The prediction is weak, however, given that the memory can function in basically the same fashion with a single, conventional post-synaptic threshold, albeit with reduced capacity.
Unlike our weak prediction of a compound dendritic learning threshold, which could be falsified without dire consequences for the model, the prediction that synapses involved in an online familiarity memory should have a prescribed lifetime in the potentiated state, after which they are actively depotentiated, is a more deeply rooted feature of our model. This prediction is also a nearly inevitable consequence of the statement of the learning problem itself: any online recognition memory whose memory retention is much shorter than the animal's lifetime will be "full" at all times, except for a transient period at the beginning of the animal's life when the memory is first filling up. Once it reaches its chronically full state, each time a new pattern is written into the memory by strengthening synapses, as a matter of homeostatic necessity the equivalent of one stored pattern must be erased by weakening synapses, and in the interests of optimal performance, that one erased pattern should be the oldest stored pattern. The alternative–partially degrading many patterns of varying ages–is a poor strategy for a recognition memory, since any pattern whose signal strength is prematurely degraded to the point where it falls below the recognition threshold is functionally lost, yet its unerased detritus continues to uselessly consume space in the memory. Furthermore, since it is most efficient from a resource allocation point of view to store memory traces that are just strong enough to cross the recognition threshold, and no stronger, the system cannot abide gradual attrition of pattern traces. Thus the problem statement itself, and simple logic, dictate that a memory network in the brain devoted to online familiarity/recognition memory should attempt to target the oldest information for erasure as each new pattern is stored. It is difficult to imagine how selective erasure of old information could occur unless synapses keep track of their ages, and unless a dendrite is able to target its oldest synapses for depression as it undergoes each new learning event.
Age-based depression of synapses was previously explored as a strategy for increasing online learning capacity in the context of a 1-layer Willshaw network [5]. It is only in the context of a 2-layer memory, however, in which synaptic learning probabilities can be driven down to extremely low values without compromising signal strength, that synapses are given the opportunity to actually grow old [7].
In the 2-layer dendrite-based memory scheme we have studied, storage capacity is increased (~linearly) by increasing the number of dendrites, without altering the synapse model or the plasticity rule. As an alternative, Stefano Fusi and colleagues have developed two elegant models of online learning that boost capacity instead by increasing the complexity of individual synapses [4,8]. Both models share the following basic framework: the memory consists of N synapses abstracted away from any particular network architecture; by default, every synapse is modified during the storage of every pattern; to store a pattern, synapses are strengthened and weakened in equal numbers; and all instructed weight changes during pattern storage overwrite previously stored information. The goal of these models is to carefully manage the plasticity-stability tradeoff that exists when each synapse is asked to encode information about many patterns that have been stored over time: synapses that are very plastic are good at rapidly storing new information but poor at preserving old information, whereas synapses that are very stable are good at preserving old information but poor at rapidly storing new information (synopsis adapted from [8]).
In the "Cascade" model [4], synaptic weights are binary valued (strong and weak), but can exist in states of varying lability/stability. The state diagram within each synapse operates according to two main principles. First, repeated potentiation instructions push a strong synapse into an increasingly stable strong state, that is, a state that shows an increasing resistance to depression. Similarly, repeated depression instructions by the learning rule have the effect of pushing a weak synapse into an ever more stable weak state, one that increasingly resists potentiation. Second, at "deeper" levels of the cascade, corresponding to more stable strong and weak states, the transitions to even deeper levels corresponding to even more stable states, and the transitions in synaptic weight value from strong to weak or weak to strong, all become increasingly improbable, so that synapses in deeper cascade states remain stable over longer and longer time scales. The variation in these transition probabilities across cascade levels can be considerable: according to [4] the optimal cascade model with 10^6 synapses has 15 cascade levels. With this many levels, the most labile synapses at the top of the cascade change weight with probability 1 (i.e. deterministically) in response to a weight change instruction, whereas the most stable synapses deep in the cascade only change weight with probability 1/16,384 in response to a weight-change instruction. Thus, a weak synapse in its most stable state would need to receive ~10,000 potentiation instructions in a row in order to reach a 50% chance of actually undergoing potentiation.
These two operating principles of the Cascade model are clearly distinguishable from those governing synaptic plasticity in our model. First, in the Cascade model, all synaptic state transitions are probabilistic, whereas in the dendrite-based model, all synaptic state changes are deterministic: during learning, weak synapses receiving the instruction to potentiate do so fully and immediately, and during forgetting, strong synapses that reach the end of their lifetimes are fully and immediately depressed. The logic of synapse durability is also different in the Cascade vs. dendrite-based models. In the Cascade model, when a synapse is first potentiated, it is in its most labile strong state, and therefore most vulnerable to depression. In the dendrite-based model, a synapse that has just been potentiated is in its most durable state, in the sense that it will withstand the largest number of consecutive learning events in which it does not participate before it "ages out" and finally succumbs to synaptic depression.
In the Benna and Fusi model [8], the machinery contained within each synapse consists (metaphorically) of a chain of connected fluid-filled beakers. The first beaker represents the synapse’s (graded) strength value by the level of virtual liquid relative to equilibrium, and the last beaker is tied to the equilibrium liquid level. Synaptic potentiation occurs deterministically, and consists of adding a fixed amount of liquid "weight" to the first beaker; synaptic depression consists of removing that amount of liquid from the first beaker. The equilibration of liquid levels in the beaker chain following an instructed weight change, and particularly the equilibration of the first beaker, captures the time course of the memory decay at each synapse. In the example shown in [8], a synapse consisted of a chain of 12 virtual beakers that doubled in capacity with each step down the chain (so that the last beaker had a capacity 2,048 times that of the first beaker), and whose fluid levels were governed by differential equations with pre-determined rate constants linking each pair of buckets. As a practical matter, the authors found the number of discrete levels per beaker could be reduced linearly from 35 in the first (smallest) beaker, corresponding to 35 levels of visible synaptic weight, down to 2 levels in the last (largest) beaker. This parameterization yielded a total of ~10^14 possible memory states within each synapse. Interestingly, unlike the cascade model whose synapses only change state in response to plasticity instructions (which can occur asynchronously), the chain-of-beakers model, if taken literally, continues to equilibrate—i.e. forget—even during periods when the rate of new learning slows or stops, such as during quiet wakefulness or sleep. Thus, an additional layer of mechanism is presumably needed to modulate the inter-beaker flow rates in a coordinated fashion depending on the external learning rate.
In summary, both of these models [4,8] achieve longer memory lifetimes by increasing the complexity of the synapse model as the size of the memory increases. In terms of cost, the machinery inside these more complex synapses requires more parameters (>10), and those parameters must span large dynamic ranges (>1000) to reach realistic memory sizes.
How does a dendrite-based model grow storage capacity without increasing the complexity of the individual synapses? Within virtually any recognition memory model, the conceptually simplest way to increase storage capacity is to reduce the fraction of synapses that are modified during the storage of each pattern (the signal), while correspondingly reducing the response of the memory to random input patterns (the noise). Practically, this can be achieved by sparsifying the input patterns inversely with pattern size as the memory grows larger. Thus, if the memory increases in size from N to c·N synapses, in order to increase capacity c-fold, the pattern density 'a' must be reduced c-fold so that the same number of synapses is activated by each pattern as before. Assuming the learning rule instructs each activated synapse to become strong if it was weak, a·N/2 weak synapses would be potentiated on average (under the assumption that half of the synapses are strong), and an equal number of strong synapses would be depressed to maintain homeostasis (drawn from the N/2 strong synapses). To a rough approximation, this leads to a capacity of ~1/a patterns. Thus, if a = 1% of synapses are changed during the storage of each pattern, then after ~100 patterns are stored, the memory will have turned over completely. This simple scaling approach runs into the biological plausibility problem that very large capacities require very low pattern densities, and very low depression probabilities. To achieve a capacity of 100,000 patterns, for example, only 1 in 100,000 input neurons could be active, and synaptic depression would occur in only 1 in 100,000 strong synapses. Reliably controlling such small activity and plasticity probabilities could be difficult to achieve in neural tissue.
As an alternative both to this very simple sparsification approach, and to the "complex synapse" approach developed by Fusi and colleagues, adding a layer of dendritic learning units allows the memory to push further into the sparse plasticity regime without the need for very low pattern densities or plasticity probabilities. Relative to a flat (1-layer) memory model, dendritic learning thresholds can restrict learning to just a few dendrites from a very large pool. For example, in a simulation of a 5 million-synapse network discussed previously, with a moderate pattern sparseness level of a = 3%, the dendrite learning probability after optimization was PL = 0.0005, (corresponding to 1–2% of neurons in the network having one dendrite that crosses the learning threshold). Beyond the sparsification of learning attributable to dendritic learning thresholds, learning is sparsified even further by the fact that within each learning dendrite, only the active 3% of synapses receives (and obeys) the instruction to potentiate or refresh, and that same small fraction of synapses is depressed. Thus, in the above scenario, relative to a 1-layer network with the same coding density of 3%, the existence of a dendritic learning threshold sparsifies learning by a factor of 1/PL = 2000, significantly boosting capacity without requiring extreme, biologically unrealistic coding sparseness.
In our model the formation of new memories is achieved through long-term potentiation (or rejuvenation) of a few activated synapses on a few strongly activated dendrites that undergo learning events. The "forgetting" of old memories involves heterosynaptic depression of the least-recently-potentiated/rejuvenated synapses in the same dendrites that are undergoing learning. Given the pressure to keep memory traces at their bare minimum strength, when our model is optimized for capacity, synaptic changes are exceedingly sparse, involving only a small fraction of the synapses on a minute fraction of dendrites. (The finding that memory capacity is optimized by sparse patterns has also been reported for 1-layer models: [2,111–114]). For example, in a memory network containing ~5 million synapses, under conditions that optimize storage capacity (i.e. with dendrites containing ~256 synapses, and patterns of 3% density), we found that each time a pattern is learned, only 150 of the 5 million synapses learn (0.003%), less than half of which are overtly strengthened (i.e. some are only rejuvenated), and those few altered synapses are confined to just 10 of the 20,000 dendrites contained within the network. If we consider extremely sparse synaptic plasticity to be a prediction of our model, could such sparse changes be detected experimentally? The likelihood of detecting changes in this few dendrites seems higher when it is considered that 20,000 dendrites corresponds to 500–1,000 neurons. We would thus expect that 10 (i.e. 1–2%) of the neurons in the network would contain a dendrite that participates in learning. In vivo imaging techniques with a field of view containing hundreds of neurons should make this level of detection possible.
What role might structural plasticity play in online learning? We previously explored the role that active dendrites might play in familiarity-based recognition in the very different scenario where patterns can be trained repeatedly [46,115]. The opportunity for repeated, interleaved training of patterns gives the system time to exploit wiring plasticity mechanisms [116], wherein existing connections between axons and dendrites can be eliminated and new ones formed in such a way that correlated inputs end up forming contacts onto the same dendrites. This type of wiring plasticity is not an option in an online learning scenario, since each pattern is experienced only once, such that all learning-related synaptic changes must be immediate–or at least immediately induced. We showed that correlation-based sorting of inputs onto different dendrites using a Hebb-type learning rule increased the storage capacity of a neuron by more than an order of magnitude compared to a neuron with the same total number of synaptic inputs that lacked dendrites. Furthermore, as here, we found that dendrites of intermediate size optimized capacity–though for different reasons.
It is interesting to note that in our current model, structural turnover of weak synapses has no effect on what is stored in the memory, as long as new weak synapses are added to the system at the same rate that existing weak synapses are removed. If weak synapses form a substantial fraction of the total synapse population–we have assumed 50% here (but the percentage may actually be closer to 90% in CA1 –see [117])–then high rates of spine elimination and new spine formation could be tolerated within the memory area without any loss of stored information–again, as long as the turnover is restricted to weak synapses. What would be the advantage of eliminating existing weak connections and forming new ones? Under the assumption that input axons are uncorrelated, as we have assumed in this work for simplicity, we can see no advantage to this type of structural turnover. However, if meaningful correlations between input axons do exist, then structural turnover could be a sign that wiring plasticity mechanisms are attempting to co-locate correlated synapses on the same dendrites [118,119], which could lead to a significant capacity advantage [46,115,116].
Familiarity-based recognition is a very basic form of memory, and is most closely associated with the perirhinal cortex [10,87,88]. However, currently available data regarding the responses of familiarity (vs. novelty) neurons in the PRC is complex, and not easily related to our findings here (see S1 Text for an in depth discussion). Further work will be required to determine whether the dendrite-based architecture of Fig 2b will be helpful in explaining familiarity-based recognition processes in the brain.
What can the dendrite-based architecture we have studied here tell us about other types of memory systems? A trivial extension of our architecture in which N copies of the memory network are concatenated would allow the construction of a full N-bit binary online associative memory. This type of memory would behave exactly as ours, but would allow an arbitrary N-bit output pattern to be one-shot associated with each input pattern, as in a Willshaw network. In this scenario, only the subset of the N networks whose outputs are instructed to be 1 would learn each input pattern, while any networks instructed to produce 0 responses would simply ignore the input pattern. If the output patterns are sparse (which they needn't be), only a small fraction of the networks would need to participate in the learning of each association.
It might also be desirable to assign extended lifetimes to particularly important patterns; this could be accomplished in either of two ways: 1) Extended-lifetime synapses could be established during the learning of important patterns, so that the synapses representing those patterns would remain invulnerable to depotentiation for longer times, or even permanently. Doing so would of course reduce the lifetimes of other patterns in the memory. 2) The memory could be composed of multiple subnetworks having a range of pattern lifetimes, and important patterns could be stored in longer-lifetime (i.e. larger capacity or more rarely used) networks. The decision as to which or how many networks participate in the storage of each pattern could be gated by an "importance" signal provided by another brain area.
In other cases it might be valuable to store different trace strengths for different patterns, rather than uniform, bare-bones recognition traces for all patterns. Note this goal is inconsistent with the goal to maximize storage lifetimes for all patterns, but could also be useful in certain ecological situations. Our simple architecture allows for this directly: nothing is to prevent a larger or small number of dendrites from being used in the learning of any particular pattern, such that it's memory trace would be stronger or weaker than the norm. Regardless of trace strength, a pattern’s lifetime would remain roughly the same, since lifetimes are determined mainly by the lengths of the dendritic age queues, which do not depend on the number of dendrites used for storage. The trace strength assigned to each pattern could again be determined by a signal generated by another brain area, whose effect is to raise or lower dendritic learning thresholds.
In yet another scenario it might be useful to store gradually decaying memory traces so that trace strength can represent recency of learning (which is again a different goal than maximizing recognition capacity). A graded recency signal can be efficiently produced by storing each pattern simultaneously in multiple networks with a range of capacities/sizes/memory lifetimes. Early in its storage lifetime, the pattern would evoke a memory trace from all networks, so that it's total trace strength would be high, but as time progresses, and its trace progressively expires from the lower-capacity networks, its overall trace strength would gradually decay. This use of such a tiered system to achieve a graded decay time course is more resource-efficient than certain other forms of trace decay that have been considered in the online memory literature, in that the stored information in a tiered network with synapse age management expires in a controlled fashion [109].
Finally, it will require future work to determine which of our results can carry over to Hopfield-style recurrent networks [120–123] constructed from neurons with thresholded dendrites, where the goal in that case would be to maximize recall capacity. In one obvious difference, the ability to recall entire patterns from partial cues requires that the entire patterns be stored (in stark contrast to the need to generate only a reliable familiarity signal), so synapse resource consumption per pattern will be much higher than in the basic familiarity network. Furthermore, the need to modify recurrent synapses during the initial learning of a pattern implies that the participating neurons must fire action potentials during initial learning in order to activate those recurrent connections, which implies that their dendrites must cross both the learning and firing thresholds during learning. Interestingly, this requirement would seem to render such a memory useless for familiarity-based recognition, since the neurons that participate in the learning of a pattern must already fire on a pattern’s first presentation to the memory. This incompatibility could be one reason why the functions of familiarity and recall memory have been assigned to distinct areas within the medial temporal lobe [87,88].
As discussed in the main text, after a certain number of learning events has occurred following the storage of a pattern feature in a dendrite, the strong synapses encoding the stored feature begin to “fall off” the end of the dendrite’s age queue, and the memory trace in the dendrite is effectively lost. We refer to the number of learning events that can be endured before this loss occurs as the length of the age queue L. If we assume that the frequency of learning events is constant across dendrites in the network, given that the queue length L is also constant across dendrites, most of the strong synapses encoding a particular pattern’s features will be depressed roughly simultaneously (in different dendrites), leading to a relatively rapid decay of the network’s overall response r to that pattern. The value of L is therefore a measure of the length of time that a pattern’s trace persists in the memory, and is therefore effectively a measure of capacity in units of dendritic learning events.
L can, in principle, be determined by framing learning as a Markov process with the state diagram shown in Fig 3. Consider a single synapse on a given dendrite. If p→ is the (L+1)×1 vector containing the probability that, at a given time, this synapse is in each of the L+1 states shown in Fig 3, and T is the L+1×L+1 matrix containing the state transition probabilities, then with each learning event, p→ will change as p→→Tp→. After many learning events, p→ will approach the equilibrium distribution, characterized by the condition that learning leaves it unchanged: p→∞=Tp→∞. Using the fact that for the equilibrium distribution p→∞,fs of the synapses must be strong, one can solve for L (since the (L+1)×1 vector p→∞ implicitly depends on L). Using the eigenvectors and eigenvalues of T, one can also compute the distribution p→t after any number of learning events. However, while the Markov approach is very general, the simple dynamics of the age queue allow for a more direct and transparent derivation of L.
To find L, we might naively divide the total number of strong synapses per dendrite (fS∙K) by the average number of synapses potentiated in each dendrite that experiences a learning event μpot. where μpot≈θLpreμburst. In words, μpot is approximately equal to the total number of spikes impinging on all activated synapses on the dendrite, given by the threshold value θLpre (since in most cases learning dendrites will have just crossed this threshold), divided by the average number of spikes per participating synapse μburst. This gives L≈fS∙K∙μburstθLpre. However, this would underestimate L because synapses that are only juvenated (i.e. that were already strong) do not contribute to the aging of synapses further along the age queue, so that the average rate of progression along the age queue slows as strong synapses grow older. To estimate L more accurately, consider the equilibrium distribution of synapse ages in the queue of a single dendrite (blue histogram in Fig 3). The age of the right-most column of the age histogram is an indicator of the expected age (measured in learning events) at which the synapses encoding a pattern are depressed and moved to the unordered collection of weak synapses. During each learning event, a large fraction (fage) of synapses in each column that were not activated move rightward to the next older column, while a small fraction (1-fage) are juvenated (promoted to the first column). This process leads to a bias towards younger synapses in the queue, and can be well-approximated by a finite geometric sequence with length L, decay ratio fage, sum fS∙K (note the sum of the columns is the total number of strong synapses), and first column height μpot (the average number of synapses that learn per dendrite per learning event), so that:
fS⋅K=μpot⋅1−fageL+11−fage.
Assuming that the synapses in a dendrite are all equally likely to be potentiated (ignoring the effects of the postsynaptic threshold–see below), with μpot≈θLpreμburst, then we have that fage≈1-θLpreK∙μburst and can solve the above equation for L. Note that L counts the number of dendritic learning events before a memory is eroded, whereas memory capacity C should count the number of training patterns. Thus, to approximate C, we must multiply L by the approximate number of patterns per dendritic learning event, or “learning interval” 1PL, where PL is the probability that an arbitrary dendrite learns a particular pattern. Although PL is conceptually simple, its expression is complicated since it depends on pattern density, noise level, two learning thresholds, dendrite size, and fS (see expression below). Collecting these results, we can approximate memory capacity by
C≈LPL=1PL⋅[log(1−fS)log(1−θLpreK⋅μburst)−1].
For simplicity, the expression for L in the capacity equation does not include the effect of the postsynaptic threshold θLpost, which makes strong synapses more likely to learn, lowers fage and increases absolute capacity. The synapse age distribution remains roughly geometric, however (see Fig 5b), and we observed that the qualitative behavior of the system depends only weakly on θLpost, justifying its omission from the analysis.
Synaptic activation on a dendrite is governed by 4 binomial random variables: as, the number of active strong synapses; ss, the number of spikes received by strong synapses; aw, the number of active weak synapses; and sw, the number of spikes received by weak synapses. These random variables have the distributions shown below. Learning occurs when presynaptic activation crosses the presynaptic learning threshold, or ss+sw>θLpre, and postsynaptic activation crosses the postsynaptic learning threshold, or ss>θLpost. Using the distributions for as,aw,ss, and sw, and the fact that PL=pss+sw>θLpre,ss>θLpost,we can write an explicit expression for PL:
as~Bi(fs·K,fA)
ss~Bi(Nburst·as,pburst)
aw~Bi((1−fs)K,fA)
sw~Bi(Nburst·aw,pburst)
PL=∑i∈[0,fs·K]j∈[0,(1−fs·K)]k∈[θLpost+1,Nburst·i]l∈[θLpre−k+1,Nburst·j]Bi(Nburst·j,pburst)[l]⋅Bi(Nburst·i,pburst)[k]⋅Bi(fs·K,fA)[i]⋅Bi((1−fs)K,fA)[j]
where BiN,p[k] is the binomial pdf with parameters (N,p) evaluated at k. A simpler alternative to evaluating this expression directly is to estimate it by generating a large number of samples of as,aw,ss, and sw according to the above distributions, and directly observing the fraction of cases that cross both learning thresholds .
Once the capacity formula is used to calculate how long a given memory trace will last, we must verify that during its lifetime, the trace is sufficiently strong. We do this by checking whether the error tolerances ϵ+ and ϵ- are met immediately after training.
First, we compute ϵ+, the probability that an untrained pattern will be recognized. To be recognized, a pattern must activate at least θR dendrites in the network. For a randomly selected untrained pattern, the distribution of the number of activated dendrites will be approximately Poisson with mean PF⋅M, where M is the number of dendrites in the network and PF is the probability that a given dendrite fires in response to a randomly selected pattern. For a pattern to fire a dendrite, it must cause a postsynaptic activation >θF, or ss>θF, using the notation of above. Since the distribution of ss is known, it is relatively easy to write an expression for PF and ϵ+ explicitly:
PF=p(ss>θF)=∑i∈[0,fs·K]j∈[θF,Nburst·i]Bi(Nburst·i,pburst)[k]⋅Bi(fs·K,fA)[i]
ϵ+=∑r≥θRPoiss(PF⋅M)[r]
As for ϵ-, the probability that a previously trained pattern is forgotten, we approximate this quantity with ϵ-0, or the immediately post-training false negative rate (justified by the fact that during the “lifetime” of the memory, C, the trace strength is roughly constant). To calculate ϵ-0, we use the following observation: when training a new pattern, it will learn in a certain set of dendrites. Immediately after training, if the pattern is re-presentated to the network, all of these dendrites should respond, since learning has significantly boosted the pattern’s features in these dendrites. In other words, dendrite readout failures immediately after learning should be very rare. Therefore, for a pattern to be too weak for recognition immediately after training, it must have learned in too few dendrites. The number of learning dendrites for a given pattern will have a Poisson distribution with mean PF⋅M. Therefore, ϵ- can be written
ϵ−≈ϵ−0=∑l<θRPoiss(PL⋅M)[l]
If for the given settings of the learning and firing thresholds θLpre,θLpost,θF,θR, the error tolerances are met–that is, ϵ+,ϵ-<1%–then the memory lifetime is compared to the best memory lifetime found so far. Otherwise, we continue the search through threshold space.
All data contained in figures as well as simulation code is available in S1 Data file titled "Plos data/code".
|
10.1371/journal.pgen.1002703 | New Insight into the History of Domesticated Apple: Secondary Contribution of the European Wild Apple to the Genome of Cultivated Varieties | The apple is the most common and culturally important fruit crop of temperate areas. The elucidation of its origin and domestication history is therefore of great interest. The wild Central Asian species Malus sieversii has previously been identified as the main contributor to the genome of the cultivated apple (Malus domestica), on the basis of morphological, molecular, and historical evidence. The possible contribution of other wild species present along the Silk Route running from Asia to Western Europe remains a matter of debate, particularly with respect to the contribution of the European wild apple. We used microsatellite markers and an unprecedented large sampling of five Malus species throughout Eurasia (839 accessions from China to Spain) to show that multiple species have contributed to the genetic makeup of domesticated apples. The wild European crabapple M. sylvestris, in particular, was a major secondary contributor. Bidirectional gene flow between the domesticated apple and the European crabapple resulted in the current M. domestica being genetically more closely related to this species than to its Central Asian progenitor, M. sieversii. We found no evidence of a domestication bottleneck or clonal population structure in apples, despite the use of vegetative propagation by grafting. We show that the evolution of domesticated apples occurred over a long time period and involved more than one wild species. Our results support the view that self-incompatibility, a long lifespan, and cultural practices such as selection from open-pollinated seeds have facilitated introgression from wild relatives and the maintenance of genetic variation during domestication. This combination of processes may account for the diversification of several long-lived perennial crops, yielding domestication patterns different from those observed for annual species.
| The apple, one of the most ubiquitous and culturally important temperate fruit crops, provides us with a unique opportunity to study the process of domestication in trees. The number and identity of the progenitors of the domesticated apple and the erosion of genetic diversity associated with the domestication process remain debated. The Central Asian wild apple has been identified as the main progenitor, but other closely related species along the Silk Route running from Asia to Western Europe may have contributed to the genome of the domesticated crop. Using rapidly evolving genetic markers to make inferences about the recent evolutionary history of the domesticated apple, we found that the European crabapple has made an unexpectedly large contribution to the genome of the domesticated apple. Bidirectional gene flow between the domesticated apple and the European crabapple resulted in the domesticated apple being currently more similar genetically to this secondary genepool than to the ancestral progenitor, the Central Asian wild apple. We found that domesticated apples have evolved over long time scales, with contributions from at least two wild species in different geographic areas, with no significant erosion of genetic diversity. This process of domestication and diversification may be common to other fruit trees and contrasts with the models documented for annual crops.
| Domestication is a process of increasing codependence between plants and animals on the one hand, and human societies on the other [1], [2]. The key questions relating to the evolutionary processes underlying domestication concern the identity and geographic origin of the wild progenitors of domesticated species [3], the nature of the genetic changes underlying domestication [4], [5], the tempo and mode of domestication (e.g., rapid transition versus protracted domestication) [6] and the consequences of domestication for the genetic diversity of the domesticated species [7], [8], [9], [10]. An understanding of the domestication process provides insight into the general mechanisms of adaptation and the history of human civilization, but can also guide modern breeding programs aiming to improve crops or livestock species further [11], [12].
Plant domestication has mostly been studied in seed-propagated annual crops, in which strong domestication bottlenecks have often been inferred, especially in selfing annuals, such as foxtail millet, wheat and barley [11], [13], [14], [15], [16], [17]. Genetic data have suggested that domestication or the spread of domesticated traits has been fairly rapid in some annual species (e.g, maize or sunflower), with limited numbers of populations or species contributing to current diversity [10], [18], [19], [20], [21], [22]. In contrast, a combination of genetics and archaeology suggested a protracted model of domestication for other annual crops, and in particular for the origin of wheat or barley in the Fertile Crescent [11], [23]. However, the genetic consequences of domestication have been little investigated in long-lived perennials, such as fruit trees [24], [25], [26]. Trees have several biological features that make them fascinating and original models for investigating domestication: they are outcrossers with a long lifespan and a long juvenile phase, and tree populations are often large and connected by high levels of gene flow [27], [28].
Differences in life-history traits probably result in marked differences in the mode and speed of evolution between trees and seed-propagated selfing annuals [27], [28], [29]. For example, outcrossing may tend to make domestication more difficult, in part because the probability of fixing selected alleles is lower than in selfing crops [6], [13]. The combination of self-incompatibility and a long juvenile phase also results in highly variable progenies, making breeding a slow and expensive process, and rendering crop improvement difficult. The development of vegetative propagation based on cuttings or grafting has been a key element in the domestication of long-lived perennials, allowing the maintenance and spread of superior individuals despite self-incompatibility [30]. However, the use of such techniques has further decreased the number of sexual cycles in tree crops since the initial domestication event, adding to the effect of long juvenile phases in limiting the genetic divergence between cultivated trees and their wild progenitors [30], [31], [32], [33]. Thus, domestication can generally be considered more recent, at least in terms of the number of generations, in fruit tree crops than in seed-propagated selfing annuals.
Given the slow process of selection and the limited number of generations in which humans could exert selection, the protracted nature of the domestication process in trees has probably resulted in limited bottlenecks [25], [31] and in a weaker domestication syndrome [34] than in seed-propagated annuals. Nevertheless, many cultivated fruit trees clearly display morphological, phenotypic and physiological features typical of a domestication syndrome, such as large fruits and high sugar or oil content [32], [35]. Many aspects of fruit tree domestication have been little studied [25]. Consequently, most of the hypotheses concerning the consequences of particular features of trees for their domestication/diversification remain to be tested. Recent studies on grapevines, almond and olive trees have provided illuminating insights, such as the importance of outcrossing and interspecific hybridization [36], [37], [38], but additional studies of other species are required to draw more general conclusions.
Here, we investigated the origins of the domesticated apple Malus domestica Borkh., one of the most emblematic and widespread fruit crops in temperate regions [35]. A form of apple corresponding to extant domestic apples appeared in the Near East around 4,000 years ago [39], at a time corresponding to the first recorded uses of grafting. The domesticated apple was then introduced into Europe and North Africa by the Greeks and Romans and subsequently spread worldwide [35]. While the ancestral progenitor has been clearly identified as being M. sieversii, the identity and relative contributions of other wild species present along the Silk route that have contributed to the genetic makeup of apple cultivars remain largely unknown. This is surprising given the potential importance of this knowledge for plant breeding and for our understanding of the process of domestication in fruit trees.
The wild Central Asian species M. sieversii (Ldb.) M. Roem has been identified as the main contributor to the M. domestica genepool based on similarities in fruit and tree morphology, and genetic data [40], [41], [42], [43]. The Tian Shan forests were identified as the geographic area in which the apple was first domesticated, on the basis of the considerable intraspecific morphological variability of wild apple populations in this region [44], [45]. Nucleotide variation for 23 DNA fragments even suggested that M. sieversii and M. domestica belonged to a single genepool (which would be called M. pumila Mill.), with phylogenetic networks showing an intermingling of individuals from the two taxa [43]. Some authors have also suggested possible contributions of additional wild species present along the Silk Route: M. baccata (L.) Borkh, which is native to Siberia, M. orientalis Uglitz., a Caucasian species present along western sections of the ancient trade routes, and M. sylvestris Mill. (European crabapple), a species native to Europe [46], [47], [48], [49]. These hypotheses were based on the history of human migration and trade, the lack of phylogenetic resolution between M. domestica and these four wild species [41], [42], genetic evidence of hybridization at a local scale between domesticated apple and M. sylvestris [40], and the recent finding of sequence haplotype sharing between M. sylvestris and M. domestica [50]. However, such secondary contributions remain a matter of debate, mostly due to the difficulty of distinguishing introgression from incomplete lineage sorting [43], [50], [51]. The three wild species occurring along the Silk Route all bear small, astringent, tart fruits. None of these species has the fruit quality of M. sieversii, but they may have contributed other valuable horticultural traits, such as later flowering, resistance to pests and diseases, capacity for longer storage or climate adaptation. The organoleptic properties of the fruits of these wild species may also have been selected during domestication, for the preparation of apple-based beverages, such as ciders [46], [52]. Cider apples are indeed smaller, bitter and more astringent than dessert apples and bear some similarity to M. sylvestris apples. There is also evidence to suggest that Neolithic and Bronze Age Europeans were already making use of M. sylvestris [39].
In this study, we used a comprehensive set of apple accessions sampled across Eurasia (839 accessions from China to Spain; Figure 1 and Figure S1; Table S1) and 26 microsatellite markers distributed evenly across the genome to investigate the following questions: 1) Is there evidence for population subdivision within and between the five taxa M. domestica, M. baccata, M. orientalis, M. sieversii and M. sylvestris? 2) How large is the contribution of wild species other than the main progenitor, M. sieversii, to the genome of M. domestica? 3) Does M. domestica have a genetic structure associated with its different possible uses (i.e., differences between cider and dessert apples)? 4) What consequences have domestication, subsequent crop improvement and vegetative propagation by grafting had for genetic variation in cultivated apples? Most of our samples of M. domestica corresponded to cultivars from Western Europe (Figure 1 and Figure S1), as almost all the cultivars available in modern collections (including American, Australasian cultivars) are of European ancestry and this region is therefore the most relevant area for the detection of possible secondary introgression from the European crabapple.
Our sampling scheme (Figure 1 and Figure S1), based on the collection of a single tree for each apple variety, was designed to avoid the sampling of clones. However, there may still be some clonality if some varieties differing by only a few mutations were propagated by grafting. We corrected for this potential clonality, using the clonal assignment procedures implemented in GENODIVE [53]. We found no pair of samples assigned to the same clonal lineage unless using a threshold of 22 pairwise differences between multilocus genotypes, indicating that our samples did not include any clonal genotypes (the threshold corresponds to the maximum genetic distance allowed between genotypes deemed to belong to the same clonal lineage).
Many apple cultivars, including modern cultivars in particular, share recent common ancestors, and siblings or clones of wild species can also be collected unintentionally in the field. Because these features could result in a spurious genetic structure due to the presence of closely related individuals in the dataset, we checked for the presence of groups of related individuals in our dataset between M. domestica cultivars and between the individuals of each wild species. The percentage of pairs with a pairwise relatedness (rxy) greater than 0.5 (i.e., full sibs) was: 0.4% in M. domestica (N = 168 pairs), 0.3% in M. sieversii (N = 79), 0.004% in M. orientalis (N = 20), and 0.7% in M. baccata (N = 40). For M. sylvestris, no individual pair with rxy>0.5 was identified. However, the distribution of pairwise relatedness rxy among M. domestica cultivars did not deviate significantly from a Gaussian distribution centred on 0 and with a low variance (Fisher's exact test, P≈1, standard deviation = 0.11, Figure S2). This suggests that closely related cultivars are unlikely to have biased subsequent analyses of population structure. We also checked for the limited effect of relatedness on our conclusions by performing all analyses of population subdivision on both the full dataset and a pruned dataset excluding related individuals (see below).
We tested the null hypothesis of random mating within each species by calculating FIS, which measures inbreeding. All five Malus species had relatively low values of FIS, although all were significantly different from zero (Table 1), suggesting that each species corresponded to an almost random mating unit. This is consistent with the self-incompatibility system of these species and indicates a lack of widespread groups of related individuals in M. domestica. Low FIS values at species level also indicate a lack of population structure within species. The higher values of FIS observed in M. baccata probably resulted from the occurrence of null alleles, as the microsatellite markers were developed in M. domestica, to which M. baccata is the most distantly related (Table 2). The lowest FIS value was that obtained for M. domestica, reflecting outcrossing between dissimilar parents in breeding programs, or that selection targeted higher levels of heterozygosity [54].
We used the ‘admixture model’ implemented in STRUCTURE 2.3 [55] to infer population structure and introgression. Analyses were run for population structure models assuming K = 1 to K = 8 distinct clusters (Figure 2). The ΔK statistic, designed to identify the most relevant number of clusters by determining the number of clusters beyond which there is no further increase in likelihood [56], was greatest for K = 3 (ΔK = 6249, Pr|ln L = −78590). However, the clusters identified at higher K values may also reveal a genuine and biologically relevant genetic structure, provided that they are well delimited [57]. The five Malus species were clearly assigned to different clusters for models assuming K≥6 clusters and for a minor clustering solution (“mode”) at K = 5 (Figure 2). The major mode (i.e., the clustering solution found in more than 60% of the simulation replicates) observed at K = 5 grouped together M. sylvestris and M. domestica genotypes. Increasing the number of clusters above K = 6 identified no additional well-delimited clusters corresponding to a subdivision of a previous cluster. Instead, it simply introduced heterogeneity into membership coefficients, indicating that the clustering of the five Malus species into separate genepools was the most relevant clustering solution. We checked that the presence of related pairs of cultivars in our dataset did not bias clustering results, by repeating the analysis on a pruned dataset (N = 489) excluding all related individuals in wild and cultivated species (i.e., excluding all pairs with rxy≥0.5). Similar results were obtained, with the same five distinct clusters identified as for the full dataset.
We estimated the genetic differentiation between the five Malus species by calculating pairwise FST (Table 2). All FST values were highly significant (P<0.001) and seemed to indicate a West to East differentiation gradient of M. domestica with the wild species. The highest level of differentiation was that between M. baccata and the other Malus species, and the lowest level of differentiation was that between M. domestica and the westernmost species, M. sylvestris (Table 2). Malus domestica was markedly more differentiated from its main progenitor M. sieversii (FST = 0.0639) than from the European M. sylvestris (FST = 0.006) and it was only slightly less differentiated from the Caucasian M. orientalis (FST = 0.049).
We first searched for footprints of a domestication bottleneck by comparing levels of microsatellite variation in M. domestica and wild species. There was no significant difference in genetic diversity (as measured by expected heterozygosity, HE) between M. domestica and M. baccata, M. orientalis or M. sieversii, but HE was significantly higher in M. sylvestris than in M. domestica (Table 1). Significant differences in allelic richness (Ar) were found between M. domestica and M. orientalis (Wilcoxon signed rank test, P = 0.03) or M. sylvestris (P<10−8), but not between M. domestica and either M. baccata (P = 0.9) or M. sieversii (P = 0.9) (Table 1).
We used the method implemented in the BOTTLENECK program [58], comparing the expected heterozygosity estimated from allele frequencies with that estimated from the number of alleles and the sample size, which should be identical for a neutral locus in a population at mutation-drift equilibrium. Inferences about historical changes in population size are based on the prediction that the expected heterozygosity estimated from allele frequencies decreases faster than that estimated under a given mutation model at mutation-drift equilibrium in populations that have experienced a recent reduction in size. BOTTLENECK analysis showed no significant deviation from mutation-drift equilibrium in any of the five species, under either stepwise or two-phase models of microsatellite evolution (one-tailed Wilcoxon signed rank test, P>0.95). We therefore detected no signal of a demographic bottleneck associated with the domestication of apples.
We used the admixture coefficients estimated by STRUCTURE to assess the recent contribution of the various wild species to the M. domestica genepool. STRUCTURE analyses of the full dataset showed some admixture among Malus species for the minor mode separating the five species at K = 5. Admixture coefficients were higher between M. domestica and M. sylvestris (α = 0.23) than between M. domestica and respectively M. sieversii (α = 0.06), M. orientalis (α = 0.034) and M. baccata (α = 0.032).
We further analysed the contribution of each wild species to the genome of M. domestica by running STRUCTURE separately on each pair of species including M. domestica (Figure 3; Table 3 and Table S2). Malus domestica genotypes with membership coefficients ≥0.20 in a wild species genepool were considered to display introgression. Using this somehow arbitrary cut-off value, STRUCTURE analyses revealed that 26% of M. domestica cultivars displayed introgression from the European crabapple, M. sylvestris (Table 3 and Table S2). By contrast, only 2%, 3% and 0.02% of the M. domestica genotypes displayed introgression from M. sieversii, M. orientalis and M. baccata, respectively (Table 3 and Table S2). The M. domestica cultivars displaying admixture with the M. sylvestris genepool were mostly Russian (e.g., “Antonovka”, “Antonovka kamenicka”, “Novosibirski Sweet”, “Yellow transparent”), French (e.g., “Blanche de St Anne”, “St Jean”, “Api” and “Michelin”) and English (e.g., “Worcester Pearmain” and “Fiesta”). The M9 dwarf apple cultivar (“Paradis jaune de Metz”, [59]) commonly used as a rootstock also appeared to display introgression from the European crabapple (proportion of ancestry in the M. domestica genepool: 0.28; Table S2). When French cultivars were removed from the dataset (N = 89) and pairwise STRUCTURE analyses were repeated for all species pairs including M. domestica, 18% of cultivars displayed introgression from M. sylvestris, including commercial cultivars such as Granny Smith, Michelin, Antonovka and Ajmi (Figure S3) with a mean membership coefficient of M. sylvestris into M. domestica genepool of 47%. Malus sylvestris thus appears to have made a significant contribution to the M. domestica genepool through recent introgression, building on the more ancient contribution (see below) of the Asian wild species M. sieversii. We also note that a few M. domestica individuals appeared to display introgression from several wild species (Table S2), and that M. baccata ornamental cultivars, such as M. baccata flexilis, M. baccata Hansen's and M. baccata gracilis, were partially or even mostly assigned (from 32% to >80%) to the M. domestica genepool (Table S3).
Previous studies [43], [50], [60] identified the Central Asian wild apple M. sieversii as the main progenitor of M. domestica on the basis of DNA sequences. Due to the large contribution by M. sylvestris detected in our dataset, corresponding mostly to Western European cultivars, M. domestica and M. sylvestris appeared to be the most closely related pair of species in our analyses of microsatellite markers. We investigated the more ancient contribution of M. sieversii to the M. domestica genepool, by reassessing the genetic differentiation between species in analyses restricted to “pure” individuals (i.e., assigned at ≥0.9 to their respective genepools) from both wild and cultivated species. All FST values were highly significant (P<0.001), but the ranking of FST values between M. domestica and the various wild species was affected: the highest differentiation was still observed between M. domestica and M. baccata (FST = 0.22), but the lowest differentiation was observed between M. domestica and M. sieversii (FST = 0.11). Regarding the differentiation between M. sylvestris and M. domestica, we observed the opposite of what was found with the full dataset: M. sylvestris appeared to be more strongly differentiated (FST = 0.14) from M. domestica than M. sieversii. Thus, by removing signals of recent introgression between cultivated and wild species we were able to confirm that M. sieversii was the initial progenitor of M. domestica.
The finding of a significant level of introgression from wild species into cultivated apple suggested that gene flow might also have occurred in the opposite direction. STRUCTURE analyses of pairs of species confirmed this hypothesis (Figure 3), revealing possible introgression of genetic material into M. sylvestris, M. baccata, M. orientalis and M. sieversii from M. domestica (mean proportions of ancestry in the M. domestica genepool of 0.12, 0.10, 0.03 and 0.23, respectively; Table 3). Considering genotypes with membership coefficients ≥0.9 in the M. domestica genepool as misclassified, we found a total of N = 31 misclassified wild Malus individuals. These results suggest gene flow from the domesticated apple genepool could significantly affect the genetic integrity of wild apple relatives, their future evolution and, possibly, their use as resources for crop improvement.
Model-based Bayesian clustering algorithms, such as that implemented in STRUCTURE, have a high level of power only for the detection of recent introgression events [55], [61], [62]. We therefore investigated the contributions of M. sylvestris and M. orientalis to the M. domestica genepool using approximate Bayesian computation (ABC) methods that offer a more historical perspective on gene flow [63]. We used a demographic model implementing admixture events [64].
We compared several admixture models to infer what species pairs underwent introgression events and to estimate introgression rates [64]. Malus baccata was not included in these analyses because of its high level of divergence from M. domestica. We assumed, as suggested by previous studies, that M. domestica derived originally from M. sieversii. The most complex model simulated sequential admixtures between M. domestica and all wild species. Other models sequentially removed introgression with each wild species, the order being based on FST values and admixture rates inferred by STRUCTURE. The compared models were the following: (i) the model a assumed that M. domestica was derived from M. sieversii and that the ancestral M. domestica population was involved in reciprocal introgression events with M. orientalis and M. sylvestris, and subsequently introgressed back into M. sieversii (Figure 4a), (ii) model b was similar to the model a, but without introgression events from M. domestica into wild species (Figure 4b), (iii) the model c included a single introgression event, from M. sylvestris into M. domestica (Figure 4c), and (iv) the model d simulated no admixture (Figure 4d). The number of parameters estimated in the model was limited by fixing the times of admixture with M. orientalis, M. sylvestris and M. sieversii at 600, 200 and 13 generations before the present, respectively. We used the following underlying hypotheses: (i) as the juvenile period of Malus lasts five to 10 years, we assumed a generation time of 7.5 years, (ii) admixture between ancestral M. domestica and M. orientalis in the Caucasus occurred approximately 4,500 years ago, shortly before the appearance of sweet apples in the Middle East (4,000 years ago), (iii) admixture between ancestral M. domestica and M. sylvestris in Europe occurred approximately 1,500 years ago, soon after the introduction of domesticated apples into Europe by the Greeks and Romans (iv) back-introgression into M. sieversii from M. domestica occurred approximately 100 years ago, when the cultivation of modern varieties reached Central Asia.
The relative posterior probabilities computed for each model provided strongest statistical support for model c, which assumed a single introgression event, from M. sylvestris into M. domestica (Table 4; posterior probability [p] = 0.67, 95% confidence interval: 0.63–0.72). Note that the model without admixture (model d) had the lowest relative posterior probability (Table 4). In analyses under alternative admixture models (models a and b), the posterior distributions were flat for introgression between M. domestica and M. orientalis and highly skewed towards low values for introgression into M. sylvestris and M. sieversii (not shown), which is consistent with statistical support being highest for model c.
Given that the model c was clearly favoured, parameter estimates are shown below only for this model (Table 5; prior distributions in Table S4). The contribution of M. sylvestris to the M. domestica genepool was estimated at about 61% (95% credibility interval [95% CI]: 50–68%). We obtained estimates of effective population sizes of 3,520 (95% CI: 2,090–5,680) for M. domestica, 13,200 (95% CI: 6,920–19,300) for M. sieversii, 34,600 (95% CI: 15,100–48,000) for M. sylvestris, and 28,300 (95% CI: 11,700–64,000) for M. orientalis. Using a generation time of 7.5 years, the divergence between M. domestica and M. sieversii (T3) was estimated to have occurred 17,700 years ago (95% CI: 6,225–25,200), which is earlier than previously thought, but we note that the credibility interval is quite large. We estimated that M. sylvestris and M. sieversii diverged about 83,250 years ago (T1, 95% CI: 40,575–334,500), with M. orientalis and M. sieversii diverging about 20,775 years ago (T2, 95% CI: 9,900–47,775).
The results above were obtained using the full dataset. We checked the validity of our inferences by conducting analyses on the dataset without admixed and misclassified individuals and using different times of admixture, by assessing the goodness-of-fit of models to data, and by checking that sufficient power was achieved to discriminate among competing models (Text S1; Tables S5, S6, S7). Overall, ABC analyses all provided clear support for a model with contribution of the European crabapple into the domesticates, although the estimated value of the actual contribution of M. sylvestris is probably overestimated here, and should therefore be treated with caution. Indeed, the simulation of a single introgression event hundreds of years ago most likely demanded higher rates of introgression to account for the actual genetic contribution of M. sylvestris into M. domestica than would be needed under continuous gene flow over a long period.
As cider cultivars produce apples that are smaller, more bitter and astringent than dessert cultivars, we expected to observe genetic differentiation between these two groups of cultivars and a closer genetic proximity of cider cultivars to M. sylvestris [35], [65]. Neither hypothesis was supported by our data. The classification of apples into “dessert” and “cider” varieties as prior information for STRUCTURE (Locprior model) revealed a very weak tendency of cider and dessert cultivars to be assigned to different clusters at K = 2 (Figure 5), but increasing K did not further result in clearer differentiation between the two types of cultivars. At K = 2, M. domestica cider genotypes had a mean membership of 94.7%, and M. domestica dessert genotypes had a mean membership of 52.5%. However, STRUCTURE analyses without this prior information gave essentially the same clustering patterns at K = 2 (G′ = 0.95 similarity to analyses using classification to assist clustering). The weak differentiation between cider and dessert cultivars (FST = 0.02) and their high level of admixture in STRUCTURE analyses (Figure 5) indicated a shallow subdivision of the M. domestica genepool. Analyses on a pruned dataset from which closely related individuals had been removed (i.e., pairs of genotypes with rxy≥0.5; N = 172) revealed the same pattern, confirming that the presence of related cultivars in the dataset did not bias clustering analyses. STRUCTURE was also run on a dataset including all M. sylvestris genotypes, to test the hypothesis that cider cultivars would display a higher level of introgression from the European crabapple. However, the opposite pattern was observed: the proportion of genotypes displaying introgression from M. sylvestris was actually significantly higher in dessert than in cider cultivars (36.4% and 15.5% respectively, χ2 = 16.9, P = 4×10−5). Finally, little genetic differentiation was observed between groups of cultivars of different geographic origins (95% CI: −0.8–0.6, Table S8).
The apple is so deeply rooted in the culture of human populations from temperate regions that it is often not recognized as an exotic plant of unclear origin. We show here that the evolution of the domesticated apple involved more than one geographically restricted wild species. The domesticated apple did not arise from a single event over a short period of time, but from evolution extending over thousands of years. The genepool of the current domesticated apple varieties has been enriched by the contribution of at least two wild species. Malus species have a self-incompatibility system; apple domestication and traditional variety improvement have therefore been based mostly on the selection of the best phenotypes grown from open-pollinated seeds. This breeding strategy has probably favoured the incorporation of genetic material from multiple wild sources and the maintenance of high levels of genetic variation in domesticated apples, despite the extensive use of large-scale vegetative propagation of superior individuals by grafting. Our results are consistent with those reported for the few other woody perennials studied to date, such as grape [37], red mombin [26] and olive trees [36], and support the view that domestication in long-lived plants differs in many respects from the scenarios described for seed-propagated annuals.
Malus sieversii was previously identified as the main contributor to the M. domestica genome on the basis of morphological and sequence data [41], [43]. The flanks of the Tian Shan mountains have been identified as a likely initial site of domestication, based on the high morphological variability of the wild apples growing in this region, and their similarity to sweet dessert apples [44], [45]. We show here, using a set of rapidly evolving genetic markers distributed throughout the genome and a large sampling, that M. domestica now forms a distinct, random mating group, surprisingly well separated from M. sieversii, with no difference in levels of genetic variation between the domesticate and its wild progenitor. This contrasts with the pattern previously reported, based on a twenty three-gene phylogenetic network [43], where domesticated varieties of apple appeared nested within M. sieversii. After the removal of individuals showing signs of recent admixture, M. sieversii and M. domestica nevertheless appeared to be the pair of species most closely related genetically, confirming their progenitor-descendant relationship.
Apple breeding methods (grafting and “chance seedling” selection), life-history traits specific to trees and/or the genetic architecture of selected traits have likely played a role in the conservation of levels of genetic diversity in cultivated apples similar to those in wild apples. Some factors, such as “chance seedling” selection [66], may even have increased genetic diversity, by favouring outcrossing events among domesticates and introgression from wild species [39]. The low inbreeding coefficients inferred in domesticated apples and the low level of differentiation between cultivated and wild apple populations [40], [54], [67], [68] indicate a high frequency of crosses between individuals of M. domestica, M. sieversii and other wild relatives hailing from diverse geographic origins. Such a high level of gene flow has likely contributed to maintenance of a high level of genetic diversity in domesticated apples.
The grafting technique, which was probably developed around 3,000 years ago, has made it possible to propagate superior individuals clonally. The spread of grafting, together with the lengthy juvenile phase (5–10 years) and the long lifespan of apples, may have imposed strong limits on the intensity of the domestication bottleneck thereby limiting the loss of genetic diversity [27], [28], [31]. By decreasing the number of generations since domestication, these factors have probably also helped to restrict the differentiation between domesticates and wild relatives. In theory, grafting may have limited the size of the apple germplasm dispersed early on to a few very popular genotypes, thereby provoking a sudden shrink in effective population size and a loss of diversity. However, we found no evidence that the clonal propagation of apples resulted in a long-lasting decrease in population size or clonal population structure. We can speculate that this may be due to a combination of various factors such as: gene flow with wild species, small-scale propagation (many farmers producing a few grafts each), a large variation in preferences for taste and other quality characteristics between farmers and cultures, large differences in growth conditions leading to the adoption of different sets of genotypes in different regions or the typical behaviour of hobby breeders, who tend to spot particular differences and multiply them. Similarly, for grape, there are huge numbers of old varieties and as much genetic variation in cultivated varieties as in wild-relative progenitors [37].
There has been a long-running debate concerning the possible contribution of other wild species present along the Silk Route to the genetic makeup of M. domestica [40], [46], [47], [65], [69]. Our results clearly show that interspecific hybridization has been a potent force in the evolution of domesticated apple varieties. Apple thus provides a rare example of the evolution of a domesticated crop over a long period of time and involving at least two wild species (see also the cases of olive tree and avocado [24], [26], [37], [70]). A recent study argued that introgression from M. sylvestris into the M. domestica genepool was the most parsimonious explanations for shared gene sequence polymorphisms between the two species [50]. Using an unprecedentedly large dataset, more numerous and more rapidly evolving markers and a combination of inferential methods, we provide a comprehensive view of the history of domestication in apple. We confirm that M. sieversii was the initial progenitor and show that the wild European crabapple M. sylvestris has been a major secondary contributor to the diversity of apples, resulting in current varieties of M. domestica being more closely related to M. sylvestris than to their central Asian progenitor. This situation is reminiscent of that for maize, in which the cultivated crop Zea mays is genetically more closely related to current-day highland landraces than to lowland Z. mays ssp. parviglumis from which the crop was domesticated [71]. This pattern has been attributed to large-scale gene flow from a secondary source, a second subspecies of teosinte, Z. mays ssp. mexicana, into highland maize populations [71].
The usefulness of wild relatives for improving elite cultivated crop genepools has long been recognised and the exploitation of wild resources is now considered a strategic priority in breeding and conservation programs for most crops [11], [12], [44]. Domesticated apples are unusual in that the contribution of wild relatives probably occurred early and unintentionally in the domestication process, preceding even the use of controlled crosses. The use of genetic markers with lower mutation rates than our set of microsatellites might also make it possible to investigate the contribution of more phylogenetically distant apple species growing in areas away from the Silk Route to the diversification of modern apple cultivars.
The Romans introduced sweet apples into Europe at a time at which the Europeans were undoubtedly already making cider from the tannin-rich fruits of the native M. sylvestris [35], [72]. Cider is not typical of Asia [35], but it was widespread in Europe by the time of Charlemagne (9th century, [73]). Large numbers of apple trees were planted for cider production in France and Spain from the 10th century onwards [48], [52]. The very high degree of stringency of cider apples (often to the extent that they are inedible) led to the suggestion that cider cultivars arose from hybridization between M. sylvestris and sweet apples [35], [46], [65]. We show here that the genetic structure within the cultivated apple genepool is very weak, with poor differentiation between cider and dessert apples. Cider cultivars thus appear to be no more closely genetically related to M. sylvestris than dessert cultivars. As wild Asian apples are known to cover the full range of tastes [44], [46], it is possible that fruits with the specific characteristics required for cider production were in fact initially selected in Central Asia and subsequently brought into Europe. There is a long-standing tradition of cider production in some parts of Turkey [35], for instance, which is potentially consistent with an Eastern origin of cider cultivars. However, the low level of genetic differentiation between dessert and cider apples indicates that, even if different types of apples were domesticated in Asia and brought to Europe, they have not diverged into independent genepools.
This study settles a long-running debate by confirming that 1) M. domestica was initially domesticated from M. sieversii, and 2) M. domestica subsequently received a significant genetic contribution from M. sylvestris, much larger than previously suspected [35], at least in Western Europe, where originated most of our samples and most cultivar diversity. The higher level of introgression of the European crabapple into the domesticated apple in this study than in previous studies [43], [50], [51] may be attributed to the use of a larger and more representative set of M. domestica genotypes coupled with the genotyping of numerous and rapidly evolving markers known to trace back more recent events.
Our inferences also have important implications for breeding programs and for the conservation of wild species of apple. The major contribution of the various wild species to the M. domestica genepool highlights the need to invest efforts into the conservation of these species, which may contain unused genetic resources that could further improve the domesticated apple germplasm [74], such as disease resistance genes or genes encoding specific organoleptic features.
Leaf material was retrieved from the collections of various institutes (INRA Angers, France; USDA - ARS, Plant Genetic Resources Unit, Geneva, NY; ILVO Melle, Belgium) and from a private apple germplasm repository in Brittany for M. domestica (N = 368, Figure S1 including only diploid cultivars N = 299) and from forests for the four wild species (Figure 1; Table S1). Malus sieversii (N = 168) material was collected from 2007 to 2010 in the Chinese Xinjiang province (N = 26), Kyrgyzstan (N = 5), Uzbekistan (N = 1), Tajikistan (N = 1) and Kazakhstan (N = 114). Malus orientalis (N = 215) was sampled in 2009 in Armenia (N = 203), Turkey (N = 5) and Russia (N = 5). Malus sylvestris (N = 40) samples were obtained from 15 European countries. Malus baccata (N = 48) was sampled in 2010 in Russia. The origins of M. domestica cultivars were: France (N = 266), Great Britain (N = 12), USA (N = 12), Russia (N = 7), the Netherlands (N = 6), Australia (N = 4), Belgium (N = 4), Germany (N = 4), Japan (N = 3), Ukraine (N = 3), Tunisia (N = 2), Switzerland (N = 2), Spain (N = 2), New Zealand (N = 2), Israel (N = 1), Ireland (N = 1), Canada (N = 1), Armenia (N = 2) and unknown/debated (N = 34). Genomic DNA was extracted with the Nucleo Spin plant DNA extraction kit II (Macherey & Nagel, Düren, Germany) according to the manufacturer's instructions.
Microsatellites were amplified by multiplex PCR, with the Multiplex PCR Kit (QIAGEN, Inc.). We used 26 microsatellites spread across the 17 chromosomes (one to three microsatellites per chromosome), in 10 different multiplexes previously optimised on a large set of genetically related progenies of M. domestica [75]. The four multiplexes (MP01, MP02, MP03, MP04; Table S9; Lasserre P. unpublished data) were performed in a final reaction volume of 15 µl (7.5 µl of QIAGEN Multiplex Master Mix, 10–20 µM of each primer, with the forward primer labelled with a fluorescent dye and 10 ng of template DNA). We used a touch-down PCR program (initial annealing temperature of 60°C, decreasing by 1°C per cycle down to 55°C). Six other multiplex reactions (Hi6, Hi4ab, Hi5-10, Hi13a, Hi13b, Hi4b) were performed using previously described protocols [75]. Genotyping was performed on an ABI PRISM X3730XL, with 2 µl of GS500LIZ size standard (Applied Biosystems). Alleles were scored with GENEMAPPER 4.0 software (Applied Biosystems). We retained only multilocus genotypes presenting less than 30% missing data.
We checked the suitability of the markers for population genetic analyses. None of the 26 microsatellite markers deviated significantly from a neutral equilibrium model, as shown by the non significant P-values obtained in Ewen-Watterson tests [76], and no pair of markers was found to be in significant linkage disequilibrium in any of the species [77], [78]. The markers could therefore be considered unlinked and neutral.
Apple cultivars may be polyploid [79]. We therefore first checked for the presence of polyploidy individuals of M. domestica within our dataset. Individuals presenting multiple peaks on electrophoregrams were first re-extracted to eliminate contamination as a possible source of apparent polyploidy. We then checked whether they had been reported to be polyploidy in previous studies [79]. After completion of this checking procedure, we removed 69 polyploids (of the 368 samples) from subsequent analyses. We tested for the occurrence of null alleles at each locus with MICROCHECKER 2.2.3 software [80]. Allelic richness and private allele frequencies were calculated with ADZE software [81], for a sample size of 22. Heterozygosity (expected (HE) and observed (HO)), Weir & Cockerham F-statistics, deviation from Hardy-Weinberg equilibrium and genotypic linkage disequilibrium were estimated with GENEPOP 4.0 [77], [78]. The significance of differences between FST values was assessed in exact tests carried out with GENEPOP 4.0 [77], [78]. Individuals were assigned to clonal lineages with GENODIVE [53]. We estimated relatedness between pairs of cultivars and between pairs of individuals within each species, by calculating the rxy of Ritland and Lynch [82] with RE-RAT online software [83]. We tested whether the distributions of rxy deviated significantly from a Gaussian distribution with a mean of zero and a standard deviation equal to the observed standard deviation, by comparing observed and simulated distributions in Fisher's exact test (R Development Core Team, URL http://www.R-project.org).
We tested for the occurrence of a bottleneck during apple domestication with the method implemented in BOTTLENECK [58], [84]. The tests were performed under the stepwise-mutation model (SMM) and under a two-phase model (TPM) allowing for 30% multistep changes. We used Wilcoxon signed rank tests to determine whether a population had a significant number of loci with excess genetic diversity.
We used the individual-based Bayesian clustering method implemented in STRUCTURE 2.3.3 [55], [85], [86] to investigate species delimitation, intraspecific population structure and admixture. This method is based on Markov Chain Monte Carlo (MCMC) simulations and is used to infer the proportion of ancestry of genotypes in K distinct predefined clusters. The algorithm attempts to minimize deviations from Hardy–Weinberg and linkage equilibrium within clusters. Analyses were carried out without the use of prior information, except for analyses of population subdivision within the M. domestica genepool for which the “cider”/“dessert” classification of cultivars was used as prior information to assist clustering. K ranged from 1 to 8 for analyses of the five-species dataset and the M. domestica dataset, and was fixed at K = 2 for analyses of pairs of species including M. domestica and each of the wild species. Ten independent runs were carried out for each K and we used 500,000 MCMC iterations after a burn-in of 50,000 steps. We used CLUMPP v1.1.2 (Greedy algorithm) [87] to look for distinct modes among the 10 replicated runs of each K.
STRUCTURE analyses were run for the full dataset (N = 839) and for two pruned datasets excluding non-pure individuals (i.e., genotypes with <0.9 membership of their species' genepool) and related individuals (rxy≥0.5).
We used the DIYABC program [88] to compare different admixture models and infer historical parameters. We simulated microsatellite datasets for 14 loci (Ch01h01, Ch01h10, Ch02c06, Ch02d08, Ch05f06, Ch01f02, Hi02c07, Ch02c09, Ch03d07, Ch04c07, Ch02b03b, MS06g03, Ch04e03, Ch02g01) previously reported to be of the perfect repeat type [89], [90], [91]. In total, we generated 5×105 simulated datasets for each model.
A generalized stepwise model (GSM) was used as the mutational model. The model had two parameters: the mean mutation rate (μ) and the mean parameter (P) of the geometric distribution used to model the length of mutation events (in numbers of repeats). As no experimental estimate of microsatellite mutation rate is available for Malus, the mean mutation rate was drawn from a uniform distribution by extreme values of 10−4 and 10−3, and the mutation rate of each locus was drawn independently from a Gamma distribution (mean = μ; shape = 2). The parameter P ranged from 0.1 to 0.3. Each locus L had a possible range of 40 contiguous allelic states (44 for CH02C06, 42 for CH04E03) and was characterized by individual values for mutation rate (μL) and the parameter of the geometric distribution (PL); μL and PL were drawn from Gamma distributions with the following parameter sets: mean = μ, shape = 2, range = 5×10−5–5×10−2 for μL, and mean = P, shape = 2, range = 0.01–0.9 for PL. As not all allele lengths were multiples of motif length, we also included single-nucleotide insertion-deletion mutations in the model, with a mean mutation rate (μSNI) and locus-specific rates drawn from a Gamma distribution (mean = μSNI; shape = 2). The summary statistics used were: mean number of alleles per locus, mean genetic diversity [92], genetic differentiation between pairwise groups (FST; [93]), genetic distances (δμ)2 [94].
We used a polychotomous logistic regression procedure [95] to estimate the relative posterior probability of each model, based on the 1% of simulated data sets closest to the observed data. Confidence intervals for the posterior probabilities were computed using the limiting distribution of the maximum likelihood estimators [64]. Once the most likely model was identified, we used a local linear regression to estimate the posterior distributions of parameters under this model [96]. The 1% simulated datasets most closely resembling the observed data were used for the regression, after the application of a logit transformation to parameter values.
|
10.1371/journal.pcbi.1006533 | A Pareto approach to resolve the conflict between information gain and experimental costs: Multiple-criteria design of carbon labeling experiments | Science revolves around the best way of conducting an experiment to obtain insightful results. Experiments with maximal information content can be found by computational experimental design (ED) strategies that identify optimal conditions under which to perform the experiment. Several criteria have been proposed to measure the information content, each emphasizing different aspects of the design goal, i.e., reduction of uncertainty. Where experiments are complex or expensive, second sight is at the budget governing the achievable amount of information. In this context, the design objectives cost and information gain are often incommensurable, though dependent. By casting the ED task into a multiple-criteria optimization problem, a set of trade-off designs is derived that approximates the Pareto-frontier which is instrumental for exploring preferable designs. In this work, we present a computational methodology for multiple-criteria ED of information-rich experiments that accounts for virtually any set of design criteria. The methodology is implemented for the case of 13C metabolic flux analysis (MFA), which is arguably the most expensive type among the ‘omics’ technologies, featuring dozens of design parameters (tracer composition, analytical platform, measurement selection etc.). Supported by an innovative visualization scheme, we demonstrate with two realistic showcases that the use of multiple criteria reveals deep insights into the conflicting interplay between information carriers and cost factors that are not amendable to single-objective ED. For instance, tandem mass spectrometry turns out as best-in-class with respect to information gain, while it delivers this information quality cheaper than the other, routinely applied analytical technologies. Therewith, our Pareto approach to ED offers the investigator great flexibilities in the conception phase of a study to balance costs and benefits.
| Designing experiments is obligatory in the biosciences to valorize their scientific outcome. When the experiments are expensive, unfortunately, in practice often the costs emerge to be showstoppers. In this situation the question arises: How to get the most out of the experiment for your invest in terms of time and money? We approach this question by formulating the design task as a multiple-criteria optimization problem. Its solution produces a set of Pareto-optimal design proposals that feature the trade-off between information gain, as measured by different metrics, and the costs. Then, exploration of the design proposals allows us to make the best decision on information-economic experiments under given circumstances. Implemented in the field of isotope-based metabolic flux analysis, practical application of the Pareto approach provides detailed insight into the tight interplay of plenty of information carriers and cost factors. Supported by an innovative tailored visual representation scheme, the investigator is enabled to explore the options before conducting the experiment. With a practical showcase at hand, our computational study highlights the benefits of incorporating multiple information criteria apart from the costs, balancing the shortcomings of conventional single-objective experimental design strategies.
| The successful design of tailor-made cell factories in the biotechnological and pharmaceutical industries needs firm understanding of the cellular functions and their underlying molecular mechanisms [1–3]. The key to get the most insight from an experiment is a careful experimental design (ED), precisely, the selection of experimental settings and measurements that harvest a maximum of information about the quantities of interest. In this context, there is growing interest in computer-aided modeling to guide the experimental choices [4–8]. Existing design techniques can be broadly divided into statistical approaches that strive to maximize the statistical confidence of inferring model parameters and information-theoretic approaches identifying informative designs to tackle the principal identifiability problem [9–11]. These techniques have been applied in various studies to deduce information-optimal settings to tackle the following questions:
For quantify the information gain, several optimality criteria (or precision scores) have been suggested, all approximating the average statistical confidence of parameter estimates [12,13]. Typically, the information criterion to be used for the ED is decided ad hoc, since the most “suited” one is not known in advance. Favoring a single criterion in the planning phase, however, may well lead to improvements in that criterion at the expense of a decline of others, taking the risk to under-explore the design space and, eventually, deriving misleading design decisions [14]. To remedy this limitation, several information criteria could be simultaneously taken into account.
Although information remains a key criterion for science, it comes at a cost. In practice, resource-oriented considerations shape ED strategies, especially when experiments are extensive, time-consuming and labor-intense. For example when organisms exhibit slow growth rates, complicated experimental and sample preparation protocols are involved, or a large number of data has to be analyzed semi-manually. Consequently, from an economic point of view questions on the design of experiments are:
These questions motivate to explore the experimental settings to select those that are informative and offer this information in a cost-efficient manner. Clearly, such information-economic considerations need cost models that contemplate all major factors (equipment, replicates, analysis time, etc.) and relate them to the information carriers. For instance, increasing the number of samples positively affects the information gain while, at the same time, it raises the costs, implying that here the goals “maximize information” and “minimize cost” are incommensurable.
That said, finding an informative, yet cost-efficient experimental setting out of the space of alternate designs is a nontrivial task: First, the space of options may be extensively large and secondly, several related, but potentially conflicting design objectives need to be optimized simultaneously. Here, a common solution concept is to optimize a weighted sum of the single criteria [6,15]. However, in real-world scenarios the objectives are hardly expressible in the same “currency” and appropriate weights to translate between them are not known before the experiment. Consequently, in scenarios where the ability to explore the whole space of design alternatives should be maintained, a fixed-weight solution cannot be utilized [16]. To overcome the limitations of weighted-sum single-objective approaches, the ED task can be casted into a multi-objective optimization (MOO) formulation [17]. Multi-objective (MO) ED comes with an important conceptual difference, compared to single-objective ED: When objectives are conflicting, instead of one specific solution, a whole set of—in terms of the objectives—equally good, compromise EDs is obtained where none of the designs is better than the others in terms of all criteria. These compromise EDs, denoted Pareto-optimal EDs, determine the Pareto front in the objective space (Fig 1). When the objectives are not in competition, a characteristic that cannot be known for real-world problems a priori, the MO-ED task degenerates to an ordinary ED problem.
The trade-off decision on the experiment is then made after examining the Pareto front and inspecting the related Pareto-optimal designs where (expert or newly available) information or preferences can be considered in addition. However, to keep track of more than a few relations is not only intrinsically challenging, it also calls for domain-specific solutions to interrogate the high-dimensional Pareto-optimal results and to support exploration and interpretation processes.
We present a universal computational methodology for the design of informative, yet cost-effective experiments. Our approach simultaneously optimizes many, potentially contradicting information and cost metrics rather than a single one, therewith generalizing traditional ED frameworks basing on the optimization of a single information criterion. To provide a visual means for result exploration of Pareto-optimal EDs in potentially high-dimensional design and objective spaces, we suggest a flexible solution using chord diagrams.
To exemplify our information-economic Pareto approach, the MO-ED framework is implemented for 13C metabolic flux analysis (13C MFA), which provides a computationally challenging test bed owing to its enormous design space and diverse cost factors. Equipped with the computational tools, the questions raised above were addressed by a comprehensive investigation featuring the fungus Penicillium chrysogenum. In particular, two different scenarios were studied. First, all analytical platforms commonly applied for 13C MFA were profiled with respect to their information-economic characteristics, using a single information criterion. The study revealed that the specific measurement information delivered by tandem mass spectrometry (MS/MS) cannot only increase flux information, but also enabled cost savings by the choice of cheaper tracers, emphasizing the potential of our approach. In the second scenario we investigated whether including more than one information criterion could provide a benefit for the decision process. Indeed, for the P. chrysogenum showcase a variety of additional Pareto-optimal designs were offered, unlocking informed decision making. In particular for, but not limited to, the domain of 13C MFA our findings show that the use of several criteria balances shortcomings of conventional ED strategies and offers additional flexibilities for the experimenter, thus providing a methodology of direct practical relevance.
Planning cost-efficient, informative experiments requires finding the “best” experimental-analytical trade-offs that, on the one hand, maximize the information gain, possibly in view of different information facets, while, on the other hand, keep the associated costs to a minimum. Consequently, two formal ingredients are needed:
Employing these criteria in the selection procedure of the ED formally amounts to a multi-objective optimization (MOO) problem:
maxα∈ΩΦ(α,θ)s.t.g(α,θ)≥0h(α,θ)=0l≤α≤u
(1)
where the objective vector Φ is composed of a set of information and (negated) cost criteria. The objective vector is a function of the design variables α, selected from the space Ω of feasible designs. Remaining design parameters, which are constant, are collected in the vector θ. Furthermore, the bounded design variables may be subject to inequality and equality constraints.
Solving the MOO problem (1) means to find the set of all trade-off design solutions α* that minimize the objectives in Φ without being dominated by another solution [18]. Here, a specific design α1 dominates another one α2, if (and only if) α1 is at least as good as α2 in all objectives and better with respect to at least one, formally expressed by Φi (α1) ≤ Φi (α2), ∀i and ∃j: Φj (α1) < Φj (α2) (gray shaded area in Fig 1). The set of all non-dominated solutions is referred to as Pareto-optimal design set, and the corresponding achievable objective values are called Pareto front.
Clearly, the concrete formulation of the MOO problem depends on the particular application case, namely the underlying system model and the peculiarities of the experimental settings. In this work, we selected a use-case from the domain of 13C metabolic flux analysis (MFA), which is arguably the most expensive type of ‘omics’ technology, featuring dozens of design variables. Before introducing the information and intricate cost models as well as the analysis of Pareto-designs in high dimensions, the essential background to the application field is provided, in particular the formulation of the system model.
Intracellular reaction rates (fluxes) describe the trafficking of metabolites which emerges as the final outcome of all catalytic and regulatory processes acting within living cells [19]. Here, the reactions within a biochemical network are characterized by a pair of flux values, net and exchange fluxes [20], to express the respective proportions of material transported between the reaction’s educts and products. At steady-state, the in- and outflows of each intermediate metabolite are assumed to be constant and mass balanced, yielding the stoichiometric equation system for the flux vector v:
S⋅v=b,Cieq⋅v≤cieq
(2)
with the stoichiometric matrix S and the vector b containing the extracellular rates (substrate uptake, product formation or effluxes leading to biomass accumulation), accessible through extracellular concentration profiles and biomass quantification. In addition, the fluxes may be constrained in their allowable value range owing to physiological knowledge.
Since metabolic networks contain parallel paths and cycles, fluxes are not uniquely determined by Eq (2), at least not without additional assumptions. The indeterminacy implies that the flux vector v can be parametrized through a certain (non-unique) sub-set of fluxes, the so called free fluxes vfree [20]. The dimensionality of the vector vfree, i.e., dim(v) − rank(S), is referred to as degrees of freedom (DoF). To resolve the DoFs, carbon labeling experiments (CLEs) are conducted. In a CLE, isotopically labeled carbon sources, like [1-13C] glucose enriched with a 13C isotope at the first position of the carbon backbone, are fed to the cells. The labeled substrate is taken up by the cells and distributed through the metabolic pathways to all intracellular metabolites, where it gives rise to characteristic labeling enrichment patterns. Thus, the labeling patterns are the convoluted result of the routes, the 13C labeled substrate takes, as well as the underlying metabolic fluxes. In isotopic steady-state 13C MFA, as used in this work, intracellular free fluxes are inferred from the equilibrated labeling patterns and external rate measurements by means of a computational flux fitting procedure that minimizes the least-squares error between observed measurements and those that are simulated by a computational network model [21].
For the model, carbon atom transitions have to be specified for each reaction step describing the fate of each carbon atom from the reactions’ educt to its corresponding product. Mass balancing of the intracellular isotopic forms then yields a high-dimensional nonlinear algebraic equation system that relates the steady-state labeling state x, the administered labeled tracer mixture xinp, and the free fluxes vfree [20]. Given vfree and xinp, the vector of steady-state labeling states x (represented as isotopomers, cumomers, EMUs, or similar [20,22,23]) is uniquely determined by [24]:
x=x(vfree,xinp)
(3)
Note that CLEs that only differ in the tracer mixture are covered by the same formalism through duplication of the network model and equating the free fluxes.
The full system-wide labeling state x is not accessible by any current measurement technology. What can be observed are linear combinations of (relative) abundances for some of the intracellular metabolites, such as mass isotopomer distributions or positional enrichments. Fig 2 shows characteristic sets of observations, henceforth denoted measurement groups, for the analytical platforms employed in the field of 13C MFA.
All measurement groups available for an analytical device are organized in the measurement matrix Mmeasdev that, owing to Eq (3), allows to simulate the measurement vector η:
η=Mmeasdev⋅x(vfree,xinp)
(4)
which mimics the real measurements up to normalization to percentage scale [25]. Examples for measurement matrices are given in S1 Text. Real measurements are unavoidably affected by noise. In the context of 13C MFA, measurement noise is assumed to be independent, unbiased, additive, and normally distributed with expectation 0 and standard deviation σmeasdev, as represented by the measurement covariance matrix Σmeasdev [26]:
Σmeasdev=diag(σmeasdev)
(5)
Since in the CLE’s planning phase real measurements are absent, from which measurement standard deviations σmeasdev can be derived, measurement error models need to be formulated, relating the measurements with their associated errors. For labeling measurements empirical rule-of-thumb approximations of the measurement precision have been derived for specific analytical setups. For instance, Crown et al. propose a precision of 0.4 mol% for their GC-MS setup targeting proteinogenic amino acids [27]. In general, labeling errors depend on the measurement technique, the instrument, the analytic protocols, they can vary between organisms, analytes and the degree of label incorporation [28]. To arrive at realistic error approximations that allow for a fair comparison of the analytical platforms, measurements and their standard deviations were collected from published studies featuring different organisms, platforms and various labeling contents. In total, more than 900 data points for six analytical platforms, namely GC-MS, LC-MS, LC-MS/MS, 13C-NMR, 1H-NMR, and GC-C-IRMS were extracted (S1 Text). For all analytical platforms, similar to the approach by Dauner et al. for 13C-NMR [29], a regression line was fitted to the respective data set, yielding device-specific linear measurement error models. These analytics-related error models provide empirical standard deviations σmeasdev for any given measured vector η:
σmeasdev(nrep,measdev)=a(nrep,measdev)⋅(b1dev⋅η+b2dev)
(6)
where b1dev,b2dev are the device-specific regression coefficients (S1 Text). Generally, by increasing the number of repetitions nrep,measdev (i.e., technical replicates), the error estimates are believed to become more reliable. This is accounted for in the error models (6) by a scaling factor (a) which tends to 1 for the case of many repetitions (see S2 Appendix for details).
Several statistical approaches have been developed to predict the approximate amount of information to be derived from the planned CLE or CLE series. When some pre-knowledge on the expected flux map v^free is available (which we assume in this work), a widely adopted local information measure is the Fisher information matrix (FIM) [9,13,26]:
FIM=(∂η∂vfree|v^free)T⋅Σmeasdev⋅∂η∂vfree|v^free
(7)
whose inversion yields the flux covariance matrix:
Cov(v^free,α)=FIM−1
(8)
which depends on the design point (v^free) and the design parameters (α). As a precondition for stable numeric calculation of the flux covariance matrix, the FIM needs to fulfill two conditions [30]: First, its minimal singular value λmin (FIM) needs to be larger than a threshold:
λmin(FIM)>τ1>0
(9)
and secondly, its condition number has to be bounded:
cond(FIM)<τ2<∞
(10)
The fulfillment of the conditions (9) and (10) implies that the standard deviations of the free fluxes—as represented by the main diagonal of the covariance matrix—remain bounded and, thus, the flux vector is said to be statistically identifiable. First, it should be remarked, that this is a slightly stronger variant of practical identifiability as defined by Raue et al. in [31] and secondly, statistically identifiable fluxes are per se structurally identifiable [32]. If either one of the conditions (9) and (10) is violated, fluxes causing the violation have to be excluded from the FIM. Eventually, this leads to models that vary in terms of their DoFs, a fact which needs careful treatment when comparing different experimental setups with respect to their information content.
For quantifying the information content of a CLE several information quality criteria have been proposed that aggregate the covariance matrix to a single number [9,12,13]. The most prominent ones are the determinant (D), the average-variance (A), and eigenvalue (E) criteria. Ultimately, all these criteria provide a means for the shape of the confidence ellipsoid in the vicinity of a given design point (v^free in our case), each emphasizing particular geometrical aspects [12] (Fig 3).
For example, the D-criterion strives to minimize the volume of the confidence ellipsoid (or the geometric mean of the flux confidence intervals):
ΦD,p=det(Cov)2⋅p
(11)
with p the dimension of Cov (with arguments omitted for brevity) while the A-criterion aims to minimize the diagonal of the smallest bounding box that contains the confidence ellipsoid (or the arithmetic mean of the flux confidence intervals):
ΦA,p=trace(Cov)/p
(12)
Hence, the A-criterion is expected to provide designs that are more robust against flux correlations than those based on the D-criterion. Notice that the explicit consideration of the dimension p of the covariance matrix in the formulation of criteria (11) and (12) intends to make the criterion values comparable for models differing in the number of free fluxes. In contrast, the E-criterion:
ΦE=λmax(Cov)/λmin(Cov)
(13)
constitutes a dimension independent measure that strives to improve worst case designs by preventing the Fisher matrix from becoming singular. Besides these quantitative information measures, an obvious quality criterion is the number of free fluxes that can be statistically identified by the ED setting, expressed by:
ΦDoF=dim(Cov)
(14)
With these information measures at hand, the information gain of a 13C MFA study can be influenced by the targeted selection of the input mixture compositions (xinp), the measured groups observable by the analytical device (Mmeasdev), as well as the corresponding measurement errors (σmeasdev(nrep,measdev)), i.e., the interval in which the true measurements are believed to lie in to a certain probability, triggered by the number of repeats.
The choice of isotopically labeled substrate species, either in pure form or in a mixture, dictates the emerging labeling states of the observable metabolites and therefore significantly impacts flux information [25,33]. Several recent field studies yielded information-optimal tracers in a variety of biological systems and give evidence for a high diversity of flux standard deviations depending on the substrate or substrate mixture. For instance, Walther et al. showed that [1,2-13C]-labeled glucose and mixtures of [3-13C]- and [3,4-13C]-glucose increase statistical identifiability when used with fully labeled glutamate for lung cell carcinoma [34]. Crown et al. identified [3,4-13C]- and [2,3,4,5,6-13C]-labeled glucose to be favorable for elucidating reaction rates in the oxidative pentose phosphate pathway (PPP) and pyruvate carboxylase flux, respectively, based on a small scale network with two free fluxes [35]. Later on, the same group determined [1,2-13C]-, [5,6-13C]-, and [1,6-13C]-labeled glucose as best single tracers for Escherichia coli wild type [36]. A study of Metallo et al. suggested [1,2-13C]-labeled glucose to be the optimal commercial tracer for most fluxes in the PPP and glycolysis in lung carcinoma cell lines while uniformly labeled glutamine provided optimal results for tricarboxylic acid cycle (TCA) fluxes [37]. In theoretical studies, [3,4,5,6-13C]-glucose and [2,3,4,5,6-13C]-glucose resulted to have to best information yield in plants and mammalian cells, respectively [38,39]. Araúzo-Bravo et al. calculated mixtures of 70% unlabeled, 10% U-13C- and 20% [1,2-13C]-labeled glucose to be optimal for flux determination in the cyanobacterium Synechocystis sp. PCC6802 [40]. Schellenberger et al. applied a Monte Carlo sampling technique for experimental tracer design to a large-scale Escherichia coli network and found positional [1-13C] or [6-13C] labeled glucoses to be superior over a commonly used mixture of 20% uniform and 80% unlabeled glucose [41]. Here, unusual multi-positional labeling, in particular [5,6-13C]-, [1,2,5-13C]-, [1,2-13C]-, [1,2,3-13C]-, and [2,3-13C]-glucose, resulted in a higher identifiability than single positional labeling. Nonetheless, no single tracer has been found to outperform all others, an observation which was experimentally confirmed by Crown et al. comparing the outcome of 14 CLEs in Escherichia coli [27]. Importantly, the studies also disclosed a high redundancy in the measurement data, meaning that not all observations effectively contribute to the information gain, although they come at a certain cost. One option to raise flux identifiability that recently has become compelling through advances in lab standardization and miniaturization [42], is the conduction of multiple independent, so called parallel CLEs under identical conditions, each with a different tracer [43] (and references therein). Concurrent fitting of all labeling patterns with a single model obviously increases the measurement-to-flux ratio but, at the same time, also the measurement redundancies. Still, in these and other theoretical and practical studies a part of the fluxes remained non-identifiable [27,44]. Interestingly, a study of Bouvin et al. [45] exemplified, also using a MO-ED approach, that it is indeed possible to find CLEs with comparable information content, but considerably different tracer costs.
In contrast to the work on tracer design, measurement setups have not yet been the target of ED in the field of 13C MFA. The primary analytical methods that are employed are NMR and MS. For both, analytical devices differ not only in the principally observable metabolite/isotopomer spectrum, achievable fragmentation patterns (Fig 2) and the measurement accuracy and sensitivity, but also in terms of analysis speed/throughput, and purchase/maintenance costs (S1 and S2 Text). Since comparative investigations on the inter-platform information content of CLEs for 13C MFA are scarce, in essence, it is still an open question which analytical platform delivers maximal flux information and what the information benefit of multiple-device applications is compared to single-device usage.
For considering economic aspects, the cost contribution of the isotopically labeled substrates, the experimental setup and the analytical technologies are to be specified. Additionally, not only the measurement time on the device, but also spectra evaluation and proofreading processes, possibly with the need for manual post-correction, contribute to the workload. Consequently, such direct and hidden factors play a part in the overall CLE costs. Till now, if at all, only 13C labeled tracers have been considered in CLE costs examinations while further experimental-analytical efforts were neglected so far (see e.g. [45]), meaning that a fine-grained cost function which relates all cost factors to the design parameters has to be set up. The overall cost function of a 13C MFA study is composed of three parts, the experimental, the analytical, and the modeling part. However, the modeling costs such as setting up an adequate model, working through the 13C MFA workflow, calculating and interpreting results etc., heavily depend on the use case and are therefore not considered in the following.
Together, the information and cost criteria Eqs (11)–(14), (18) make up the set of goal functions out of which the objective vector Φ of the MO-ED problem Eq (1) is composed. The design vector α is subject to inequality and equality constraints such as the invertibility conditions on the Fisher matrix (9) and (10), constraints for weights, as well as constraints imposed by reasonable practical resource considerations, e.g., a maximum number of replicates. Since exact handling of integer-valued replicate numbers would result in NP-complete mixed integer nonlinear optimization problems [46], the optimization problem is relaxed by allowing the replicates to take non-integer values. The solution for the relaxed problem is then “rounded” to integers. The full formulation of the MO-ED problem is given in S2 Text.
Solving Eq (1) means to numerically approximate the (potentially infinite) design set α* by an ensemble of Pareto-optimal results [47,48], optimally uniformly distributed covering the whole Pareto front. Particularly successful among these algorithms with respect to convergence and extensity of Pareto front approximation are those based on Particle Swarm Optimization (PSO) with update mechanisms to ensure that the solution ensemble is well-dispersed over the front [49]. For this work, the jMetal (Metaheuristic Algorithms in Java) library, a suite of state-of-the-art MO algorithms is utilized [50]. jMetal is linked to the high-performance 13C MFA simulator 13CFLUX2 [51] via a Java Native Interface (JNI) that enables jMetal to call 13CFLUX2 methods. While 13CFLUX2 is used to evaluate the objectives and takes care of the feasibility of the design parameters, the solution of the MO problem is steered by jMetal routines (Fig 4).
Initially, all experimental, analytical and simulation settings as well as the network model (incl. free flux set and flux values), measurement error models and input species with their respective costs are specified. Depending on the measurement selection proposed by jMetal, the measurement error model is evaluated for the suggested mixture composition in 13CFLUX2 while also taking the number of replicates into account. In turn, statistical flux identifiability is tested and, if one of the invertibility criteria fails, the free flux set is adapted in an iterative procedure: Non-identifiable fluxes are eliminated one-at-a-time by constraining them to their nominal values beginning with the worst determined one, eventually providing the effective number of statistically identifiable fluxes, i.e, p. From the resulting covariance matrix the local information measures ΦD,p etc. are calculated. Furthermore, the expected CLE costs ΦCostsdev are evaluated according to the cost model, given the experimental specification. The objective values are then passed to jMetal, calling the SMPSO algorithm (the rationale for the choice of SMPSO and its parameters is given in S2 Text).
Starting with an initial population created randomly, the swarm is evolved driven by polynomial mutation rules that trigger the choice of the design parameters. In this way, new swarm candidates are proposed out of which Pareto-optimal solutions are selected. The best Pareto solutions are stored in an archive where for each iteration the crowding distance is used to decide which swarm individuals remain in the archive to achieve maximal coverage of the designs. For the newly generated swarm members, measurement values are predicted in silico according to the 13C MFA model using 13CFLUX2 and the corresponding standard deviations are derived from the associated error models. This process cycle is restarted with the next generation of particles until the stopping criterion (i.e., maximum number of generations) is reached. Finally, the archive containing the (best known) Pareto-optimal ensemble is returned and subjected to visual analysis.
Having the Pareto front approximation at hand, the final step of ED involves decision making on the next experiment. In the context of 13C MFA, decision making means to find the most suited experimental-analytical setup out of the range of analytical platforms, input mixture compositions, sets of observable metabolites and replicate numbers. These quantities have different contextual meanings, scales and importance, in the sense of affecting the objective values. Hence, the visual interpretation of MO-ED results faces two challenges:
To tackle these challenges a tailor-made visual interpretation workflow was created (Fig 5). The workflow is composed of three modules, applying different information visualization techniques that (a) allow for visual assessment of the Pareto front, (b) relate the objective with the most important elements of the design space, and (c) compress presentation of the less important design elements.
We developed a framework for information-economic design of CLEs, which we now put into practice.
P. chrysogenum is the primary microbial cell factory for the production of penicillin G and V. Although metabolic engineering strategies have led to strongly improved production efficiencies, the yields of P. chrysogenum are still far below the theoretical maximum [55]. In this situation, 13C MFA is a powerful technique to detect pathway bottlenecks and to guide metabolic engineering efforts. Therefore, this case study explores the Pareto-optimal experimental design spaces in an industrially relevant setting.
With the 13C MFA P. chrysogenum model of at hand, two scenarios differing in the composition of the design objectives were studied. The goal of this first scenario was to profile the analytical platforms according to their information-cost trade-offs and to explore the underlying Pareto-optimal designs. To this end, three objectives were considered, two information criteria and the cost criterion:
Thus, the objective vector is represented by:
Φ=(ΦDoFΦD,p−ΦCostsdev)T
(19)
Due to the number of objectives involved, the MO-ED problem (1) with (19) is hitherto denoted 3D-MO-ED task. Pareto-optimal solutions were calculated and objective values were recorded along with the identifiers of the statistically (non-)identifiable fluxes as well as the number of replicates for each single measurement group. Solutions obtained with models of maximal dimension, p = 21, are discussed in the following (the complete sets of Pareto sets and fronts are provided in S1 Data).
The previous study revealed detailed insights into trade-off CLE designs for P. chrysogenum that relied on the commonly used D-criterion as quantitative information measure. With our second scenario we aimed to study the impact of including additional information criteria on the MO-EDs. To this end, the objective vector is extended by A- and E-information criteria:
Φ=(ΦDoFΦD,pΦA,pΦE−ΦCostsdev)T
(20)
The MO-ED scenario (1) with (20), henceforth referred to as 5D-MO-ED, was performed along the same lines as the 3D-MO-ED study. In the following, selected results are presented and related to the outcomes of the previous ED results. Detailed results are given in the S5 Text.
Intracellular fluxes are of special importance, as they describe the trafficking of metabolites which emerges as the final outcome of all catalytic and regulatory processes acting within living cells. Model-based 13C MFA is the gold standard for the quantification of intracellular metabolic fluxes. A smart combination of tracers and measured labeling patterns, i.e., the tracer composition, measurement groups, number of replicate measurements, are the key to accurate flux determination. Since 13C MFA studies remain complicated and costly, experimental design can safeguard against sub-optimal resource utilization. Using the concept of single-objective ED previous studies provided valuable indications for informative tracer mixtures. These studies have been performed for specific measurement setups and without considering economic aspects in experiment and analytics. To exploit the full power of ED in 13C MFA, here we generalized existing work by simultaneously taking several information quality measures as well as several experimental-analytical cost contributors into account. With a large-scale 13C MFA model and realistic measurement setups at hand, widely used analytical platforms were compared with respect to information-economic design options. With that, tracer and measurement design was performed simultaneously rather than independently. The MO-ED technique for designing informative, yet economic experiments, as showcased with the P. chrysogenum application study, is transferable to virtually any model-based approach and experimental-analytical setup, e.g., to plan parallel CLEs.
MO-ED enables the determination of design ensembles that seek to balance mutually exclusive information- and cost-objectives. Understanding the characteristics of the Pareto sets and the relationships between the different objectives is invaluable to guide the decision process on how to perform the next experiment. However, the sheer size of the design space and the many and various properties of the design parameters pose new challenges for the exploration procedure. First, searching for Pareto-optimal sets exhaustively over the whole, high-dimensional design space is compute-intensive and requires the efficient evaluation of the system model. Here, this challenge was tackled by connecting the high-performance simulator 13CFLUX2 with the optimization library jMetal. Second, a tailored visual analysis workflow was invented that tracks down Pareto-optimal designs thereby relating tracers, measurement groups, replicate numbers, costs, and information measures by means of graphical representations, starting from most relevant (input species) to less informative features (replicates). This workflow aids the scientist to weigh the insights against the costs and, thus, guides decision making.
ED studies were performed for a reference flux distribution representing prior information about the expected fluxes. It is, however, likely that the actual flux distribution under which a CLE is conducted differs from the assumed one. Because the information criteria used in this work rely on local statistical measures, actual Pareto-optimal designs may be widely different from the suggested ones. To investigate the robustness of the 3D-MO-ED Pareto designs in terms of information gain with respect to deviations from the reference flux values, for each platform 10,000 flux distributions were randomly sampled in the bounding box of the corresponding confidence ellipsoids. For the in each case most informative 3D-MO design setting, the D-information criteria values were calculated (for instance in case of LC-MS/MS for pure [1,2-13C]-glucose). In all cases, the average information value of Pareto-optimal results remained in the upper third suggesting that the determined MO-ED designs are reasonably robust (S4 Text and S3 File).
The set of Pareto-optimal labeled tracers for CLEs was found to be remarkably similar across all investigated platforms and platform combinations, e.g., [3-13C]-, [4-13C]-, and [5-13C]-glucose rarely contribute to the designs. However, the quantitative composition of the Pareto-optimal tracers varies widely. Often used, inexpensive substrate mixtures consisting of [1-13C]-, [U-13C]-, and [12C]-glucoses provide moderate statistical identifiability for LC-MS/MS. Several former single-objective ED studies found [1,2-13C]-glucose to be particularly informative (cf. Sec ED approaches in 13C MFA revisited). Although this tracer is more expensive than standard mixtures, our results show that [1,2-13C]-glucose is beneficial to achieve a higher degree of flux confidence across all studied platforms. Our study also reveals that the use of other, more expensive substrate species such as [1,6-13C]-glucose, which seldom have been suggested by conventional ED studies before, is mandatory when a high degree of flux confidence is needed (as measured by the D-criterion), especially for GC-MS, LC-MS, 13C-NMR, GC-MS/LC-MS, 1H-NMR/13C-NMR. These findings yield a generalized view on existing work that focuses on single objective ED aspects.
The study delivers detailed experimental and analytical cost reports for all analytical platforms. 3D-MO-ED results demonstrate, not surprisingly, that the substrate species of choice are the main contributor of the costs. On the other hand, often almost the complete available measurement spectrum, including the maximal number of replicates, contributed to the Pareto-optimal designs arguably because, compared to the substrates, additional measurements come almost for free while they always increase the statistical information gain of the CLE. Only for inexpensive substrate mixtures some measurement groups did not contribute to the designs, most likely due to their redundancy. Hence, savings in analytical costs are possible but only achievable to a lesser extent. Interestingly, robust A-optimal designs emerge to be the most expensive ones across all investigated analytical platforms.
Traditional 1D 13C MFA experimental planning techniques as first proposed by Möllney et al. [25] capture the “value” of a CLE in a single scalar measure of information content which is of limited value. By generalizing the 1D formulation to nD, strikingly, our study demonstrated that the latter gives a much more comprehensive view on Pareto-optimal designs, therewith opening up new possibilities for the experimenter in the planning phase of an experiment. The competition between single criteria is reflected in diverse, partly orthogonal designs. For instance, A-/D- and E-/D-optimal designs but no A/E-optimal designs co-exist for LC-MS/MS. The ability to account for a range of information criteria allows to pro-actively countering undesired side effects caused by (a priori unknown) flux correlations and, thus, could increase the design’s reliability. Importantly, these results were found to be specific to the analytical platform under consideration. Clearly, this wealth of additional insights comes at a computational cost. Here, the generalized ED framework has taken advantage of recent algorithmic advances in 13C MFA [23,51], which paved the way for complex field studies such as reported in this work.
Statistical flux identifiability with a comprehensive metabolic network of P. chrysogenum varies strongly among the measurement techniques. Even acknowledging long analysis times and high equipment costs, LC-MS/MS provides EDs with 50% less costs than other devices due to the use of cheaper input substrates. Simultaneously, LC-MS/MS yields up to ~300% higher information values as compared to the other techniques. Remarkably, the first scenario showed that the spread of Pareto-optimal designs has the highest coverage for LC-MS/MS, thus offering more options to the investigator than GC-MS, LC-MS, and 13C-NMR.
Eventually, the goal of 13C MFA is to measure metabolic fluxes with the highest possible precision. Hence, the question arises whether the extra effort of MO-ED pays off in practice. A use case scenario may be as follows: An ED is desired with overall equally well determined fluxes and as little flux correlations as possible. Analyzing the 5D-MO-ED results with a high E-criterion value, corresponding designs may yield large flux confidence regions. In contrast, A-optimal designs indeed deliver superior designs in the sense of overall flux precision. However, by inspecting the costs associated with A-optimal designs, it becomes apparent that CLEs with A-criterion values overrun the budget. In this situation, alternate A-optimal designs satisfying certain cost constraints can be readily identified and even further ranked by their E- and/or D-criteria values. Having localized the desired Pareto-set(s), the associated designs can be further explored in depth providing detailed specifications of substrate composition and the measurement setup. Operated in that way, we believe MO-ED to become a useful new tool for prospective and rational planning of experiments under full cost control.
Besides deploying the framework to further application fields, there are several options to follow up this work. Technically, dependencies of the MO-EDs on the local design points should be diminished, e.g., by incorporation of global sensitivity analysis [60] or other more advanced design techniques [61] into the framework, to handle scenarios when pre-knowledge on the model parameters is absent. Practically, introducing interactive features to the visual analysis such as browsing, querying, filtering, or sorting could boost the quick understanding relationships within and in-between Pareto sets.
|
10.1371/journal.pntd.0001672 | Cross Neutralization of Afro-Asian Cobra and Asian Krait Venoms by a Thai Polyvalent Snake Antivenom (Neuro Polyvalent Snake Antivenom) | Snake envenomation is a serious public health threat in the rural areas of Asian and African countries. To date, the only proven treatment for snake envenomation is antivenom therapy. Cross-neutralization of heterologous venoms by antivenom raised against venoms of closely related species has been reported. The present study examined the cross neutralizing potential of a newly developed polyvalent antivenom, termed Neuro Polyvalent Snake Antivenom (NPAV). NPAV was produced by immunization against 4 Thai elapid venoms.
In vitro neutralization study using mice showed that NPAV was able to neutralize effectively the lethality of venoms of most common Asiatic cobras (Naja spp.), Ophiophagus hannah and kraits (Bungarus spp.) from Southeast Asia, but only moderately to weakly effective against venoms of Naja from India subcontinent and Africa. Studies with several venoms showed that the in vivo neutralization potency of the NPAV was comparable to the in vitro neutralization potency. NPAV could also fully protect against N. sputatrix venom-induced cardio-respiratory depressant and neuromuscular blocking effects in anesthetized rats, demonstrating that the NPAV could neutralize most of the major lethal toxins in the Naja venom.
The newly developed polyvalent antivenom NPAV may find potential application in the treatment of elapid bites in Southeast Asia, especially Malaysia, a neighboring nation of Thailand. Nevertheless, the applicability of NPAV in the treatment of cobra and krait envenomations in Southeast Asian victims needs to be confirmed by clinical trials. The cross-neutralization results may contribute to the design of broad-spectrum polyvalent antivenom.
| Snake envenomation is a serious public health threat in the rural areas of Asia and Africa. To date, the only proven treatment for snake envenomation is antivenom therapy. Owing to the difficulties in the diagnosis of the biting species, there is a need to develop polyvalent antivenoms that could cross-neutralize venoms of medically important venomous snakes in the various regions. Recently, Thai Red Cross Society from Thailand has developed a new polyvalent antivenom for treatment of cobra and krait venoms. The polyvalent antivenom, termed “Neuro Polyvalent Snake Antivenom (NPAV),” is raised against venoms of two Thai cobras and two Thai kraits. Our results indicated that the polyvalent antivenom can effectively neutralize venoms from many Southeast Asian cobras, kraits and king cobra but is less effective against Indian cobra venoms. Studies using anesthetized rats showed that NPAV can effectively protect against cobra venom-induced cardio-respiratory depressant and neuromuscular blocking effects, confirming that the antivenom can effectively neutralize the major lethal toxins of common cobra venoms. This new antivenom may find potential application in the treatment of elapid bites in Southeast Asia, especially Malaysia, a neighboring nation of Thailand.
| The global figure of snake envenoming cases has been estimated to be greater than 1.8 million annually, with an annual death toll of more than 90,000. Most of the snake envenoming cases occurs in South Asia and Southeast Asia (estimated 720,000 cases, 53,000 fatality), followed by Africa (estimated 420,000 cases, 32,000 fatality) [1], and the main biting species are snakes from the Elapidae and Viperidae families. Among members of the Elapidae family, the cobras and kraits are the main causes of snake envenoming [2], [3]. There are about 34 species of cobras belonging to 7 genera (Aspidelaps, Boulengerina, Hemachatus, Naja, Ophiophagus, Pseudohaje and Walterinnesia). The genus Naja distributed extensively across large regions of the Africa (13 species) and Asia (12 species) [4]. Ophiophagus hannah or commonly known as the king cobra, is the only member of the Ophiophagus genus and is found only in Asia. Bungarus (the kraits), are represented by 12 species and their distribution is confined to the Indian subcontinent, Southeast Asia, as well as Southern China and Taiwan [4]. Cobra and krait envenomations are generally characterized by neurotoxic envenoming [5].
Antivenom therapy is the only effective treatment for snake envenomation. Monovalent antivenoms are raised with venom from one particular species and hence generally only effective in the treatment of envenomation caused by the particular species. Because of the difficulties in accurate diagnosis of the biting species, polyvalent antivenoms that offer paraspecific protection against several venomous snake bites have also been developed and become commercially available. It has been argued that monovalent antivenoms are generally more effective than polyvalent antivenoms, though this has not been firmly established. At present, several types of polyvalent antivenoms against Afro-Asian venomous snakes are available in the market, produced mainly by Asian or African commercial pharmaceutical firms or government institutions [5], [6]. There is, however, a lack of rigorous evaluation of the paraspecific protective actions of these commercially available polyvalent antivenoms. Recently, Thai Red Cross Society produced a new polyvalent antivenom that offers protection against neurotoxic envenomations by elapids in Thailand. This polyvalent antivenom, termed Neuro Polyvalent Snake Antivenom (abbreviated as NPAV) is raised against venoms of four medically important cobras and kraits in Thailand, i.e. Naja kaouthia (Thai monocellate cobra), Ophiophagus hannah (king cobra), Bungarus candidus (Malayan krait) and Bungarus fasciatus (banded krait). In this paper, we report evaluation of the cross-neutralizing potential of NPAV against heterologous venoms of common Afro-Asian cobras (Naja spp. and Ophiophagus hannah) and Asian kraits (Bungarus spp.) We also compared the efficacy of the polyvalent antivenom versus the relevant monovalent antivenom. The results will provide preliminary information as to whether the polyvalent antivenom could find therapeutic application for cobra and krait envenomations outside of Thailand, as well as contribute to the design of a broad-spectrum, Pan-Asian polyvalent antivenom [7].
Venoms of Naja sputatrix, Naja siamensis, Naja kaouthia (Thailand), Naja philippinensis, Naja oxiana, Naja atra, Naja naja (Sri Lanka, sample 1), Naja naja (India, sample 1 and 2), Naja melanoneuca, Naja nigricollis, Naja nubiae, Naja katiensis, Naja haje, Bungarus multicinctus and Bungarus caeruleus were purchased from Latoxan (Valence, France). Venoms of Naja sumatrana, Naja kaouthia (Malaysia), Ophiophagus hannah, Bungarus candidus, Bungarus fasciatus and Bungarus flaviceps were pooled samples obtained from several adult individuals captured in Malaysia whilst Naja naja (Sri Lanka sample 2) was a pooled sample obtained from several adult individuals captured in Sri Lanka. After extraction, the venoms were instantly lyophilized. Two antivenoms were studied: (a) Neuro Polyvalent Snake Antivenom (NPAV) (Lyophilised; Batch no. 0030208; Exp. Date April 21st, 2013), a purified F(ab′)2 obtained from serum of equines hyperimmunized against a mixture of four venoms: Naja kaouthia (Thai monocellate cobra), Ophiophagus hannah (king cobra), Bungarus candidus (Malayan krait) and Bungarus fasciatus (banded krait); (b) Naja kaouthia monovalent antivenom (NKMAV) (Full name: Cobra antivenin; Lyophilised; Batch no. 0090406; Exp. Date August 31st, 2014), a purified F(ab′)2 obtained from serum of equines hyperimmunized specifically against the venom of Thai N. kaouthia. Both of these antivenoms are produced by Queen Saovabha Memorial Institute (QSMI), the Thai Red Cross Society from Bangkok, Thailand. For neutralization studies, both antivenoms were reconstituted in the same manner: 10 mL of normal saline was added to 1 vial of the freeze-dried antivenom. According to the attached fact sheet, 1 mL of the NPAV antivenom is able to neutralize the following amount of snake venoms: 0.6 mg each of N. kaouthia and B. fasciatus venoms, 0.4 mg of B. candidus and 0.8 mg of O. hannah venoms; while 1 mL of the NKMAV can neutralize 0.6 mg of N. kaouthia venom.
Albino mice (ICR strain, 20–25 g) and male Sprague Dawley rats (250–300 g) were supplied by the Laboratory Animal Centre, Faculty of Medicine, University of Malaya. The animals were handled according to the guidelines given by CIOMS on animal experimentation [8]. All experiments involving animals were approved by the Animal Care and Use Committee (ACUC) of the University of Malaya (Ethical clearance letter No. PM/03/03/2010/FSY(R)).
Protein content was determined by Bradford method [9]. All measurements were performed in triplicate. Bovine serum albumin (Sigma, USA) was use to generate a standard curve.
SDS-polyacrylamide gel electrophoresis (SDS-PAGE) was conducted according to the method of Studier [10], using the Bio-Rad broad-range prestained SDS-PAGE standards (6.5–200 kDa) and 15 µL of each antivenom sample (3 mg/mL) was loaded in the gel (12.5%). High performance gel filtration chromatography of the reconstituted antivenom (100 µL, 10 mg/mL) was performed using a Superdex 200 HR 10/30, 13 µm SEC 10×300 mm (GE Healthcare, Sweden). Elution buffer was 100 mM sodium phosphate, 0.15 M NaCl, pH 7.4 at a flow rate of 0.75 mL/min. Protein was monitored by absorbance measurement at 280 nm. The column was calibrated using the following protein standards obtained from Bio-Rad (BIO-RAD Gel filtration Standard): thyroglobulin (670 kDa), γ-globulin (158 kDa), ovalbumin (44 kDa) and myoglobin (17 kDa).
The median lethal dose, LD50, of the venom was determined by intravenous or intramuscular injection into ICR mice (20–25 g, n = 4). The survival ratio was recorded after 48 h to determine the LD50.
In vitro neutralization of lethality was conducted as described by Ramos-Cerrillo et al. [11]. Briefly, a challenge dose of the venom in 50 µL saline was pre-incubated at 37°C for 30 min with various dilutions of the reconstituted antivenom (NPAV or NKMAV) in normal saline, to give a total volume of 250 µL. The mixture was subsequently centrifuged at 10000× g before being injected into the caudal vein of the mice. The number of survival after 48 h was recorded. Generally, the challenge dose used was 5 LD50. However, if 200 µL of the reconstituted antivenom (maximum permitted volume to inject into the mouse) failed to give full protection of the mice, a lower challenge dose of 2.5 LD50 was used instead. The antivenom was considered ineffective when none of the animals injected with the pre-incubated mixture (containing 2.5 LD50 challenge venom in 50 µL saline and 200 µL of the undiluted reconstituted antivenom) survived. Neutralizing potency of the antivenom was expressed as ED50 (the amount of reconstituted antivenom in µL or the ratio of mg venom/mL reconstituted antivenom that gives 50% survival of the animals tested) as well as in term of ‘neutralization potency’ (P, the amount of venom that is completely neutralized by a unit volume of antivenom) calculated according to Morais et al. [12].
This was carried out by intramuscular injection of 5 LD50 or 2.5 LD50 of the venom into mice (n = 4) followed by intravenous injection of 200 µL of appropriately diluted reconstituted antivenom, 10 min later. The number of survival after 48 h was recorded.
The study was conducted on three groups of rats (n = 3, 250–300 g) anesthetized with intraperitoneal injection of urethane (1.4 g/kg, i.p.) to the point of loss of the eyelid reflex and the pedal withdrawal reflex on painful stimuli. The anesthetized animals were surgically prepared for the simultaneous measurement of blood pressure, heart rate, respiratory rate and muscle twitch tension. Data collection and analysis were conducted using PowerLab 4/30 data Acquisition system equipped with LabChart software (AD Instruments, Australia). Rats in group 1 (termed ‘saline/-’ group) were injected with 50 µL saline intramuscularly at 0 min and served as control; Rats in group 2 (termed ‘NsV/-’ group) were injected with 6 mg/kg N. sputatrix venom (dissolved in 50 µL saline) intramuscularly; and rats in group 3 (termed ‘NsV/NPAV’ group) were injected with the same dose of venom intramuscularly followed by intravenous administration of 3 mL of the reconstituted NPAV (1 mL each at 10 min, 30 min and 50 min post-injection of the venom). The volume of antivenom administered and the time points were chosen to ensure no disturbance of the blood pressure, heart rate and respiratory rate occurred.
LD50 of the venoms and ED50 of antivenoms are expressed as means with 95% confidence intervals (C.I.). LD50, ED50 (median effective dose) and the 95% confidence intervals (C.I.) were calculated using the probit analysis method of Finney [13] with the BioStat 2009 analysis software (AnalystSoft Inc.). The statistical analysis for pharmacological study was conducted using SPSS. The data (expressed as mean ± S.D.) were analyzed using one-way ANOVA, with Tukey's post hoc multiple-comparison test, with P<0.05 as significant.
The protein contents of reconstituted NPAV and NKMAV were 20.3 mg/mL and 12.5 mg/mL, respectively. The SDS-PAGE patterns of NPAV (Neuro Polyvalent Snake Antivenom) and NKMAV (Naja kaouthia monovalent antivenom) indicated that there was no distinct band of high or low molecular weight proteins (Fig. 1). The same was also observed in the gel-filtration chromatographic profiles of the two antivenoms (Fig. 2). Based on ‘area-under-the-curve’ comparison, the quantity of F(ab′)2 in both these antivenoms were comparable (91–96%) whilst the quantity of respective dimers and low molecular weight proteins in NKMAV were slightly higher than those in NPAV (Figure 2). No high molecular weight aggregates were detected in both of these antivenoms.
The results of the in vitro neutralization of venom lethality by the NPAV and the monovalent N. kaouthia antivenom (only for selected venoms) are shown in Table 1 and 2, respectively. The results showed that NPAV was able to confer protection/cross-protection against the venoms of all krait as well as almost all the Afro-Asian cobra venoms examined (except for the African spitting cobra N. katiensis), although the neutralizing potency range varied from low to high. The NKMAV was able to neutralize the venoms of six Asian cobras (N. kaouthia [Thailand & Malaysia], N. sputatrix, N. sumatrana, N. siamensis and O. hannah) tested, with efficacy comparable to that of NPAV, but failed to neutralize the venoms of B. fasciatus and B. candidus. It is interesting to note that our results on the neutralization potentials of NPAV against N. kaouthia, B. fasciatus, B. candidus and O. hannah venoms are much higher than stated in the antivenom fact sheet provided, in particular against O. hannah venom. It should be noted, however, that the definition of neutralization potential used in the fact sheet has not been clearly stated.
Table 3 shows the in vivo neutralization of lethality of four venoms from Asian cobras and krait (N. sputatrix, N. kaouthia (Thailand), N.kaouthia (Malaysia) and B. candidus) by NPAV. It is interesting to note that the i.m.LD50 values determined were comparable to the corresponding i.v.LD50 values. Our results showed that NPAV effectively neutralized all four venoms in vivo, with ED50 comparable to the corresponding ED50 in the in vitro neutralization assays.
Figure 3 shows that NPAV was able to fully protect against the N. sputatrix venom-induced cardio-respiratory depressant and neuromuscular blocking effects in the anesthetized rats. The mean blood pressure (BP), heart rate, respiratory rate and muscle twitch tension of the control anesthetized rats (‘saline/-’ group) remained constant for at least 6 hours after the initial stabilization. Following an intramuscular injection of the venom at 4×LD50 dose (6 mg/kg), however, there was an immediate small decrease (about 20%) in the BP, which remained constant thereafter for the next 90 min (Figure 3A). During this period, the heart rate remained essentially unaffected (Figure 3B). And then the BP and heart rate both began to fall precipitously from 90 min onward. On the other hand, the respiratory rate and muscle twitch tension were stable only for the first 60 min after venom administration. Both these parameters, however, began to decrease sharply thereafter (Figure 3C and 3D) and death occurred at 125–130 min after the venom injection.
In a parallel series of experiment conducted to examine the ability of NPAV to protect against the cardiovascular depressant and neuromuscular blocking effects of N. sputatrix venom, 1 mL each of NPAV was administered via the left jugular vein, at 10, 30 and 50 min, respectively, following the i.m. injection of the venom. The results showed that the administration of the antivenom effectively reversed the cardio-respiratory depressant and neuromuscular blocking effects induced by the N. sputatrix venom. The heart rate and muscle twitch tension of the antivenom-treated animals (‘NsV/NPAV’ group) were restored to the same levels as the ‘saline/-’ control group (P>0.05, not significantly different between ‘saline/-’ group and ‘NsV/NPAV’ group by one-way ANOVA), whereas the blood pressure and respiratory rate were restored to 80% of the control group (P<0.05 between the two groups) and remained at that level throughout the monitoring period.
In this report, the neutralization capabilities of the antivenoms are expressed in three different ways: the commonly used median effective dose ED50 in µL antivenom, ED50 in mg/mL and potency, P, as defined by Morais et al. [12]. For ED50, expressing the value in term of mg venom neutralized per mL antivenom is a more realistic assessment of the neutralization capabilities of the antivenom than in term µL antivenom or number of mouse LD50, because the LD50's of the various cobra venoms differ substantially. For example, expressed in term of µL antivenom, the ED50's of NPAV against N. kaouthia (Malaysia) venom and N. philippinensis venom are comparable, but when expressed in term of mg/mL, NPAV was obviously much more effective against N. kaouthia (Malaysia) venom then against N. philippinensis venom.
Because of the high lethality of certain venoms, a 2.5 LD50 instead of the standard 5 LD50 was used as the challenge dose. Since the ED50 value of an antivenom is highly dependent on the challenge dose, ED50 obtained from the two different challenge doses cannot be compared directly. As such, we also expressed neutralization capability of the antivenom in terms of P (potency), which is the mass of venom that is completely neutralized per unit volume of antivenom, as defined by Morais et al. [12]. Potency gives a better estimate of the relative efficacy of antivenoms than comparing ED50 values when different challenge doses were used in the determination of ED50, as P is theoretically independent of the amount of challenge doses. Nevertheless, since the relationship between antivenom neutralizing capability versus venom challenge dose is not necessarily linear due to the complexity of venom and antibodies composition, comparing the P values of antivenom when using different challenge doses does have its limitation.
The relative merit of monovalent antivenoms versus polyvalent antivenoms has been the subject of much discussion [14] and there are some authors who suggested that monovalent antivenoms are generally more potent than polyvalent antivenoms and less likely to cause adverse reactions as it may involve administration of a lower quantity of antivenom IgG that with a polyvalent antivenom. Several studies have shown, however, that this is not necessarily true [14], [15]. Our results here demonstrated that the quality and neutralization capabilities of polyvalent antivenom are not necessary inferior to that of monovalent antivenoms. Here we compared the protein composition of the N. kaouthia monovalent antivenom(NKMAV) and Neuro Polyvalent Snake Antivenom (NPAV), as well as the in vitro neutralization potency of the two antivenoms against venoms from five common Asiatic cobras and O. hannah (Table 2). The results show that both antivenoms are devoid of high molecular weight aggregates, the compounds that are usually associated with adverse reactions. The protein contents of the two antivenoms are both relatively low (20.3 mg/mL and 12.5 mg/mL respectively, for NPAV and NKMAV). The neutralization potencies of NKMAV and NPAV against N. kaouthia (Thailand) venom (the venom used to raise both antivenoms) are essentially the same. This, however, is not surprising since according to the manufacturer, monovalent antivenoms (including NKMAV) were later added to the purified polyvalent antivenom to ensure the neutralizing potency of the NPAV was comparable with the neutralizing potencies of the respective monovalent antivenoms. What is interesting is that, the neutralization potencies of both antivenoms against four Asiatic cobra venoms tested are comparable. NPAV, however, is much more potent than NKMAV in neutralizing O. hannah venom. This is to be expected as O. hannah venom was included in the immunogen mixture used in raising NPAV. These observations are in accordance with the conclusion drawn by Raweerith and Ratanabanangkoon [15].
NPAV was raised using the four common elapid venoms in Thailand: N. kaouthia (Thailand), O. hannah, B. candidus and B. fasciatus. Our results showed that the polyvalent antivenom could effectively neutralize venoms of the same four species that originated from Malaysia. It is interesting that the polyvalent antivenom could effectively neutralize venom of N. kaouthia from Malaysia too. According to Wüster and Thorpe [16], the composition of the venom of Thai N. kaouthia is substantially different from that of the Malaysian N. kaouthia, the former appears to be more neurotoxic, while the latter more necrotic. It was suggested that this difference in the toxin composition may result in antivenom incompatibility. Our results, however, showed that while the two N. kaouthia venoms did differ substantially in their venom i.v. LD50, the polyvalent antivenom that was raised from Thai N. kaouthia venom (together with other Thai elapid venoms) could effectively neutralize the venom of Malaysian N. kaouthia, albeit with a moderately lower ED50. The same cross-neutralization potency was also observed with the monovalent N. kaouthia antivenom.
In addition to the four elapid venoms mentioned, NPAV also effectively neutralized the venoms of the other Malaysian kraits B. flaviceps and B. fasciatus, as well as the medically important Equatorial spitting cobra N. sumatrana. Thus, the results of our in vitro neutralization studies suggest that NPAV, which is prepared from Thai elapid venoms, can be useful in the treatment of elapid envenoming in Malaysia, since the polyvalent antivenom can effectively neutralize venoms of all medically important elapids in Malaysia.
Our results showed that NPAV was effective (P>0.5 mg/mL) against other Asiatic cobra venoms, including the venoms of the three Southeast Asian spitting cobras N. siamensis, N. sumatrana and N. sputatrix, as well as the venoms of the Central Asia cobra N. oxiana and Chinese cobra, N. atra. Earlier report claimed that Chinese cobra N. atra venom was poorly neutralized by other commercial cobra antivenoms [17]. NPAV, however, was only weakly effective against the venom of N. philippinensis, the highly neurotoxic (lethal) Philippine cobra. The low neutralization potency of 0.11 mg/mL would mean that more than 50 vials of the antivenom may be required for the victim, as the amount of venom injected during a cobra bite can be more than 50 mg (dry weight). Indian polyvalent antivenom (raised from N. naja venom) has also been reported to show poor neutralizing ability against N. philippinensis venom.
NPAV was only moderately effective against N. naja venoms from the Indian subcontinent with the exception of one Sri Lankan N. naja venom sample (mixtures collected from several adult N. naja). The variation suggests geographical differences in N. naja venom.
Thus, our results suggest that NPAV is effective (P>0.5 mg/mL) in neutralizing the venoms of Asiatic cobras from Southeast Asia (except the Philippines cobra), Central Asia and China, but only moderately effective against venoms from the cobras from the Indian subcontinent. These observations can be used as the basis for the design of a polyvalent antivenom with a broader spectrum of cross-neutralization. For example, if venoms of N. philippinensis, N. naja and B caeruleus were added to the immunogen mixture to raise NPAV, the resulting polyvalent antivenom may well be an effective Pan-Asian polyvalent cobra and krait antivenoms.
Five common African cobra venoms, including venoms from three spitting cobras (N. nigricollis, N. nubiae and N. katiensis) were selected for this study. NPAV could neutralize effectively the venoms from N. melanoleuca, but is moderately effective against that of the spitting cobras N. nigricollis and N. nubiae, and weakly or not effective against venom of the highly lethal N. haje and the spitting cobra N. katiensis. We have not carried out a thorough study of the neutralization potency of NPAV against African cobra venoms, as many of the venoms are not available to us. Nevertheless, the results showed that there are still substantial cross-neutralizations between major venom toxins of the Asiatic and the African Naja, despite the fact that the Asian group lies in a more distant branch from the African group in the phylogenetic dendrogram [18].
It is interesting to note that the NPAV could effectively neutralize the venoms of the three Southeast Asian kraits (B. fasciatus, B. candidus and B. flaviceps), but only moderately or weakly against the venoms of the other two kraits (B. multicinctus and B. caeruleus). The results indicate that krait venoms share enough common antigens among them to enable the NPAV raised against B. fasciatus and B. candidus to neutralize all 5 krait venoms tested. However, the low to moderate potency of NPAV against B. multicinctus and B. caeruleus venoms indicates significant differences in antigenicity of some of the venom toxins.
For all four elapid venoms tested, the i.m LD50 values are comparable to the i.v. LD50, indicating that the main venom toxins could diffuse effectively from muscle to circulation, presumably because these toxins are mainly low molecular weight proteins (phospholipases A2 and the three-finger toxins). Also, the neutralization capability (as measured by ED50) of NPAV in the in vivo assay is comparable to that of in vitro assay, suggesting that neutralization potential of antivenom against elapid snakes measured by the usual in vitro neutralization assay does provide a good indication of its effectiveness in the in vivo situation.
To further examine the in vivo neutralization capability of the NPAV, we examined its ability to protect against N. sputatrix cobra venom-induced cardio-respiratory depressant and neuromuscular blocking effects in anesthetised rats. The antivenom (3 mL in total) was administered in 3 separate injections to minimize disturbing the cardio-respiratory parameters. It is well established that the major lethal toxins of N. sputatrix venom consist of polypeptide neurotoxins, polypeptide cardiotoxins and phospholipases A2 [19]. The venom-induced cardio-respiratory depressant effect was likely due to the combined actions of polypeptide cardiotoxins and phospholipases A2 whereas the neuromuscular blocking effect was likely due mainly to the action of the venom polypeptide neurotoxins, although venom phospholipases A2 may also play a role. Thus, the ability of NPAV to effectively reverse the N. sputatrix venom-induced cardio-respiratory depressant and neuromuscular blocking effects in the rats demonstrated that the antivenom did contain specific antibodies that could effectively neutralize the major lethal toxins of N. sputatrix venom.
In conclusion, the in vitro and in vivo neutralization studies indicated that Neuro Polyvalent Snake Antivenom (NPAV) effectively neutralized venoms from many Southeast Asian Naja, Bungarus and Ophiophagus hannah but less effective against the venoms of Naja from the India subcontinent and Africa, as well as the Asiatic N. philippinensis. This cross-neutralization information can be used as the basis for the design of broader-spectrum polyvalent cobra antivenom. The abilities of NPAV to protect against N. sputatrix venom-induced cardio-respiratory depressant and neuromuscular blocking effects confirmed that the antivenom effectively neutralized the major lethal toxins of Naja venoms. The antivenom may find potential application in the treatment of elapid bites in Southeast Asia, especially Malaysia, a neighboring nation of Thailand. Nevertheless, the applicability of NPAV in the treatment of cobra and krait envenomations in Southeast Asia needs to be confirmed by clinical trials, as it is known that antivenom that has been proved effective in murine model is not necessarily effective in treating human victims [20].
|
10.1371/journal.ppat.1005832 | Cytomegalovirus Infection Leads to Development of High Frequencies of Cytotoxic Virus-Specific CD4+ T Cells Targeted to Vascular Endothelium | Cytomegalovirus (CMV) infection elicits a very strong and sustained intravascular T cell immune response which may contribute towards development of accelerated immune senescence and vascular disease in older people. Virus-specific CD8+ T cell responses have been investigated extensively through the use of HLA-peptide tetramers but much less is known regarding CMV-specific CD4+ T cells. We used a range of HLA class II-peptide tetramers to investigate the phenotypic and transcriptional profile of CMV-specific CD4+ T cells within healthy donors. We show that such cells comprise an average of 0.45% of the CD4+ T cell pool and can reach up to 24% in some individuals (range 0.01–24%). CMV-specific CD4+ T cells display a highly differentiated effector memory phenotype and express a range of cytokines, dominated by dual TNF-α and IFN-γ expression, although substantial populations which express IL-4 were seen in some donors. Microarray analysis and phenotypic expression revealed a profile of unique features. These include the expression of CX3CR1, which would direct cells towards fractalkine on activated endothelium, and the β2-adrenergic receptor, which could permit rapid response to stress. CMV-specific CD4+ T cells display an intense cytotoxic profile with high level expression of granzyme B and perforin, a pattern which increases further during aging. In addition CMV-specific CD4+ T cells demonstrate strong cytotoxic activity against antigen-loaded target cells when isolated directly ex vivo. PD-1 expression is present on 47% of cells but both the intensity and distribution of the inhibitory receptor is reduced in older people. These findings reveal the marked accumulation and unique phenotype of CMV-specific CD4+ T cells and indicate how such T cells may contribute to the vascular complications associated with CMV in older people.
| Cytomegalovirus (CMV) is a member of the herpesvirus family and most humans carry chronic CMV infection. This drives the development of large expansions of CD8+ CMV-specific T cells, which increase further during ageing. CMV infection is associated with vascular disease and increased risk of mortality in older people, which may be related to damage from this CMV-specific immune response. Here we used a set of novel reagents called HLA class II tetramers to make a detailed study of CMV-specific CD4+ T cells. We show that CMV-specific CD4+ T cells are found at remarkably high frequencies within blood, representing up to a quarter of all such white cells. In addition they demonstrate a range of unique features. Firstly they carry a chemokine receptor that directs the cells to activated endothelial cells within blood vessels. Secondly, they express epinephrine receptors which would allow them to respond rapidly to stress. Finally, these CD4+ T cells are unique as they are strongly cytotoxic and equipped with the ability to directly kill virally-infected cells. HLA class II tetramers therefore reveal a profile of unique features which provide insight into how CMV-specific CD4+ T cells may be involved in vascular immunopathology.
| The effective control of infectious agents requires a range of different arms of the immune system. CD4+ T cells play a pivotal role in orchestrating these events, including support for both antibody production and the expansion and effector function of CD8+ T cells. However it is now well established that CD4+ T cells can also exert crucial effector functions which may be mediated by cytokine production or direct cytotoxicity [1–4]. In chronic viral infections such as cytomegalovirus (CMV) these effector functions are important for control of lytic replication and suppression of viral reactivation. Human leukocyte antigen (HLA) class I tetramers have made a huge contribution to the study of antigen-specific CD8+ T cell immune responses through their ability to allow the visualisation and phenotypic analysis of cells isolated directly from blood and tissue [5]. In contrast, the study of antigen-specific CD4+ T cells has been limited by the relative lack of HLA class II tetramers. Although virus-specific CD4+ T cells can be detected relatively easily by their functional response following exposure to antigen, this alters their phenotype and transcriptome and does not permit analysis of the resting cellular profile. As such, much less is known about the profile and function of antigen-specific CD4+ T cells.
CMV is a β-herpesvirus that establishes a lifelong chronic infection and which is well controlled in healthy people. Initial infection leads to the expansion of very large numbers of virus-specific T cells within peripheral blood which are maintained throughout life and increase even further with age [6–8]. The virus has evolved multiple mechanisms to evade HLA class I and class II-restricted T cell immune responses and a state of functional latency is established after infection, which is associated with intermittent episodes of viral replication (reviewed in [9,10]). HLA class I tetramers have revolutionized the study of the CMV-specific CD8+ immune response and have been pivotal in defining the unique immunodominance of the virus, the phenotype of virus-specific cells and unique aspects of their transcriptome [11]. CMV-specific CD4+ T cells are also critical effector populations in the control of CMV infection where they maintain function of virus-specific CD8+ T cells and suppress viral replication at specific tissue sites [12–17]. Indeed, delayed reconstitution of CMV-specific CD4+ T cells correlates with viral reactivation and CMV disease in organ transplant recipients and is associated with prolonged urinary viral shedding in children undergoing primary infection [18]. In relation to murine CMV, mice lacking CD4+ T cells are impaired in their ability to clear virus from the salivary gland which is an important site of viral latency [19,20].
CMV-specific CD4+ T cells have been identified in vitro using cell culture and epitope screening technology. Indeed, the use of peptide pools spanning the whole viral proteome has shown a very broad and strong CD4+ T cell response against many viral proteins of which the most immunodominant are glycoprotein B (gB) and the major tegument component phosphoprotein 65 (pp65) [21]. These studies have shown that the CMV-specific CD4+ T cell response is of unusually strong magnitude and increases further during ageing [15,22–24]. However, such analyses have relied on the interrogation of cells that have been stimulated with antigen for several hours prior to analysis and the almost complete absence of HLA class II tetramers has greatly limited the ability to determine the profile of virus-specific CD4+ T cells directly ex vivo.
HLA class II tetramers have recently been used to identify T cells recognising influenza [25,26], hepatitis C virus [27,28], HIV [29] and Epstein-Barr virus [30]. Here we have used three HLA class II tetramers to carry out the first comprehensive analysis of the phenotypic and transcriptional profile of unmanipulated CMV-specific CD4+ T cells. We show that CMV-specific CD4+ T cells are found at very high frequencies within peripheral blood, exhibit a highly differentiated and cytotoxic phenotype which would target them to activated endothelium through CX3CR1-fractalkine binding. These features reveal the extraordinary magnitude of the CMV-specific CD4+ T cell pool that must be maintained to suppress viral reactivation and indicate potential mechanisms that may underlie the development of vascular disease during chronic CMV infection.
Glycoprotein B and pp65 are the two most immunodominant target proteins for CMV-specific CD4+ T cells. We obtained HLA-peptide tetramers that contained three epitopes derived from gB or pp65, the gB-derived DYSNTHSTRYV peptide restricted by HLA-DRB1*07:01 (DR7) as well as two pp65-derived epitopes, AGILARNLVPMVATV and LLQTGIHVRVSQPSL, which are restricted by HLA-DRB3*02:02 (DR52b) and HLA-DQB1*06:02 (DQ6) respectively (Table 1). These epitopes are subsequently named by the first three amino acids of their respective peptide sequence throughout this paper.
To confirm the specificity of all three tetramers we initially used the reagents to stain CD4+ T cell clones specific for the cognate HLA class II-peptide complex. This confirmed strong and specific binding whilst very little background was observed following staining of peripheral blood molecular cells (PBMCs) from CMV-seronegative individuals who expressed the appropriate HLA allele contained within each tetramer (Fig 1). The sensitivity of detection of virus-specific T cells through use of tetramer staining was defined by mixing aliquots of peptide-specific T cell clone (5%, 1%, 0.5%, 0.25% and 0.1%) with PBMCs taken from a CMV-seronegative individual. This approach showed that CMV-specific T cells could be identified reliably at frequencies as low as 0.01–0.05% of the total CD4+ T cell population (Fig 1A)
We next went on to use the HLA class II tetramers to enumerate CMV-specific CD4+ T cells within the peripheral blood of healthy donors. PBMCs were isolated from 73 CMV-seropositive individuals between the age of 24 and 88 years who all expressed the appropriate HLA class II allele contained within the tetramer (Table 1). These were then stained directly with the HLA class II tetramer and analysed by flow cytometry (Fig 1B and 1C). CMV-specific CD4+ T cells were observed in 74% of the donors that were tested. The median frequency of CMV-specific CD4+ T cells was 0.45% of the total CD4+ subset, and this varied between 0.75%, 0.21% and 0.66% for T cells specific for the LLQ, AGI and DYS epitopes respectively. The proportion of CD4+ epitope-specific T cells ranged from 0.01% up to a remarkable value of 24% of all CD4+ T cells within one individual. Interestingly, 26% of the donors (19/73) carried peptide-specific T cell populations representing over 1% of the total CD4+ T cell pool.
An increase in the number of CMV-specific CD8+ T cells in association with aging, sometimes termed ‘memory inflation’, has been demonstrated through the use of HLA-peptide tetramer staining. We therefore analysed the frequency of virus-specific CD4+ T cells in relation to age (Fig 1D). Although the very largest tetramer-staining populations were indeed identified in the older donors in our study, no clear increase in CMV-specific CD4+ T cells was observed with age as many younger donors also carried substantial frequencies of CMV-specific CD4+ T cells.
CMV-specific T cells have been detected previously by Interferon (IFN)-γ production following antigen stimulation [23,31] and we were interested to compare the relative number of CD4+ T cells identified by HLA class II tetramers compared to this functional response. Analysis of intracellular cytokine staining (ICS) for IFN-γ production after stimulation with CMV peptide was therefore performed within a panel of donors. A strong correlation was observed between these two values (rSpearman = 0.83; p = 0.008; Fig 1E) although it was of interest that the number of cells detected by tetramer staining was greater than the value obtained by cytokine detection. This indicates that both the number of virus-specific cells has been underestimated in previous studies using cytokine detection and that peptide-specific T cells display other functional responses in addition to production of IFN-γ.
To further investigate the functional properties of CMV-specific CD4+ T cells following activation, we next went on to stimulate PBMCs with peptide prior to assessment of the profile of cytokine production using ICS. The predominant expression pattern was of combined IFN-γ, TNF-α and MIP-1β (CCL4) production, with a further subset which failed to generate MIP-1β (Fig 2). Of note, the proportion of TNF-α+ cells in most donors exceeded that of IFN-γ+ cells. Interestingly, in three individuals the proportion of IFN-γ+ cells was only between 36–65% and large populations of cells were observed which produced IL-4, usually in isolation and sometimes in combination with other cytokines. Indeed, in one donor these comprised up to 60% of peptide-specific cells. Virtually no CMV-specific CD4+ T cells produced IL-17A or IL-10 in response to antigen stimulation and no significant differences were observed between T cells recognising the gB or pp65-derived epitopes.
We next undertook a more detailed analysis of the phenotype of CMV-specific CD4+ T cells. Expression of CCR7 and CD45RA was used to define naïve, central memory (CM), effector memory (EM) and revertant CD45RA+ effector memory (EMRA) cells. A median of 88% of CMV-specific CD4+ T cells displayed an EM phenotype (CCR7-CD45RA-) with only 3.3% of cells expressing a CCR7+CD45RA- profile typical of CM cells (Fig 3A). In addition, re-expression of CD45RA was found on only a minor subset (1.8%) of effector memory cells, in marked contrast to the profile observed commonly on CMV-specific CD8+ T cells [6]. The surface expression of the additional differentiation markers CD27, CD28, CD57 and CD45RO was then assessed in order to undertake a more detailed phenotypic analysis. A Boolean gating strategy was used to investigate expression patterns of these markers and determine the differentiation hierarchy of the CMV-specific CD4+ T cells (Fig 3B). As anticipated, the majority of the global CD4+ T cell population in each donor expressed a naïve phenotype but CM and EM populations were also observed in comparable proportions (S1 Fig). The differentiation pattern of the CMV-specific T cells was very different (Fig 3B) and this, when combined with current understanding of T cell biology [32,33], allowed us to model the profile of CMV-specific CD4+ T cell differentiation (Fig 3D). This revealed that CMV-specific CD4+ T cells can be detected in several stages of differentiation and reveals progression through a dual CD45RA+CD45RO+ stage prior to loss of CD45RA expression and attainment of CCR7+CD45RO+ central memory phenotype. Further differentiation led to downregulation of CCR7 followed by a sequential loss of CD27 and CD28 expression. Indeed, 64% of cells exhibited a predominant CD27-CD28- profile whereas 22% (12/55) displayed a largely CD27-CD28+ profile. CD57 expression is a predominant feature of CMV-specific T cells and was observed almost exclusively on CD27-CD28- cells, with minor expression on the CD27-CD28+ population. A final stage of differentiation, in a minority of cells, was the re-expression of CD45RA which coincided with complete loss of CD45RO expression.
Further examination of virus-specific populations within individual donors revealed a moderate degree of heterogeneity in relation to differentiation status. Indeed it was noteworthy that 9% (5/55) of responses were characterised by a dominant central memory phenotype whilst only 3 responses exhibited late stage CD45RA+ effector memory (EMRA) differentiation (Fig 3B). As such we next went on to examine potential factors that might be related to the differentiation profile of virus-specific CD4+ T cells. Interestingly, this was not correlated with the magnitude of the immune response, a pattern that is different to the profile for virus-specific CD8+ T cells where clonal expansion is associated with a greater degree of differentiation [6]. In contrast, antigenic specificity may be an important factor as CD4+ T cells specific for DYS (glycoprotein B) displayed a more differentiated phenotype compared to pp65-specific T cells (Fig 3C). As such, loss of CD28 or gain of CD57 expression was seen on 62% and 41% of gB-specific T cells respectively, compared to only 30% and 18% of CD4+ T cells specific for the epitopes from pp65. Moreover, an EMRA phenotype was observed only on the CD4+ T cells which were specific for DYS.
The availability of HLA-peptide tetramers allows the direct analysis of antigen-specific T cells without prior stimulation and this was felt to be particularly valuable in the assessment of CD4+ T regulatory function as FoxP3 expression can be induced following activation through the TCR [34]. In order to investigate whether CMV-specific CD4+ T cells contain natural T regulatory cells we stained cells directly ex vivo with HLA class II tetramer, anti-CD4, anti-CD25, anti-CD127 and intracellular anti-FoxP3. However, virtually no CMV-specific T cells were found to exhibit a CD4+CD25+CD127low/-FoxP3+ T regulatory phenotype (S2 Fig).
The availability of HLA class II-peptide tetramers allowed us to undertake direct purification and transcriptional analysis of CMV-specific CD4+ T cells, an approach that has been important in relation to determining novel features of the equivalent CD8+ T cell subset [11]. CMV-specific CD4+ T cells were isolated from the blood of five CMV-seropositive donors by staining with tetramer followed by high purity cell sorting. Two of these populations were specific for epitope DYS and three recognised the peptide LLQ. Effector memory T cells isolated from CMV seronegative individuals were used as a comparator group.
The pattern of normalised gene expression was compared initially between the combined transcriptome of the CMV-specific T cell samples and the effector memory population from CMV-negative donors. Global expression patterns were broadly similar between the two groups, reflecting the shared effector memory phenotype. However 55 mRNA transcripts differed by at least two-fold expression between the two groups, of which 35 were upregulated in CMV-specific T cells and 20 genes were lower within this group (Fig 4A and S1 Table). We also compared the individual transcriptional profiles of DYS- and LLQ-specific T cell populations and here 12 of the 55 genes that exhibited differential expression between the combined profile of CMV-specific and control EM cells were also differentially expressed in both the DYS- and LLQ-specific T cells. 36 genes were altered only within the DYS-specific populations and 7 genes exhibited differential regulation within LLQ-specific T cells alone (Fig 4B, S2 and S3 Tables), probably reflecting the more marked differentiation profile observed for the DYS-specific population. Relative expression levels (aquantile normalised expression) for selected transcripts are depicted in Fig 4C comparing DYS and LLQ-specific CD4+ T cells, as well as CD4+ EM cells from CMV seronegative donors. An increase in relative transcription levels was often observed for LLQ-specific T cells which was then further enhanced in DYS-specific T cells explaining why more significant differences in gene expression were observed in comparisons between DYS-specific T cells only and EM T cells.
The function of many proteins encoded from the genes upregulated in CMV-specific T cells is related to cytotoxic function, such as granzymes B, H and A, granulysin and perforin. Expression of the chemokines CCL3 (MIP-1α) and CCL4 (MIP-1β) was strongly increased and indicates an important role for CMV-specific CD4+ T cells in attracting cells of the innate immune system to the site of viral recognition. The increased pattern of transcription of CX3CR1 in DYS-specific T cells is of particular note as this chemokine receptor has been shown to be a discriminative marker for CMV-specific CD8+ T cells and is thought to attract cells to areas of stressed endothelium which express the membrane-bound ligand fractalkine [11]. In addition we observed marked overexpression of ADRB2, the gene encoding the β2-adrenergic receptor, on these cells which forms an important link between the sympathetic nervous system and the immune system. Additional upregulated genes of interest in CMV-specific T cells included the G protein coupled-receptor GPR56 and fibroblast growth factor-binding protein 2 (FGFBP2), both of which have been previously associated with cytotoxic activity, and the secreted extracellular matrix protein SPON2. As changes in level of transcription do not always translate into the same changes at protein level, further analysis would be needed to confirm some of these observations.
Several genes were downregulated in CMV-specific T cells of which the most striking pattern was seen for ADAMTS6, a member of the ADAMTS family (a disintegrin and metalloproteinases with thrombospondin). These secreted proteins have roles in mediating cell adhesion and proteolytic shedding and it is of interest that ADAMTS6 expression is increased by TNF-α [35]. The physiological importance of this will require further investigation as the substrate for ADAMTS6 is currently unknown. CD27 expression was also reduced, reflecting its marked reduction at the cell surface, and levels of TNFRSF4 (OX40) transcript, which is induced following cell activation, was also low suggesting that CMV-specific CD4+ T cells are largely resting in the steady state.
To further validate results from the microarray analysis, and investigate differences observed between LLQ- and DYS-specific CD4+ T cells, we performed qPCR analysis for genes that were found to exhibit differential expression on microarray. These include the chemokines CCL3, CCL4 and CCL5, GZMB and perforin, as well as CX3CR1 and the ADRB2 gene. For all seven genes we confirmed increased levels of transcription in CMV-specific T cells compared to the CD4+EM population (Fig 4D). The pattern of expression was broadly reflective of that seen in the microarray analysis. Particularly high transcript levels in CMV-specific T cells were observed for GZMB (30 to 50-fold increase), CCL4 (6 to 19-fold increase) and CX3CR1. Of note, in DYS-specific T cells transcription of CX3CR1 was found to be 1000-fold higher than in EM cells and levels were also 12-fold higher in LLQ-specific T cells. Gene expression of ADRB2 was also increased in both LLQ- and DYS-specific T cells.
We next went on to investigate the protein expression of four genes whose transcription had been revealed to be strongly upregulated by microarray analysis. As such, tetramer staining was combined with antibodies to granzyme B, perforin, FasL and CX3CR1. CMV-specific CD4+ T cells were found to possess a very strong cytotoxic phenotype with up to 96% of cells staining positive for granzyme B on direct ex vivo analysis (Fig 5A). This pattern was particularly strong for glycoprotein B-derived DYS-specific T cells which exhibited a median expression level of 78%, compared to 61% and 45% of the pp65-derived LLQ- and AGI-specific T cells respectively. Perforin expression was also found on 57%, 39% and 27% of these three populations respectively and all T cells that expressed perforin showed co-expression of granzyme B. A very strong correlation was observed in relation to the expression of CX3CR1 with cells of a cytotoxic phenotype (Fig 5A). In CMV seronegative individuals the proportion of CD4+ EM T cells expressing markers of cytotoxicity was only around 1%.
Given the clinical importance of CMV infection in older people we further analysed the expression of granzyme B, perforin and CX3CR1 in relation to the age of the donor (Fig 5B). Interestingly, the substantial cytotoxic potential of CMV-specific CD4+ T cells was found to increase even further with aging and this was particularly the case for pp65-specific T cells, within which perforin expression increased from 18% within donors aged 20–35 years compared to 43% in those aged over 60 years. The cytotoxic profile of DYS-specific T cells also tends to increase with age but this effect was less marked as the cytotoxic phenotype was already strongly established in these cells at an early age. The expression of CX3CR1 was consistently high on CMV-specific T cells from donors at all age groups with 80–90% of both gB and pp65-specific cells carrying this receptor. Expression of FasL was detected on only a very small proportion of CD4+ T cells, with a median of 0.7% of CMV-specific cells and 0.48% of the global effector memory CD4+ pool expressing this marker (S3 Fig).
To investigate whether CMV-specific CD4+ T cells are indeed capable of killing target cells directly ex vivo, we isolated virus-specific cells from peripheral blood using tetramers and determined lysis of autologous or HLA-matched lymphoblastoid cell lines (LCLs) loaded with cognate peptide. CD4+ effector memory T cells (CCR7-CD45RA-) from two CMV-seronegative individuals were selected as controls. CMV-specific CD4+ T cells displayed remarkable cytotoxic capacity and this was seen for cells specific for all three peptide targets (Fig 5C). At an effector:target (E:T) ratio of 1:1 LLQ- and AGI-specific T cells lysed around 77% of target cells within 20 hours. DYS-specific T cells were able to kill 56% of target cells even at an E:T ratio of 0.5:1. No killing was observed by CD4+EM cells isolated from CMV seronegative individuals.
Previous studies have shown that CMV-specific CD4+T cells can express the co-stimulatory molecule NKG2D [36] which is almost always present on cytotoxic NK cells and CD8+ T cells. However these analyses relied on functional activation with viral antigen and we therefore examined NKG2D expression on unmanipulated cells through the use of HLA class II tetramers. Interestingly, NKG2D expression was negligible on the CD4+ EM T cell population in CMV seronegative individuals, with a median expression of only 0.66%, but was observed on 23% of CMV-specific CD4+ T cells (Fig 5D). Importantly, this proportion did not show any increase in relation to aging.
Relatively little is known about the pattern of expression of inhibitory markers such as PD-1 or Tim-3 on antigen-specific CD4+ T cells. Expression of PD-1 on virus-specific CD8+ T cells has been associated with functional impairment of immune responses against HIV or HCV, both chronic infections in which antigen load may remain high for prolonged periods of time [37,38]. However in healthy individuals PD-1hi CD8+ T cells usually demonstrate an effector memory phenotype and do not necessarily exhibit functional ‘exhaustion’ [39].
To get a better understanding of PD-1 expression on CD4+ T cell populations we initially analysed the memory phenotype of PD-1+ cells within the global CD4+ T cell subset and the CMV-specific T cell population. This showed that within the CD4+ T cell population 62% of PD-1+ cells carried an EM phenotype and 22% were CM cells (Fig 6A). Of the PD-1+ CMV-specific CD4+ T cells 94% displayed a CCR7-CD45RA- EM phenotype and virtually no cells had a CCR7+CD45RA-CM phenotype (Fig 6B).
We therefore examined the pattern of expression of PD-1 on CMV-specific CD4+ T cells and contrasted this with the pattern of staining on the CD4+EM T cell population. PD-1 expression was observed on a median 47% of CMV-specific CD4+ T cells and exhibited a remarkable range of expression, with cells from some donors showing hardly any evidence of PD-1 expression whereas virtually 100% of cells were positive in other individuals (Fig 6C). When PD-1 expression was evaluated in relation to peptide specificity it was clear that PD-1 expression was markedly reduced on DYS-specific CD4+ T cells, where it was observed on 29% of virus-specific cells, compared to 51% and 47% of AGI- or LLQ-specific T cells, respectively. Interestingly, PD-1 expression was very low on the CD4+EM T cell population and was observed on less than 10% of cells (Fig 6C). When analysing the median fluorescence intensity (MFI) of PD-1 on these cell populations we found the same trends within virus-specific T cells, and much lower MFI values were detected on CD4+EM T cells (Fig 6D). Of interest, Tim-3 expression was not detected on a significant proportion of any CD4+ T cells (S4 Fig).
It might be anticipated that the proportion of T cells expressing an inhibitory marker would increase with age. Therefore we examined both the percentage of PD-1+ cells and the MFI of PD-1 expression on CMV-specific T cells, and the CD4+EM T cell subset, in relation to age. The proportion of PD-1+ cells did not change in relation to the age of the donor (Fig 6E) however a non-significant trend was observed towards lower levels of PD-1 protein expression (MFI) on the surface of PD-1+ CMV-specific CD4+ T cells in older adults (Fig 6F).
In this study we have used HLA class II-peptide tetramers to identify and characterise CMV-specific CD4+ T cells, without the need for functional identification, for the first time. This allowed a comprehensive analysis of the resting phenotype and transcriptional profile of antigen-specific CD4+ T cells recognising glycoprotein B and pp65, the two viral proteins recognised most frequently by CMV-specific CD4+ T cells [21]. The CD4+ T cell response against cytomegalovirus is of interest for several reasons, related primarily to the unusual magnitude and phenotype of virus-specific cells, and also their potential role in the vascular damage that is reported in association with CMV infection in older people [40,41]. Here we identified CMV epitope-specific responses ranging from 0.01% up to a remarkable 24% of the total CD4+ repertoire. The median peptide-specific response was 0.45% of the CD4+ T cell repertoire which is considerably higher than CD4+ T cell responses against viruses such as influenza, hepatitis C Virus and Epstein-Barr Virus [30,42,43].
Interestingly, the frequency of cells identified by tetrameric staining was greater than that detected by expression of IFN-γ following peptide stimulation. As such, although this confirms the strong Th1 profile of CMV-specific CD4+ T cells, it also reveals that a considerable proportion of peptide-specific CD4+ T cells demonstrate an alternative cytokine profile. Indeed, TNF-α was the cytokine most frequently produced by CMV-specific CD4+ T cells. In addition, a proportion of virus-specific T cells was found to secrete IL-4 in some donors although the great majority of CMV-specific CD4+ T cells display a polarized Th1 phenotype [44]. IL-4 production by CMV-specific CD4+ T cells has previously been reported through ELISPOT analysis [45]. Of interest, although IL-4 production within pp65-specific T cells was usually observed with co-expression with IFN-γ, IL-4 production by gB-specific cells was seen in the absence of other cytokines. As such these observations indicate novel functional roles for IL-4 in the setting of chronic infection which warrant further investigation [46]. Functional assays using stimulation with whole viral lysate have demonstrated an increase in virus-specific CD4+ T cells with age [22] and although a similar trend was observed in our analysis it is likely that demonstration of such T cell ‘memory inflation’ will require analysis of responses directed against a wide range of different epitopes. Many younger donors already had high frequencies of epitope specific T cells in the periphery. This may be a reflection of the duration of viral carriage as some individuals acquire the virus very early in life whereas others only do so at a later time.
Almost all CMV-specific CD4+ T cells expressed an effector memory phenotype, although in contrast to CMV-specific CD8+ T cells, very few had reverted to re-expression of the CD45RA isoform rather than CD45RO. In vitro studies suggest that re-expression of CD45RA occurs during prolonged absence of antigenic stimulation [47] which may indicate that CMV-specific CD4+ T cells undergo antigen recognition in vivo more frequently than CD8+ subsets. Latent CMV is maintained within cells of the monocytic lineage and viral reactivation occurs during maturation into dendritic cells [48]. As such, CMV-specific CD4+ T cells are likely to undergo regular antigenic stimulation and this may serve to retain expression of the CD45RO isoform. Tetramer staining also allowed a detailed analysis of the membrane phenotype of CMV-specific CD4+ T cells and indicated that CD27 expression is lost early during differentiation and is then followed by loss of CD28 in the majority of the population. Loss of CD28 expression on CD4+ T cells is very unusual and, indeed, the CD4+CD28- phenotype is virtually unique to CMV-specific T cells [49] and indicates that alternative mechanisms of T cell co-stimulation become important in the CMV-specific immune response, potentially through molecules such as 4-1BB [50]. CD57 is a poorly characterised molecule, but is again highly specific for CMV-specific T cells and was found to be expressed reciprocally with CD28, suggesting that it may itself have a potential role in co-stimulation. NKG2D is one such potential alternative costimulatory molecule and can synergise with TCR-dependent activation of CMV-specific CD4+ T cells to enhance a range of effector functions [36,51,52]. Indeed, we found that NKG2D was expressed on 23% of CMV-specific CD4+ T cells, a remarkably high level for a molecule typically associated with expression on NK and CD8+ T cells. The high degree of differentiation seen for CMV-specific CD4+ T cells was largely independent of the frequency of epitope-specific T cells, indicating that alternative factors, such as the environment in which T cell priming occurs or the availability of antigen, may influence their phenotypic profile (reviewed in [53]). However the antigenic specificity of virus-specific CD4+ T cells did have a marked influence on the differentiation status such that CD4+ T cells specific for glycoprotein B-derived epitope DYS consistently displayed a more differentiated phenotype than pp65-specific cells. Glycoprotein B is highly unusual in that it has direct access to the HLA class II antigen-processing pathway of infected cells. This mechanism is likely to mediate frequent reactivation of gB-specific T cells and could explain the more ‘driven’ differentiation phenotype of CD4+ T cells specific for peptides from this protein [54,55].
A unique aspect of our study was the direct isolation of virus-specific CD4+ T cells without the need for antigenic stimulation, such that we were able to analyse their resting transcriptional profile. Even though differences to CD4+EM T cells of CMV seronegative individuals were generally modest, the most striking observations were seen in the expression of genes related to cytotoxic function and chemotaxis. Remarkably, the profile of cytolytic genes upregulated in CMV-specific CD4+ T cells closely corresponds to the pattern seen in CD8+ cytotoxic T cells (CTL). These included granzymes A, B and H, as well as perforin, granulysin and NKG7. Importantly, we were also able to determine that Fas ligand is not expressed by virus-specific CD4+ T cells, indicating that the delivery of mediators from cytotoxic granules is the dominant mechanism of target cell lysis. Remarkably, the use of tetrameric staining reveals that this profound cytotoxic potential of CMV-specific T cells is observed within resting cells in the bloodstream. Furthermore isolated virus-specific CD4+ T cells very efficiently kill antigen-loaded target cells directly ex vivo suggesting that they are primed for rapid target cell lysis in the event of an episode of viral reactivation.
CMV-specific CD4+ T cells also expressed a distinctive profile of chemokines and additional proteins that indicate an important role in chemotaxis and tissue migration. Indeed, one of the most fascinating features of CMV-specific CD4+ T cells is their high level of expression of CX3CR1, a chemokine receptor that binds CX3CL1 (fractalkine) and has already been identified as a specific marker for CMV-specific CD8+ T cells [11]. The CX3CR1/CX3CL1 axis plays an important role in both the adhesion and transmigration of lymphocytes to endothelial cells during inflammation [56,57]. Interestingly, endothelial cells are a principal target tissue for CMV infection and the expression of CX3CR1 on CMV-specific T cells may therefore act to localise adaptive immunity to sites of viral reactivation. This mechanism of endothelial-targeting may be highly relevant to the potential development of endothelial immunopathology mediated by CMV-specific cytotoxic T cells [58,59]. Indeed a close link has been observed between augmented CMV-specific immune responses and a range of inflammatory conditions (reviewed in [60]) and the proportion of CD4+CD28- T cells has been shown to correlate directly with cardiovascular mortality in some studies [61].
CMV-specific CD4+ T cells also exhibit a very similar chemokine secretion profile to that of virus-specific CD8+ cells, with production of the inflammatory mediators CCL3 (MIP-1α), CCL4 (MIP-1β) and CCL5 (RANTES), all of which would support recruitment of innate immune cells such as macrophages and NK cells. The pattern is associated with high level expression of IFN-γ and TNF-α but very little production of IL-2 [23,62] and is again typical of a differentiated Th1 profile.
An additional interesting observation from the microarray data was the detection of high levels of ADRB2 mRNA in glycoprotein B-specific T cells. ADRB2 encodes the β2-adrenergic receptor which allows cells to respond to systemic production of epinephrine and forms an important link between the sympathetic nervous system and the adaptive immune response [63,64]. The CMV genome contains promoter elements that can bind epinephrine leading to increased viral gene transcription [65] and physiological stress is established as a risk factor for reactivation of many herpesviruses. Interestingly, CX3CR1+ T cells have been shown to be the major T cell subset released into the circulation following administration of epinephrine, in a mechanism partially mediated by reversal of their resting adherence to endothelium [66]. This may provide a mechanism of rapid mobilisation for CMV-specific T cells in response to stress induced viral reactivation, but this will need further investigation.
T cell activation and effector function are finely tuned events and CD4+ T cells show an extremely high sensitivity for their cognate antigen [67]. Moreover lytic synapse formation has a very low threshold for activation [68,69]. A proportion of CD27-CD28- CMV-specific CD4+ T cells has previously been described to exhibit regulatory function in vitro [70] but FoxP3+ cells were not detectable in this study and antigen-specific T cells were not shown to produce IL-10. This suggests that regulatory mechanisms are likely to be focussed at the level of the effector T cell and here PD-1/PD-L1 interactions play a key role as a negative feedback mechanism controlling TCR-dependent effector function of T cells [71]. High levels of PD-1 have been detected on CD8+ effector memory T cells [39] and similarly we find that PD-1 expression on CD4+ T cells is also largely confined to this subset. PD-1 expression was seen on nearly half of all CMV-specific CD4+ T cells and could serve to limit T cell activation. As such it is noteworthy that the level of membrane PD-1 expression (MFI) on virus-specific CD4+ cells tends to decrease during aging and may therefore lower the activation threshold of T cells at a time when their cytotoxic potential is actually increasing. Together these factors may influence virus-host balance across the life course and may contribute towards immunopathology.
In summary our data reveal that people who are chronically infected with cytomegalovirus, which represents the great majority of the global population, harbour substantial populations of virus-specific CD4+ T cells within their bloodstream. These highly differentiated cells display a strongly cytotoxic phenotype, may be targeted to activated endothelium and have the potential to respond to physiological ‘stress’ through detection of epinephrine (Fig 7). In addition, this cytotoxic profile increases further with age whilst the level of inhibitory PD-1 on the surface declines. These findings reveal the exceptional evolutionary adaptation of the CD4+ T cell response towards the control of CMV. In addition they shed light on the potential mechanisms whereby CMV infection may serve to mediate tissue damage, most particularly vascular disease, and indicate a range of potential novel immunotherapeutic targets.
The study was approved by the West Midlands (Black Country) Research Ethics Committee (REC 07/Q2702/24) and all donors gave written informed consent before participation. The donor cohort included samples from laboratory personnel, the blood transfusion service and healthy older adults recruited from the ‘Birmingham 1000 Elders group’ (REC 2002/073).
A total number of 73 CMV-seropositive donors, aged between 24–88 years, with appropriate HLA-genotype were included in the study. PBMCs were isolated from 50ml heparinized blood by density gradient centrifugation using Lympholyte-H cell separation media (Cedarlane) and aliquots of 10x106 cells were cryopreserved in RPMI1640 (Sigma-Aldrich) containing 20% fetal calf serum (FCS) and 10% DMSO and stored in liquid nitrogen until use.
To identify donors with the appropriate HLA-genotype, genomic DNA was isolated from PBMCs using the GenElute Blood Genomic DNA Kit (Sigma-Aldrich) according to manufacturer’s instructions. Typing for HLA class II alleles was performed by PCR technique as described previously [72].
Phycoerythrin (PE)-conjugated custom-made HLA class II tetrameric complexes were purchased from the Benaroya Research Institute at Virginia Mason (Seattle, Washington). Three tetramer complexes were used in this study: They were comprised of the CMV gB-derived epitope DYSNTHSTRYV in the context of HLA-DRB1*07:01 (DR7) [73], or pp65-derived epitopes AGILARNNLVPMVATV within HLA-DRB3*02:02 (DR52b) [74] and LLQTGIHVRVSQPSL within HLA-DQB1*06:02 (DQ6) [75].
Initially, the specificity of each tetramer was confirmed by screening against CD4+ T cell clones recognizing the tetramer’s cognate HLA class II-peptide complex or against PBMCs from a CMV-seronegative donor expressing the appropriate HLA-allele. Optimal tetramer concentration and staining times were distinguished at the outset and constant conditions used throughout the study.
To identify cytokine-producing T cells following activation, 1x106 PBMC were resuspended in RPMI1640 (Sigma-Aldrich) supplemented with 10% FCS and 1% Penicillin/Streptomycin and stimulated with peptide (5μg/ml final concentration) overnight at 37°C with 5% CO2 in the presence of BrefeldinA (10μg/ml final concentration; Sigma-Aldrich). Unstimulated cells and cells stimulated with Staphylococcus Enterotoxin B (final concentration 0.2μg/ml; Sigma-Aldrich) served as controls. Following overnight incubation the cells were stained with LIVE/Dead fixable violet or aqua stain (Invitrogen) as described above, T-cells were then identified by staining with anti-CD4-PE (RPA-T4, BD Bioscience) or anti-CD3 AF700 (SK7, Biolegend) and anti-CD4 APC-Cy7 and B cells excluded by staining with anti-CD19 pacific blue (H1B19, eBioscience; dump channel). Fixing was carried out with 4% paraformaldehyde (in PBS; Sigma-Aldrich) for 15 min at RT before permeabilising with 0.5% Saponin (in PBS; Sigma-Aldrich) for 5 min. Intracellular cytokines were then stained with the following antibodies: anti-IFN-γ FITC antibody (25723.11, BD Bioscience) or anti-IFN-γ PE-Dazzle (4S.B3), anti-TNF-α PE-Cy7 (Mab11), anti-IL-10 PE (JES3-9D7), anti-IL-4 BV421(MP4-25D2; all Biolegend), anti-IL17A (eBio64DEC17, eBioscience) and anti-MIP-1β (24006, R&D) and followed by a final wash in staining buffer. Acquisition was carried out on an LSR II flow cytometer and DIVA software (BD Bioscience) collecting 300,000 live lymphocytes and data analysed using FlowJo software version 7.6.5 (Tree Star). For the analysis single, viable, CD19- lymphocytes were gated before identification of cytokine-producing CD4+ T cells. For the multi-cytokine panel Boolean gating was used to determine all possible combinations and further analysis was carried out using SPICE version 5.3.
To analyse the transcriptional profile of CMV-specific T cells we sorted DYS- and LLQ-specific CD4+ T cells from CMV-seropositive healthy donors and for comparison CD4+ T cell subsets from CMV-seronegative healthy individuals. For this PBMCs were isolated from 120 mL of heparinised blood by density gradient centrifugation and the CD4+ T cell population was enriched by negative selection (StemCell Technologies) according to manufacturer’s instructions. CD4+ T cells from CMV-seropositive donors were then stained with LIVE/Dead fixable far red stain (Invitrogen), washed and re-suspended in 600 μL of human serum prior to incubation with PE-conjugated HLA class II tetramer (HLA-DR7:DYS or HLA-DQ6:LLQ) for 1 h at 37°C and 5% CO2. After washing, cells were stained with anti-CD4 PE-CF594 (RPA-T4, BD) and re-suspended in RPMI + 10% FCS. From the single, viable lymphocyte population CD4+tetramer+ cells were then sorted on a MoFlow Cell Sorter (Beckman Coulter) consistently reaching a purity of 98–99%. Following CD4-enrichment cells of CMV-seronegative donors were stained with LIVE/Dead fixable far red stain (Invitrogen), anti-CD4 PE-CF594 (RPA-T4, BD Biosciences), anti-CCR7 FITC (150503, R&D) and anti-CD45RA PE (HI100, eBioscience). Based on their expression profile effector memory cells (CCR7-CD45RA-) were then sorted to high purity.
Total RNA of the sorted cells was extracted using an RNeasy Plus Mini kit (Qiagen) according to manufacturer’s instruction. The RNA integrity was checked and it was subsequently labelled before hybridisation to Agilent human gene expression 8x60k microarrays (G4858A) according to manufacturer’s instructions following the standard Agilent Low Input Quick Amp labelling protocol. Due to low mRNA yield, CD4+CD45RO+ T cells sorted from a CMV-seronegative donor served as the reference sample on the two-colour microarray slide. These were not taken into account for the analysis. The Microarrays were carried out by the Functional Genomics, Proteomics and Metabolomics Facility at the School of Biosciences, University of Birmingham.
Microarray data was analysed with the R Limma Package (Bioconductor) [78–80]. Normalisation was performed with the Loess (intra-array) and Aquantile (inter-array) methods. An adjusted p-value (Benjamini and Hochberg's method) of 0.05 and below was taken as significant for differences in gene expression or otherwise a 2-fold change in expression levels. Further analysis of output data was completed in Excel (Microsoft Corp). Hierarchical clustering was performed on the MultiExperiment viewer version 4.9 (MeV, [77]. Functional enrichment analysis was completed using DAVID version 6.7 [81,82].
Microarray data are available in the ArrayExpress database (www.ebi.ac.uk/arrayexpress) under accession number E-MTAB-4510.
cDNA generated from the same samples used for the microarray analysis was used to analyse transcription of a selected number of genes that were differentially expressed between CD4+EM cells and CMV-specific CD4+ T cells. TaqMan Gene Expression Assays for CCL3 (Hs04194942_s1), CCL4 (Hs04421399_gH), CCL5 (Hs00982282_m1), GZMB (Hs00188051_m1), PRF1 (Hs00169473_m1), ADRB2 (Hs00240532_s1), CX3CR1 (Hs04187059_m1) and GAPDH (4310884E) were bought from Thermo Fisher Scientific. Specific target amplification was carried out on the cDNA using 2× TaqMan PreAmp Master Mix (Life Technologies) and 0.2× primer mix (20× TaqMan assays diluted in water). Reactions were subjected to 95°C for 10 min, followed by 12 cycles of 95°C for 15 s and 60°C for 4 min. These pre-amplified samples were then diluted 1:5 with water prior to Q-PCR analysis using the TaqMan Gene Expression Assays. The relative transcription was calculated comparing with average transcription level of the three control CD4+EM T cells and GAPDH assays served for normalization. Assays were performed in duplicate for two to three donors each (EM, LLQ and DYS). Transcription levels in CD4+EM T cells were set to 1.
To analyse the cytotoxic capacity of CMV-specific CD4+ T cells directly ex vivo, CD4+TM+ cells were separated as described above and then co-cultured over night with CFSE labelled (0.5μM; Invitrogen) autologous or HLA-matched LCLs which were loaded with the cognate peptide. CFSE-labelled LCL alone served as control. In addition CD4+ CCR7-CD45RA- (EM) cells of two CMV seronegative individuals were sorted and served as effector cells. Killing of target cells was assessed on a BD Accuri flow cytometer (BD Biosciences) by quantifying live CFSE-labelled LCL using counterstaining with Propidium Iodide (PI) to identify dead cells. All conditions were carried out in triplicate. For the analysis initially the T cell population and the LCL population were gated on, using FSC and SSC scatter plots, to determine the true ratio of effector to target cells. Then the number of live LCLs was determined using CFSE and PI and percent killing of target cells was calculated.
Statistical analysis of flow cytometry data was performed using GraphPad Prism5. The non-parametric Mann-Whitney U-test was applied for comparison of two groups, and the Kruskal-Wallis test (with Dunn’s multiple comparison) for comparison of more than two groups of continuous measurements. To analyse the strength of associations between variables Spearman’s rank test was used. A p-value of less than 0.05 was considered statistically significant.
|
10.1371/journal.pntd.0005701 | The interplay of climate, intervention and imported cases as determinants of the 2014 dengue outbreak in Guangzhou | Dengue is a fast spreading mosquito-borne disease that affects more than half of the population worldwide. An unprecedented outbreak happened in Guangzhou, China in 2014, which contributed 52 percent of all dengue cases that occurred in mainland China between 1990 and 2015. Our previous analysis, based on a deterministic model, concluded that the early timing of the first imported case that triggered local transmission and the excessive rainfall thereafter were the most important determinants of the large final epidemic size in 2014. However, the deterministic model did not allow us to explore the driving force of the early local transmission. Here, we expand the model to include stochastic elements and calculate the successful invasion rate of cases that entered Guangzhou at different times under different climate and intervention scenarios. The conclusion is that the higher number of imported cases in May and June was responsible for the early outbreak instead of climate. Although the excessive rainfall in 2014 did increase the success rate, this effect was offset by the low initial water level caused by interventions in late 2013. The success rate is strongly dependent on mosquito abundance during the recovery period of the imported case, since the first step of a successful invasion is infecting at least one local mosquito. The average final epidemic size of successful invasion decreases exponentially with introduction time, which means if an imported case in early summer initiates the infection process, the final number infected can be extremely large. Therefore, dengue outbreaks occurring in Thailand, Singapore, Malaysia and Vietnam in early summer merit greater attention, since the travel volumes between Guangzhou and these countries are large. As the climate changes, destroying mosquito breeding sites in Guangzhou can mitigate the detrimental effects of the probable increase in rainfall in spring and summer.
| An unprecedented dengue outbreak occurred in Guangzhou, 2014, with 38,036 reported cases in contrast to 73,179 cases in all of mainland China from 1990 to 2015. In an earlier analysis using a deterministic model, we concluded the early timing of local transmission to be the most important determinant of this outbreak. Here we use a stochastic model to explore the reasons why the outbreak happened earlier in 2014. Our results identified the higher number of imported cases in May and June to be the most probable explanation. Based on the investigation of the determinants of success rate and final epidemic size, this work provides suggestions for reducing dengue outbreak potential and epidemic size in the future. More attention should be paid to imported case detection and vector control measures in early summer, because this is the time when successful invasion can result in high incidence of infection and the success rate of each imported case begins to rise. Destroying mosquito breeding sites can reduce the maximum water level of the system and attenuate the role played by climate. In addition, interventions within 10 days after the introduction of imported cases is still effective in preventing further transmission.
| As the most important mosquito-borne disease globally, the incidence of dengue has increased 30-fold in the past 50 years [1]. In 2013 alone, dengue was responsible for 576,900 years of premature life lost and 1.14 million disability-adjusted life-years globally [2]. However, dengue was not a serious concern in mainland China before 2014. There were only 73,179 cases from 1990 to 2015, among which 47,056 (64.3%) occurred in 2014, and 80.8 percent of these cases were contributed by Guangzhou [3]. Because of its warm and wet climate, as well as the close ties with dengue endemic Southeast Asian countries, Guangzhou has a high risk of local dengue transmission. The introduction of dengue virus (DENV) by imported cases is required annually for local epidemics in Guangzhou, since the dengue virus cannot disseminate to the salivary glands of Aedes albopictus when temperature is below 18°C [4] and adult mosquitoes rarely survive the winter.
Many studies have attempted to explain the unprecedented dengue outbreak in Guangzhou, 2014. Possible explanations include more imported cases, higher mosquito abundance due to a more favorable climate, higher larval development and adult survival rates caused by urbanization, and lack of diagnostic experience [5, 6]. The result of our earlier analysis using a deterministic mathematical model identified the early timing of local transmission and climate as the most important determinants of the large final epidemic size (FES) [7], but the model could not explain why local transmission occurred earlier that year. Because a stochastic model can produce different results every run under the same conditions, an imported case can lead to local transmission in some runs but not so in others. Therefore, a stochastic model can shed more light on the outbreak probability and can provide greater insight into the various explanations of earlier local transmission, such as a more favorable climate for mosquito growth and more imported cases. Clearly, understanding the reasons underlying the early and large outbreak can help to prevent dengue outbreaks in the future, and reduce potential economic and health impacts.
Stochastic models have been widely used to study the invasion of a disease into non-endemic areas, such as dengue virus transmitted by Ae. aegypti mosquitoes [8], dengue virus in Madeira [9], and Chikungunya virus in the United States [10]. However, none of these models considered the importance of water availability in restricting the environmental carrying capacity and density-dependent rates.
In this paper, we simplify our deterministic model by leaving out vertical transmission from infected adults to eggs since it was found to be unimportant in the earlier analysis [7], and then extend it to incorporate stochastic effects. A hybrid deterministic/stochastic algorithm originally used in simulating chemical reactions [11–13] is used to simulate the dengue transmission dynamics. This algorithm can take advantage of the low computational burden of deterministic models while still incorporating stochastic effects. Six different scenarios are constructed to explore the determinants of early outbreak in 2014 and the effectiveness of interventions (insecticide spraying and pooled water removal) in reducing dengue transmission risk. We also address the determinants of the outbreak size and FES, as well as the implications for prediction and prevention.
The Ethics Committee of the Guangzhou Center for Disease Control and Prevention reviewed and approved this study. All patient data was made anonymous prior to the analysis.
Guangzhou, with a population of 13.1 million at the end of 2014 [14], is the largest city in South China. The climate is characterized by warm and humid summers and mild and dry winters. The climate data downloaded from China Meteorological Data Sharing Service System (CMDSSS) indicates that the annual mean temperature from 1985 to 2014 was 22.4°C. January had the lowest average temperature of 13.7°C, while July and August had the highest at 28.9 and 28.7°C, respectively (Fig 1A). The average annual accumulated precipitation was 1,821 mm, of which 1,490 mm (81.9%) occurred between April and September (Fig 1B). The warm and wet summer is favorable for the growth of Ae. albopictus, the sole vector of dengue transmission in Guangzhou [15].
Because low temperature is unable to support virus replication in Ae. albopictus in the winter [4], the occurrence of dengue epidemics in Guangzhou depends highly on imported cases from surrounding endemic countries, such as Thailand, Singapore, Malaysia, Philippines, Vietnam, Cambodia, and Indonesia. Serving as a transportation hub for these countries increases the outbreak risk of Guangzhou further (Fig 1C).
We collected daily reported dengue case numbers in 2013 and 2014 from the Guangzhou Center for Disease Control and Prevention (Guangzhou CDC). These data are also available from the Health and Family Planning Commission of Guangdong Province (http://www.gdwst.gov.cn/) in the transmission season. Both passive and active surveillance systems are used in China. Health-care providers are required to diagnose the disease according to the National Diagnostic Criteria for Dengue Fever (WS216-2008) and then report the new cases to the web-based National Notifiable Infectious Disease Reporting Information System within 24 hours. After the confirmation of a case, active detection is initiated and conducted by various means, such as interviewing residents in the same community, checking employee attendance records in the same work place, and checking the outpatient medical records in nearby health facilities [16]. Cases are categorized further into imported and indigenous cases based on the travel and mosquito biting history. Imported cases are those who have traveled to dengue endemic regions and been bitten by mosquitoes less than 15 days before the symptom onset [6]. This dataset was used to calibrate the deterministic model.
Entomological surveillance data, including the Breteau index (BI) and mosquito ovitrap index (MOI) for 2013 and 2014, were also obtained from Guangzhou CDC. BI is a representation of larva abundance, which is defined as the number of water containers infested with larva per 100 houses inspected, while MOI is a proxy for adult abundance defined as the percentage of positive ovitraps in a specific area. The daily temperature, precipitation and evaporation data for Guangzhou from 1985 to 2014 were downloaded from CMDSSS (http://data.cma.cn/). The 30-yr daily average climatic factors were calculated to represent a typical year and were used in Scenario 2014 to Avg described later. Climate datasets were used as inputs to the model.
Human population data were extracted from the Guangdong Statistical Yearbook [17–19] to estimate the birth and death rates, which were treated as known parameters in the deterministic model. In addition, the initial value for the susceptible human population in the model was set to the population at the end of 2011, because we assumed that all residents were susceptible to dengue virus infection since no big outbreaks had occurred in Guangzhou from 1978.
To investigate the tourist exchange between Guangzhou and dengue endemic countries, we also extracted the number of direct air travelers between Guangzhou and different countries from on-flight origin and destination (OFOD) data provided by International Civil Aviation Organization (ICAO) (https://www4.icao.int/NewDataPlus/). Since this dataset leaves out indirect air travelers or travelers by sea or by land, we also collected the number of foreign tourists staying overnight in Guangzhou and the number of tourists who traveled from Guangzhou with tour groups from the Tourism Administration of Guangzhou (http://www.gzly.gov.cn/index.html) as a complement (S1 Table).
To simulate the population dynamics and dengue virus infection status of both Ae. albopictus and humans, we constructed a compartment model shown in Fig 2 which is based on Ross-Macdonald model [20, 21]. Several modifications were made to the basic model to emphasis the influence of temperature and precipitation, such as modelling water level and aquatic stages explicitly, and using temperature- and density-dependent rates. The model has been described in detail elsewhere [7]. Because vertical transmission was found to be unimportant in the 2014 outbreak [7], here we removed the vertical infected egg, larva, pupa and emerging adult compartments to simplify the model structure and to reduce computational complexity. Immature life stages eggs (E), larvae (L), and pupae (P) were included to incorporate the effects of water availability and temperature on vector abundance. Emerging adults (Aeu) were also incorporated as a separate compartment because they do not bite. Then, according to infection status, the biting mosquito population was divided further into susceptible (As), exposed (Ae) and infectious (Ai) adults, and the human population was divided into susceptible (Hs), exposed (He), infectious (Hi) and recovered (Hr) individuals.
Transition rates in the model depend on temperature and water availability. Temperature can influence the development rate from eggs to larvae, larvae to pupae, and pupae to emerging adults (Events 2, 4 and 6 in Fig 2), the mortality rate of larvae, pupae and adults (Events 3, 5, 7, 9, 13 and 16), the biting rate of adults (Events 12 and 19), the extrinsic incubation period (EIP) of dengue virus (Event 14), and the average number of eggs and duration of each gonotrophic cycle (Events 10, 11 and 15) [23–25]. Therefore, temperature-dependent functions whose form is based on [23, 26] were used to describe these rates, and the coefficients in these functions were estimated from experiments conducted on Ae. albopictus strains from South China [25, 27]. Furthermore, the development rate from eggs to larvae, larvae to pupae, and the mortality rate of larvae also depend on the larval density (Events 2–4), which further depend on the current available water level determined by rainfall, evaporation, the minimum and maximum water level (ωmin and ωmax) of the system, and interventions (eg. emptying water containers) [28]. Minimum water level represents standing waters that are difficult to evaporate, such as the water in containers with lids or in shaded areas, while the maximum water level is the upper limit of the water in the system beyond which water will overflow. Since these two parameters cannot be obtained from the literature or surveys, they were estimated by the deterministic model. More information for the model structure can be found in the S1 File.
As in the previous paper [7], we still include the arrival of imported cases, intervention, the spillover effect and diapausing eggs in the current model. The model was run from 2012 to 2014, though we only fit the model to the observed daily reported cases of 2013 and 2014. Year 2012 was included to get a reliable mosquito population for 2013 and 2014, because the initial value for eggs can only affect the mosquito abundance of the first simulated year, but not in later years [29].
There are 19 parameters in our model and their values are highly uncertain. In addition, we are more confident about the pattern of the daily reported new cases and mosquito surveillance data than the exact value of these data on any given day. Therefore, instead of employing Markov Chain Monte Carlo (MCMC) fitting procedures common in epidemiologic mathematical modelling [9, 30], we used a parameter estimation strategy originally called regional sensitivity analysis (RSA) [31] to fit the model to the pattern of the surveillance data. The detailed process was described elsewhere [7], and only a brief description is provided here.
To calibrate the deterministic model, a broad range for each parameter was specified based on the literature or our best knowledge. Classification criteria were defined to describe whether the model output mimicked the observed timing and approximate number of reported cases at the onset, peak and end phase of the epidemics (see details for the criteria in S1 File). A Monte Carlo procedure was then followed in which random samples were drawn from the specified parameter space and used to run the mode. The output of each realization was classified into one of two subgroups “pass” or “fail” according to whether the result met all of the classification criteria. After accumulating a sufficient number of parameter vectors in the “pass” group, the characteristics of the cumulative univariate distributions were used to trim the range for each parameter to get a smaller parameter subspace with a higher passing rate.
A total of 5 trimming cycles were conducted until the pass and fail univariate distributions offered no further guidance on identifying subspaces with higher pass rates. The results are presented in S1 File. Unlike MCMC fitting, which gives only one parameter set with a confidence interval, RSA gives multiple parameter sets matching the passing criteria. Here we obtained 5,320 parameter sets out of 100,000 runs after the last trimming cycle (Cycle 5) in contrast with the 83 passing sets from 800,000 trials initially (Cycle 1). Fig 3A shows the trajectories and the envelope of the number of daily new cases produced by the 5,320 sets in Cycle 5. As expected, these passing parameter sets produce trajectories which mimic the pattern of the field data successfully, except that the peak number of cases in 2014 is a little lower and occurs earlier than observed. Although the observed initial exponential growth rate of 2014 is higher than that of 2013, the model produces very similar values, presumably because we assume that the intrinsic incubation period, recovery period, transmission probability from vector to human and from human to vector, and reporting rate remain the same from year to year. This lack of fit suggests difference in dengue virus virulence or reporting rate between these 2 years, which may need further investigation in the future.
The number of larvae and adult mosquitoes output by the model shows the same patterns as the mosquito surveillance data for both BI and MOI. Since these entomological datasets were not used in model calibration, they support the validity of our model and its parameterization.
In transitioning from a purely deterministic model to one suitable for exploring stochastic effects, we must take account of the earlier finding that there are extended regions of the parameter space that led to good fits to the calibration data for the deterministic model. Since thousands of simulations were needed for each parameter set, it is impractical to use all the 5,320 passing sets of Cycle 5. As a result, 100 parameter sets were sampled randomly to represent the space of good fitting vectors (S2 Table). To reduce the computational costs further, a simulation procedure developed specifically for hybrid deterministic/stochastic models was utilized.
The hybrid deterministic/stochastic algorithm can add stochasticity to the dengue transmission model and obtain estimates of the success invasion rate of one imported case and the distribution of FES. This method has been widely used in simulating chemical reactions. It can improve the calculation efficiency significantly while still producing a similar pattern as that achieved by using the exact method [11–13, 32, 33]. The transitions between different compartments here, analogous to the chemical reactions in those studies, were partitioned into “fast” and “slow” subsets according to their transition rates. Fast events happen frequently and have lower level of stochasticity, thus they are simulated by the deterministic model using ordinary differential equations (ODEs) (Events 1–11, 15, 17 and 18 in Fig 2), while slow events happen infrequently so they must be modeled stochastically (Events 12–14, 16, and 19–24). Since there were only small numbers of exposed and infected mosquitoes (Ae and Ai), and of exposed, infected or recovered humans (He, Hi and Hr) at the beginning of the outbreak, the mortality rates of these compartments (Events 13, 16, 20, 22, and 24) and the transition rates from these compartments (Events 14, 21, and 23) were relatively low. Hence these events are considered as slow events. In addition, the transition rate of Event 12 (bαhvHiAs/N), which is the mosquito infection via human contagion, is also low, because b is the temperature-dependent biting rate which ranges from 0 to 1; αhv is the transmission probability of dengue virus from infected human to susceptible adult mosquito which also ranges from 0 to 1; Hi is the number of infected humans which is a small number at the beginning of the outbreak; and As/N is the ratio of mosquitoes to humans, which is also small when compared with mosquito abundance or the human population. Therefore, Events 12 and 19 are slow events, for the same reason.
For the stochastic simulation of the slow events, instead of the widely utilized Gillespie’s stochastic simulation algorithm (SSA) in the chemical studies, here we used the adaptive tau-leap algorithm [34–36], which is more commonly used in epidemiological models to reduce the computational burden [10, 34, 35]. To test the validity of the tau-leap algorithm, Gillespie’s SSA was also tried for two randomly chosen parameter sets. The Kolmogorov–Smirnov two-sample test for both parameter sets indicated no significant difference between the FES distributions produced by different algorithms. Therefore, only the tau-leaping algorithm was applied in the following analysis. The time step tau was initially set to be 1/5 day, which is appropriate for a population of the order of millions [37]. If high numbers of events occurred in this time interval which led to negative population sizes, a new value of tau was adopted as tau/2 to shorten the time interval and avoid negative population sizes. The details of the implementation of the model are shown in S1 File.
Dengue transmission risk is mainly determined by the mosquito-to-human ratio, temperature, and the immune status of human population [38]. Since the human population and its immune status do not vary much from year to year, the only differences between 2014 and other years were mosquito abundance and temperature. Mosquito abundance further depends on the availability of breeding sites, represented by water level in the model, and human interventions. The most common interventions in Guangzhou are insecticide spraying and emptying water containers, which can reduce the abundance of adults or aquatic stage of mosquitoes immediately. Emptying water containers can also affect the abundance through water availability and the density-dependent rate.
Six different scenarios were designed here to investigate the role of climate and human interventions in determining the potential and FES of the dengue outbreak in 2014 (Table 1). The first scenario used the observed climate data of 2014 and serves as the baseline for comparison (Scenario 2014). Then, to study the role of climate only, we replaced the climate files of 2014 in Scenario 2014 with those of 2013 (Scenario 2014 to 2013) and the 30-yr average (Scenario 2014 to Avg). The 30-yr average was used here to represent the climate in a typical year. In addition to climate, year 2014 also differed from year 2013 in the initial water level, which was determined by the climate and interventions in the previous year. The initial water level in 2014 is lower than 2013, because more water was removed in 2013 than in 2012 by more frequent interventions. The number of reported dengue cases in Guangzhou increased from 139 in 2012, to 1,249 in 2013, and then to 37,341 in 2014, which led to increasing intervention frequency over these years. Therefore, we compared the results of Scenario 2013 and Scenario 2014 to determine the combined effect of climate and intervention. Moreover, aiming to understand the outbreak potential under only natural conditions, we also designed two additional scenarios without interventions (Scenario 2014 w/o intervention and Scenario 2013 w/o intervention). The previous deterministic model [7] suggested that the interventions can reduce the FES effectively, so here these two scenarios were included to estimate the quantitative effects of intervention in reducing dengue transmission risk.
Turning to the role of the timing of imported cases, the introduction date was varied between March 21st and November 26th (the mosquito growth season) at 10-day intervals (a total of 25 different introduction dates). As with the deterministic model, the stochastic model was also run from Jan 1st, 2012 to December 31st, 2014. Five hundred iterations were run for each combination of scenario and introduction date. The FES of the last simulated year was recorded for each iteration. Therefore, a total of 7.5 million simulations were conducted (100 sample sets, 25 introduction dates, 6 scenarios, and 500 repetitions).
The successful invasion rate of an imported case was defined as the proportion of repetitions with a FES greater than 0. The average final outbreak size only accounts for successful invasions, and can be calculated as the mean of FESs greater than 0 in 500 repetitions, while the average final epidemic size was the mean FES of all invasions.
R 3.2.3 [39] was used for simulation, data analysis and visualization. The package deSolve was used for solving the ODEs [40], Rcpp, foreach, and doSNOW for parallel computing to increase the computational speed [41, 42], and ggplot2 for data visualization [43].
There are three possible explanations for the early timing of the local transmission in 2014, a higher success rate for a single imported case, more imported cases in early summer, or by chance alone. These possibilities are explored in this section, as well as the distribution and the mean of FES.
In order to determine whether the early outbreak in 2014 was caused by a particularly favorable climate for mosquito growth, we held all the other conditions constant, such as initial water level and the timing of interventions, but only changed the temperature, precipitation and evaporation of 2014 to those of 2013 and then to the 30-yr average (Scenario 2014 to 2013 and 2014 to Avg). Fig 4A shows that if an imported case is introduced to the system on the same day in early summer (the shaded rectangle), Scenario 2014 has the highest success rate. This implies that the climate in 2014 was more favorable for an early outbreak. For all the three scenarios, the success rate increases in late spring and early summer, remains at a high level from May to July, and then decreases until it reaches and stays at a low level in October. The biggest difference in the success rate between these scenarios occurs between May and July, the same period as the peak of success rate.
Fig 4B shows the comparison between two scenarios with observed climate and intervention (Scenario 2014 and 2013) and two scenarios with observed climate, but without intervention (Scenario 2014 w/o intervention and 2013 w/o intervention). The result for Scenario 2014 is depicted here again as a baseline for comparison. Although the success rate of Scenario 2014 is higher than that of Scenario 2014 to 2013 and 2014 to Avg when only the climate is considered (Fig 4A), it is lower than that of Scenario 2013 when the combined effect of lower initial water level and favorable climate is considered. Therefore, though the climate, especially adequate rainfall, in the early summer of 2014 is more favorable for mosquito growth and dengue transmission, the lower initial water level of 2014 caused by emptying water containers in late 2013 appears to have reduced the risk to a level even lower than that of 2013. In other words, the success rate of a single imported case in causing local transmission was not significantly higher in the time window of interest in 2014 and suggests other explanations, such as more imported cases in that time period or by chance alone.
The difference between Scenario 2014 and 2014 w/o intervention, as well as between Scenario 2013 and 2013 w/o intervention shows the effectiveness of intervention in reducing dengue transmission probability (Fig 4B). Similar to the three scenarios shown in Fig 4A, the success rates of Scenario 2013, 2014 w/o intervention, and Scenario 2013 w/o intervention also increase in the early summer until May. However, the duration of the peak differs. The success rate of Scenario 2014 stays at the peak until July, Scenario 2013 until September, while the two scenarios without intervention continue until October. The intervention started on September 27th in 2013 and July 25th in 2014, so they have no influence on the success rate in the beginning of these years. The success rates are the same for Scenario 2014 and 2014 w/o intervention, as well as for Scenario 2013 and 2013 w/o intervention in the late spring and early summer. However, the success rate drops 20 days before the first intervention, which indicates that the intervention can reduce the transmission probability of a case imported up to 20 days ago. This could possibly due to the death of exposed or infected mosquitoes in the intervention or reduced virus transmission from imported or secondary cases to mosquitoes caused by mosquito population loss.
Besides the success rate of one imported case on any given date, the outbreak probability, which is the probability that at least one indigenous case occurs in a particular year, also depends on the number of imported cases. It is clear that the outbreak probability will increase when there are more imported cases, although the increase may be quite small depending on the timing pattern. The outbreak probability can be calculated from the success rate of all imported cases in that year, once the timing of each imported case is known. Fig 5 illustrates the number of identified imported cases into Guangzhou from other countries or Chinese cities in 2013, 2014, and the mean, median and range of 2001 to 2014. Since the detection and reporting rate of imported cases are likely to increase after August, 2014, when attention began to be paid to the serious dengue outbreak, only data from January to August are shown here. The number of imported cases known to have entered Guangzhou in May and June, 2014 is four times higher than the average or median of that in 2001 to 2014, which may be the main determinant of the early local dengue virus transmission in 2014.
A different FES can be produced by stochastic effects in each simulation even with the same input. Fig 6 shows the histograms of the FES + 1 under different introduction dates when using the combination of sample set 1 and Scenario 2014 as an example. The FES for an imported case arriving May 11th, 2014, which is around the estimated timing of the first successful invasion in 2014, can range from 0 to 132,457 with a mean of 22,790 and a standard deviation of 25,219. These histograms also show that the FES of successful invasions decreases with the delay of the imported cases. In other words, though the success rate of an early imported case is low, the FES could be extremely high once it succeeds.
Average final outbreak size is the mean FES of successful invasions, while average FES is the mean of all simulations. Since the average FES equals the product of the outbreak probability and the final outbreak size, and outbreak probability has been discussed above, here we first look into the average final outbreak size and then the FES. Because all scenarios show the same pattern, and while the results under the last three scenarios are several magnitudes higher than the first three scenarios, to make the figure readable, only the results of the first three scenarios were shown here.
Fig 7A shows that the average final outbreak size decreases exponentially with introduction date, and the decrease rate is affected by climate. Scenario 2014 has the highest average final outbreak size, and Scenario 2014 to Avg has the lowest. According to Fig 4A, the mosquito abundance and success rate is higher in Scenario 2014, so each case can infect others more easily under this scenario, which will then lead to a higher final outbreak size. Unlike the monotonically decreasing average final outbreak size, the average FES peaks in mid-April (Fig 7B), similar to the conclusion from the deterministic model [7]. In addition, the average FES given by the stochastic model is also comparable to that given by the deterministic model. Though the outbreak size is extremely high before April, the success rate is quite low, while after May, though the success rate reaches its peak, the outbreak size drops dramatically. Therefore, the average FES has the highest value in mid-April, and the peak time changes only slightly with climate. Similar to Fig 4A and Fig 7A, Scenario 2014 has the highest average FES.
One limitation of the previous deterministic model was that it failed to account for the effect of chance [7]. The transmission dynamics is highly affected by chance at the onset phase of an epidemic, when the populations of infected humans and vectors are low. Under the same situation, an imported case can sometimes result in local transmission, while not at others. However, every imported case can lead to local transmission in the deterministic model when the basic reproduction number R0 is greater than one. Therefore, though the deterministic model concluded the early timing of local transmission to be the main determinant of the large FES in 2014, it was unable to answer the question that why it happened earlier in 2014. Here stochasticity was added to the previous model to investigate the successful invasion probability when the importation happens at different times and under different climate and intervention scenarios in 2014.
In [9], a stochastic model suggested that the slightly warmer summer was adequate to explain the large dengue outbreak in Madeira, 2012. However, in the Guangzhou outbreak, the favorable climate was only a necessary, but not sufficient condition, because human activities such as intervention and travel also matter. We find the higher number of imported cases in May and June to be the most important determinant of the early outbreak. Although the excessive rainfall in 2014 did increase the successful invasion rate, this effect was cancelled out by the low initial water level due to the interventions in late 2013. Because of a small outbreak in 2013, regular interventions were conducted from late September to early November, which markedly reduced the initial water storage in 2014. The entomological surveillance data shown in Fig 3C supports this interpretation. Specifically, the MOI, which represents the adult mosquito abundance, was lower in 2014 than in 2013, despite the more favorable climate. The increased number of imported cases in late April to early July was also identified as the driving force of the 2014 Guangzhou dengue outbreak by a deterministic model, which incorporated the number of imported case by a fitting function [44]. The result of a classification tree also suggested the number of imported cases, monthly average BI, and temperature as the most important determinants of the dengue outbreak occurrence in Guangzhou from 2002 to 2013 [45].
As noted in Fig 8A and 8B, dengue endemic countries Thailand, Singapore, Malaysia and Vietnam had the highest tourist exchange with Guangzhou. Therefore, more attention should be paid to identifying imported cases from these countries when they have significant dengue outbreaks, especially in late spring and early summer. Though the successful invasion rate is likely to be low in this period, the final outbreak size can be extremely large once it succeeds (Fig 7A). In all countries, the dengue incidences in 2014 were lower than 2013, except for Malaysia (Fig 8C). Although the incidences are somewhat comparable in time, it is questionable to compare the incidence of different countries, because the reporting rates may vary significantly between countries.
A phylogenetic analysis suggested that the dengue virus isolated in Guangzhou, 2014 was most similar to the isolates in Singapore, Malaysia, Indonesia and Thailand [46, 47]. The importation index, which incorporates both the incidence in country of origin and the travel volume between the target city and the country of origin [48, 49], also shows that Singapore, Thailand and Malaysia had the highest importation indices in 2013, and the order changed to Malaysia, Singapore and Thailand in 2014. However, because of the differences in reporting rates between countries, this result is inconclusive. When compared across time, the importation index of 2014 is lower than that of 2013, but more imported cases were observed in 2014, especially in May and June (Fig 5). Considering the epidemic had not started yet at that time, the excessive number of reported imported cases was unlikely to be caused by increasing reporting rate. One possible explanation is spatial heterogeneity. The importation index assumes that human and vector populations are well-mixed, which means that the contact rate between every person and every vector is the same. However, the actual contact rate can be affected by tour route, local transmission hotspots, as well as personal habits, such as whether take measures to avoid mosquito bites. The other explanation is the different reporting rate. For example, if the reporting rate is lower in Malaysia than in other countries, then its actual contribution to the total importation index should be higher, which may make the pattern of the total importation index more similar to that of Malaysia.
The first step of a successful invasion is infecting adult mosquitoes (Event 12), which is described in the model by transition rate bαhvHiAs/N. Both the biting rate b and the transmission probability from human to vector αhv ranges from 0 to 1, Hi equals to 1, and the total human population N changed little in the study time period. Hence, the size of the susceptible mosquito population, As, has the greatest influence on this transition rate. As can be seen from Figs 4, 9A and 9B, it is obvious that the success rate and As have almost the same pattern. Furthermore, we calculated the cross-correlation function (CCF) between the difference of success rate and the difference of As. First differences were taken here to remove trend of the time series. The result shows that success rate at day t is positively correlated with As at day t + 10 for all parameter sets under all scenarios. The average result for Scenario 2014 is shown in Fig 9C. Since the average recovery period is assumed to be 6 days in the model [7], and we use a 10-day time window for the imported case, the 10-day time lag between As and success rate is reasonable.
When we explore the individual results for each of the 100 parameter sets, we found that the gap between the success rate under different climate scenarios is more marked when the maximum water level ωmax in that parameter set is larger (Fig 9D). This is because precipitation can affect the carrying capacity through water availability only before ωmax is reached, since water will overflow after reach this threshold. Therefore, if the ωmax is larger, it takes longer to reach this threshold, which will then lead to larger difference in the mosquito abundance and success rate. This phenomenon implies that reducing ωmax by destroying breeding sites can mitigate the effect of climate on dengue transmission dynamics.
Regression models are widely used to predict the occurrence of dengue outbreaks, monthly reported cases and the FES [5, 6, 45, 50]. Here we suggest the factors to include in these models according to the results of our stochastic model.
Once given the timing of each imported case, the outbreak probability can be calculated from the success rate curve (Fig 4), which depends mainly on mosquito abundance. Therefore, to predict the occurrence of dengue outbreaks in subtropical areas, variables, such as the number of imported cases and mosquito abundance, should be included. Mosquito abundance is related to temperature, water availability and interventions. Water availability can be estimated from rainfall, evaporation, intervention, and the water limitation parameters ωmax and ωmax. Without all these data, water availability can be indirectly represented by precipitation, and relative humidity. However, precipitation and relative humidity do not reflect the effects of intervention and the relationship between water level and precipitation or relative humidity is non-linear. The water level increases with precipitation in spring and early summer, but when it reaches its threshold ωmax, the water level cannot increase further and remains at this value until evaporation outweighs precipitation or the occurrence of an intervention. In addition, the relationship between water level and mosquito abundance is complicated, which depends on previous water level, the environmental carrying capacity and the previous abundance of aquatic stages. As a result, mosquito indices might still be preferred among all these variables. By including the monthly maximum and minimum temperature, average BI, and number of imported cases, a classification tree can be used to predict the dengue outbreak occurrence in Guangzhou [45]. For tropical areas, since local epidemics continue throughout the entire year, there is no need to predict the outbreak occurrence.
The FES depends on both the final outbreak size and outbreak probability. The final outbreak size decreases exponentially as a function of the introduction date, while temperature, water storage, and human interventions can modify the decrease rate through changing the force of infection. Earlier introduction in subtropical areas means a longer period of transmission and more cases, because dengue transmission is terminated by the low temperature every winter. The force of infection can be affected by mosquito abundance, biting rate and virus virulence. Although both the length of transmission and the force of infection can affect the final outbreak size exponentially, the former fluctuates over a much wider range. Sometimes, though the climate is favorable for vector growth, no big outbreak occurs because of the low level or late importation. Thus the imported cases introduce high uncertainty to the prediction models, and a successful model should ideally include the number of imported cases and indigenous cases in the previous month to reduce this uncertainty. Time series regression models with the number of imported and indigenous cases in the previous month, lagged temperature, precipitation and relative humidity can give reasonable estimations of the number of new cases [6, 50]. Climate conditions in tropical areas can support overwinter local transmission, so the role of imported cases here is not as important as that in subtropical areas. Climate and interventions are the only crucial factors in tropical areas. Prediction models based on only climate can have good prediction power [51–53]. When considering the spatial aspects of the transmission, heterogeneity in the distribution of population, open water, vegetation, and host immunity can also influence the outbreak probability and FES for both subtropical and tropical countries [54, 55].
Currently, the most common interventions in China are chemical insecticide spraying and environmental management, such as water container emptying. Also, releasing mosquito larvae-eating fish Gambusia affins and wolbachia-infected male mosquitoes are sometimes used in small scale tests.
Since the adult abundance and success invasion rate of imported cases increase sharply in April and May (Fig 4), the interventions should begin before or in this time period. However, the regular interventions in Guangzhou started on September 27th and July 25th in 2013 and 2014, respectively, both of which were at least one month later than the onset of local epidemics. In addition, interventions out of the transmission season, such as before the next transmission season can also be important, because it can reduce the water storage, increase the time needed to accumulate water in the next transmission season, and postpone the outbreak, which, in turn, can then decrease the final outbreak size significantly.
According to result of the PRECIS (Providing Regional Climate for Impact Study) regional climate modeling system, the spring and summer rainfall of South China will increase in the future [56], which may increase the average FES by causing earlier outbreaks. However, destroying mosquito breeding sites can mitigate this detrimental impact of climate change, since it can reduce the maximum water level ωmax and narrow the gaps in the success rate or final outbreak size between different climate scenarios (Fig 9D).
Increasing international travel, urbanization, global warming, and changes in precipitation patterns pose higher dengue outbreak risk for Guangzhou and other subtropical non-endemic areas. Moreover, secondary infection can worsen the situation by increasing the incidence of life-threatening dengue hemorrhagic fever and shock syndrome [57]. Under this situation, vector control and management of imported cases at entry points should be implemented strictly, especially around mid-April, to reduce possible health and economic loss caused by dengue or other mosquito-borne diseases, such as yellow fever and Zika. Since the travel volumes between Thailand, Malaysia, Singapore and Guangzhou are the highest, more attentions should be paid to the detection of imported cases when dengue outbreaks occur in these countries, especially at early summer. Besides international imported cases, cases from other Chinese cities should also be considered. Since Guangzhou is normally the first city to have dengue outbreak sin mainland China, it is unusual for Guangzhou to have domestic imported cases in the early phase of an epidemic, although it happens on occasion. For example, the epidemic in Zhongshan started 2 weeks earlier than in Guangzhou [50]. On the other hand, the onset of an epidemic in Guangzhou signals that surrounding cities of the need to pay attention to case detection and vector control immediately. For example, from Figs 4B and 9C, it can be seen that, after detection, the model suggests that interventions should be conducted around the residences of the reported cases whose symptoms began less than 10 days ago, even if they have recovered.
The model proposed here represents the transmission of only one serotype DENV-1, since 5,947 out of 6,024 cases (98.7%) tested in 2014 were infected by DENV-1, and only 74 and 3 were infected by DENV-2 and -3, respectively [48]. When adapted for use in other subtropical areas, the single serotype assumption may or may not be appropriate. Secondly, we assumed that dengue virus virulence and reporting rate were the same for 2013 and 2014, which led to similar initial exponential growth rates for these two years. However, from the daily reported new case data, the growth rate of 2014 was slightly higher than that of 2013 (Fig 3), the reasons for which deserve further investigation. Thirdly, the estimated success rates illustrated in Fig 4 are estimates of the upper bound on what might be expected in practice, because it estimates the success rates for imported cases who spend the whole viremic period in Guangzhou. However, an imported case can enter Guangzhou at the middle or even the last day of the viremic period and thereby have less time to infect Ae. albopictus and result in a lower successful invasion rate. In this model, we also assumed that humans and the mosquitoes are well-mixed, which means the contact rate is the same for every host and vector. However, studies of Ae. albopictus indicate that their flying range is as small as 300 meters, and that they normally they stay near their breeding sites for their whole life [58]. Therefore, a spatially-explicit model is needed to better simulate the spatial distribution of dengue cases and study the risk factors related to the observed distribution patterns in the future.
|
10.1371/journal.pntd.0001001 | Evolutionary History of Rabies in Ghana | Rabies virus (RABV) is enzootic throughout Africa, with the domestic dog (Canis familiaris) being the principal vector. Dog rabies is estimated to cause 24,000 human deaths per year in Africa, however, this estimate is still considered to be conservative. Two sub-Saharan African RABV lineages have been detected in West Africa. Lineage 2 is present throughout West Africa, whereas Africa 1a dominates in northern and eastern Africa, but has been detected in Nigeria and Gabon, and Africa 1b was previously absent from West Africa. We confirmed the presence of RABV in a cohort of 76 brain samples obtained from rabid animals in Ghana collected over an eighteen-month period (2007–2009). Phylogenetic analysis of the sequences obtained confirmed all viruses to be RABV, belonging to lineages previously detected in sub-Saharan Africa. However, unlike earlier reported studies that suggested a single lineage (Africa 2) circulates in West Africa, we identified viruses belonging to the Africa 2 lineage and both Africa 1 (a and b) sub-lineages. Phylogeographic Bayesian Markov chain Monte Carlo analysis of a 405 bp fragment of the RABV nucleoprotein gene from the 76 new sequences derived from Ghanaian animals suggest that within the Africa 2 lineage three clades co-circulate with their origins in other West African countries. Africa 1a is probably a western extension of a clade circulating in central Africa and the Africa 1b virus a probable recent introduction from eastern Africa. We also developed and tested a novel reverse-transcription loop-mediated isothermal amplification (RT-LAMP) assay for the detection of RABV in African laboratories. This RT-LAMP was shown to detect both Africa 1 and 2 viruses, including its adaptation to a lateral flow device format for product visualization. These data suggest that RABV epidemiology is more complex than previously thought in West Africa and that there have been repeated introductions of RABV into Ghana. This analysis highlights the potential problems of individual developing nations implementing rabies control programmes in the absence of a regional programme.
| Rabies virus (RABV) is widespread throughout Africa, with the domestic dog being the principal vector. Dog rabies is estimated to cause 24,000 human deaths per year in Africa, however, this estimate is still considered to be conservative. Two sub-Saharan African RABV lineages (Africa 1 and 2) are thought to circulate in western and central Africa. We confirmed the presence of RABV in a cohort of 76 brain samples obtained from rabid animals in Ghana collected from 2007 to 2009. In addition we developed and tested a novel molecular diagnostic assay for the detection of RABV, which offers an alternative RABV diagnostic tool for African laboratories. Our analysis of the genetic sequences obtained confirmed all viruses to be RABV, however, unlike previous studies we detected two sub-Saharan African RABV viruses (Africa 1 and 2) in this cohort, which included a single virus previously undetected in West Africa. We suggest that there has been repeated introduction of new RABVs into Ghana over a prolonged period from other West African countries and more recently from eastern Africa. These observations further highlight the problems of individual developing nations implementing rabies control programmes at a local, rather than regional level.
| Viruses belonging to the genus Lyssavirus, family Rhabdoviridae, cause the disease rabies. Rabies virus (RABV) is enzootic throughout Africa with the domestic dog (Canis familiaris) being the principal vector [1]. Sylvatic rabies is also reported in a number of wildlife hosts, particularly in southern Africa [2], [3], [4], [5]. Rabies remains the only disease known to have a 100% mortality rate and has a high DALY (disability adjusted life years) score compared with other ‘neglected zoonoses’ [1], [6], [7]. Dog rabies is estimated to cause 24,000 (7000–46000, 95% percentiles) human deaths per year in Africa [1], however, this figure is still considered to be a conservative estimate as rabies cases in humans are widely under-reported in parts of Africa [8], [9].
Rabies has been present within the dog population of Ghana for decades [10], [11]. Previously, control methods including dog vaccination and stray dog removal have been intermittent and not sustained. Unfortunately, as in several other developing African countries, rabies diagnostics within the Ghanaian veterinary services remains limited to non-Lyssavirus species specific staining techniques, including the Sellers' stain and fluorescent antibody test (FAT) [12]. Currently, only individual owners vaccinate their dogs for their (owner and dog) protection. Between 1970 and 1974, an average of 72 cases of canine rabies were reported annually throughout the country [10]. Between 1977 and 1981 this number increased to over 100 cases annually, with an incidence of human rabies cases rising to 27 in 1981 [11]. Since 1981 there have been no further published reports of rabies in Ghana, and rabies viruses from the country have not been included in phylogenetic analyses of rabies in Africa [13], [14]. The virus is believed to cause disease in approximately 0–60% of those patients that are exposed depending on route of exposure [8]. Despite this, 123 clinically-confirmed human cases were recorded by public health officials between 2000 and 2004 (unpublished results). Moreover, ‘suspect’ human rabies cases are rarely confirmed using a laboratory-based diagnosis, relying solely on a clinical diagnosis [9].
The first phylogenetic study of rabies viruses from sub-Saharan Africa established three genetically distinct lineages (Africa 1, 2, and 3) [15]. Sub-lineage Africa 1a dominates northern and eastern Africa, but has also been detected in Nigeria, Gabon and Madagascar, suggesting a very broad distribution. Sub-lineage 1b is found in eastern, central and southern Africa and lineage 2 is present in an uninterrupted band across West Africa as far east as Chad [13], [16]. Africa 1 and 2 lineages have been detected in a range of domestic and wild carnivore species. While domestic dogs appear to be the only population essential for maintenance of canid variants in some parts of Africa [17], [18], wild canids have been suggested to contribute to sustaining canine rabies cycles in specific geographic loci in South Africa and Zimbabwe [19], [20], [21]. A third lineage (Africa 3) is thought to be maintained within viverrid species in southern Africa [22], [23], [24]. This phylogenetic distinction has been supported by studies investigating rabies across Africa [13], [25], epidemiological studies of rabies within specific countries [3], [16], [18], [26], studies on wildlife populations [5], [27], [28] and investigations into the origin of human rabies [29], [30]. More recently another distinct lineage, Africa 4, has been identified in northern Africa [31].
The principal objectives of this study were to characterise the lyssaviruses causing rabies in Ghana and to understand the evolutionary history of the circulating viruses. We also assessed the performance of a novel isothermal amplification technique for the detection of rabies virus for use in African laboratories. The low threshold of technology required to use this technique for diagnosis of animal diseases in Africa has been advocated [32], [33].
The Republic of Ghana is on the southern coast of West Africa (Figure 1). It shares borders with Togo (east), Ivory Coast (west), and Burkina Faso (north). Ghana has several ecosystems broadly attributed to the patterns of rainfall and geological topology [34]. The south eastern coastline consists of mostly low plains and scrubland, and separates the upper and lower Guinea African forest systems. Southwest and south central Ghana is a semi-deciduous forested plateau. Savannah dominates the northern part of the country. There are geographical features that may represent barriers to rabies spread in Ghana. The highest point in Ghana is only 885 m above sea level along the eastern border, however, the world's largest artificial lake, Lake Volta, separates much of eastern Ghana from the rest [34]. Ghana's population has rapidly increased in the last few decades. A census in 1961 recorded 6.7 million people, however, the current estimate is approximately 24 million [35].
Brain samples were derived from dogs (74) and cats (2) brought to the central diagnostic veterinary laboratory (Veterinary Services Laboratory, VSL) in the capital of Ghana, Accra, on suspicion of being rabid (Table S1). The samples used in this study were obtained by the Ghanaian government's veterinary services laboratory from naturally infected rabid animals in Ghana. No samples were obtained from, nor animals used in, an experimental study. All samples were obtained from animals within 142 km of Accra. Infection with RABV was suspected from clinical signs and from test results using either Sellers' staining (n = 69) of Negri bodies or the FAT (n = 7) in the VSL [12]. The panel were assigned numbers randomly and transferred from the VSL to the Veterinary Laboratories Agency (VLA), Weybridge, UK, where further molecular analysis was undertaken.
Total RNA was extracted from each brain sample using Trizol (Invitrogen) following the manufacturer's protocol. Pellets were resuspended in 10 µL of HPLC grade water. Reverse transcription and polymerase chain reaction were performed using previously published methods to amplify a 600 bp region of the nucleoprotein gene [36].
A novel reverse-transcription loop-mediated isothermal amplification assay (RT-LAMP) was applied to a limited panel of ten samples systematically taken from the larger randomly numbered Ghanaian panel. Previous reports applied this technique to viruses from a range of countries [37], [38] or to fixed rabies virus [32]. The assay is composed of two sets of primers (Table 1). The first, designated Rab1, amplifies viruses belonging to the cosmopolitan lineage. The second, Rab4, amplifies viruses belonging to the arctic lineage. A reaction mixture incorporating a combination of all 12 primers amplifies viruses from both groups (data not shown). 1 µg of each RNA sample was added to a reaction mixture containing each of the 12 primers at the final concentration indicated in Table 1, Isothermal Mastermix (GeneSys Ltd) and 0.12 units Thermoscript reverse transcriptase (RT) (Invitrogen) in a final reaction volume of 25 µl. A cosmopolitan RABV obtained from a Turkish dog that had been used to develop the assay (data not shown) was included as a positive control. A no-template control sample (HPLC grade water) was used as a negative control. The reaction was incubated at 65°C for 1 hour. A 10 µl aliquot was removed and mixed with 2 µl sample loading buffer and loaded onto a 1% agarose gel containing ethidium bromide and separated at 80 volts for 1 hour. The amplification products were visualized by UV irradiation. The RT-LAMP assay was further adapted for use with a lateral flow device (LFD) for visualization of RT-LAMP products. The assay was run with the above conditions and reagents, but with the alternative loop primer sets (Table 1: Rab1 FLOOPFlc, Rab1 BLOOPBtn, Rab3 FLOOPFlc, Rab3 BLOOPBtn, Forsite Diagnostics). The LFD (Forsite) uses a mouse anti-biotin monoclonal antibody (MAb) in the “get wet” strip to indicate the LFD run succeeded and a mouse anti-fluorescein MAb to bind the LAMP product to the fluorescein tag to show a positive result. The product was diluted in 1∶500 volumes of HPLC grade water and 60 µl added to the LFD test well.
Direct consensus DNA sequencing of a 405 bp region of the nucleoprotein (N) gene was undertaken as previously described [39]. Sequences produced were edited using SeqMan (DNAstar Lasergene) and aligned (ClustalW, Megalign, DNAstar Lasergene). Further analysis of the newly derived sequences was undertaken using Bayesian Markov chain Monte Carlo (MCMC) phylogenetic analysis using BEAST software (version 1.6.1) [40] with a panel of pan-African RABV selected from GenBank (Table S2). Sequences were aligned in ClustalX2 (version 1.2). A relaxed-clock (uncorrelated lognormal) [41] was employed in conjunction with a general time reversible (GTR) model of substitution with gamma distributed variation in rates amongst sites and a proportion of sites assumed to be invariant. This method allows the evolutionary rate of each branch to vary without assuming these rates are correlated among adjacent branches. A model of constant population size was employed for the phylogeographic analysis, motivated by a preliminary analysis of the data using a non-parametric model of growth under which suggested no significant deviation from the constant size. The MCMC was run for 30,000,000 steps with parameters and trees sampled every 6,000 steps. Parameter effective sample sizes were >100 and posterior distributions were inspected to ensure adequate mixing in Tracer (version 1.5). A phylogeographic approach was not taken to analyze the correlation between lineage and distance, due to all animals reportedly originating within close proximity from central Accra. To infer the temporal and spatial diffusion of Africa 1 and Africa 2 clades into Ghana, a continuous-time Markov chain (CTMC) process over discrete sampling locations was employed in a phylogeographic analysis of each clade using BEAST. The sampling origin for each sequence was considered to be the centroid of the country from which the sequence was sampled [14]. The same models of nucleotide substitution, growth and clock rate were employed as before, but an MCMC chain length of 100 million steps was used to ensure sufficient mixing and convergence of all phylogeographic parameters, and trees were logged every 20,000 steps. An appropriate (maximum 10%) burn-in was removed from each and the sampled trees were summarized as maximum clade credibility (MCC) trees. All sequences reported in this study (Table S1) were deposited in GenBank.
Seven of the 69 samples from suspected rabies cases tested at the VSL Accra were negative by Sellers' stain, whereas each of the seven tested by FAT was positive. Due to clinical signs exhibited by the animals, all 76 samples were included for further analysis at VLA-Weybridge and were subsequently positive by RT-PCR for RABV. Sequence analysis demonstrated that all viruses belonged to lineages previously reported from Africa. Twenty-seven samples were from the Africa 2 lineage, 48 samples from the Africa 1a sub-lineage, and a solitary sequence (sample G13) belonged to the Africa 1b sub-lineage (Table S1, Figure 2).
The MCMC tree of a 405 bp region of the 76 RABV N gene sequences analyzed with 20 African RABV sequences from GenBank is shown in Figure 2. The topology is similar to other analyses of African RABV N genes [15] that included Africa 1, 2, and 3 lineages. Rabies viruses from Ghana clearly form two lineages, Africa 1 (49 viruses) and 2 (27 viruses). Within each lineage sequences are separated into sub-lineages, in the case of Africa 1, or clades in that of the Africa 2 lineage. Our analysis estimates that the Africa 2 lineage diverged approximately 181 years ago (73–313 yrs, 95% HPD).
Within the Africa 2 lineage we detected three clades in the sample of viruses from Ghana. In order to test the hypothesis that these clades entered Ghana from different West African countries and to understand these viruses' evolutionary history, we re-analyzed the Africa 2 sequence data with 139 Africa 2 sequences alone, including the eleven used previously (Table S2, Figure 3). Thirteen Africa 2 viruses form a clade with a virus from Benin, with a time to the most recent common ancestor (TMRCA) estimated between 23 and 73 years (95% HPD) and there is a considerably higher posterior probability (0.442) for the ancestor of this clade to have originated in Benin than any other sampled country (Figure 3). A further thirteen Africa 2 viruses form a clade with viruses from Niger and Burkina Faso, with a TMRCA estimated to be between 22 and 53 years. It is most likely that this clade entered Ghana from Niger (posterior probability = 0.464). A single Africa 2 virus (G6) shares a common ancestry with viruses from Ivory Coast and Burkina Faso and has a more recent ancestry of between 1 and 20 years. There is very high support for the ancestor of this clade to have originated in the Ivory Coast, before entering Ghana.
Phylogeographic analysis of the newly sequenced Ghana Africa 1 sequences with pan-African 1 sequences (Table S2, Figure 4) confirmed a monophyletic group of Africa 1a viruses (Figure 4). This clade is estimated to have emerged 23–31 years ago from Gabon (posterior probability = 0.944). The spatial analysis also provides high support for the introduction of the single Africa 1b virus from Kenya (posterior probability = 0.937) 15–22 years ago (95% HPD) (Figure 4).
For ten randomly selected samples from the cohort, RT-LAMP detected RABV from each sample with a similar banding pattern to the positive control when separated by agarose gel electrophoresis (Figure 5) or when biotinylated products were applied to a LFD (Figure 6). This group comprised three Africa 2 and seven Africa 1 viruses (Table S1). The cost of this assay was calculated at approximately $3 per assay.
Each of the 76 brain samples used in this study was positive for RABV antigen. The overall topology of the phylogenetic tree produced by our analysis of the RABV N-gene sequence data available from a sample of rabid African dogs and cats in Ghana was consistent with those previously described [13], [15], [42]. This analysis of Ghanaian rabies cases is the first phylogenetic analysis of RABV from Ghana. Where this analysis is distinct from reports of RABV in other West African nations is in the diversity of viruses detected within Ghana. The samples were all taken from a relatively small geographical region with those samples not from within the greater Accra region originating from towns relatively close to Accra. These included eight viruses from Tema and five from Cape Coast (25 and 142 km from Accra, respectively). There was no evidence of infection with Africa 3 RABV (detected in mongoose in southern Africa) [22], [23], [24], Africa 4 RABV (detected in north-eastern Africa) [31] or other Lyssavirus species such as Lagos bat virus, against which a high seroprevalence of antibodies has been detected in bats from Accra [43]. However, our analysis suggests that rabies epidemiology is much more complex than at first thought from previous studies within West Africa. Indeed, whilst West African countries typically have defined lineages circulating within them, only Nigeria and the Central African Republic have previously been described as having Africa 1 and 2 lineage viruses co-circulating within their national borders [13], [16]. We detected both in Ghana, and propose that Ghana's recent history and geography may explain why both virus lineages were detected.
Africa 2 viruses appear to have been present within the dog populations of West Africa, including Ghana, for decades. This is derived from the close relationships between the RABV characterized in Ghana and those reported in other West African countries, such as Benin, Ivory Coast, Burkina Faso and Niger. Our results support the findings of others that the Africa 2 virus lineage has been circulating within Africa for less than 200 years [13]. Within Ghana, our analysis suggests the Africa 2 clades now co-circulating in Ghana have different evolutionary histories. From the Africa 2 phylogenetic analysis (Figure 3), we hypothesize that the three Ghanaian Africa 2 clades co-circulate in Ghana, but share evolutionary histories with viruses from other West African countries. Whilst we cannot be certain of the direction of the virus spread, we believe that there have been three different introductions of Africa 2 viruses to Ghana. We found support for the hypothesis that one clade that circulates in Ghana and in the northeasterly West African countries of Niger and Burkina Faso was originally imported from Niger and subsequently entered both Ghana and Burkina Faso (Figure 3). Another clade of viruses share a common ancestry with a Beninese isolate from the east and likely entered the country from Benin or via neighboring Togo. The evolutionary history of those viruses from the east and northeast may be due to Lake Volta providing a physical obstacle to virus transmission between dog populations. Further analysis of this phylogenetic relationship is precluded, however, by the lack of additional published sequences from Benin, and none from neighboring Togo. A single virus, G6, forms a clade with isolates from the Ivory Coast. This virus appears to be a recent introduction, sharing a TMRCA of just 1 to 20 years with viruses from the Ivory Coast to the west. A possible reason for fewer viruses being from the Ivory Coast may be the large tropical forest system along the Ghana-Ivory Coast border providing a barrier to dog movements. The border with the Ivory Coast was historically the most forested area of Ghana, however rapid deforestation and increasingly easy “between country” travel may have led to the trans-boundary movements of this virus.
Due to the historical dominance of Africa 1 viruses in the northern, eastern and southern parts of Africa, we believe it reasonable to hypothesize that Africa 1 viruses have entered Ghana from those regions, and that transmission has not been from Ghana to those regions. This hypothesis is supported by the phylogeographic analysis which suggests that the virus sub-lineage Africa 1a was transmitted from central African counties to Ghana. If we accept this, the origin of the Ghanaian Africa 1a sub-lineage viruses may be explained simply by virus transmission through dog (and potentially other vector) populations from central African nations to Ghana (Figure 4). Indeed, in our analysis the Ghanaian Africa 1a viruses share an ancestry with a virus from Gabon with a TMRCA estimated to be 23–31 years ago. This would require viruses to be transmitted at an approximate rate of between 39 to 53 kilometers per year. The large number of Africa 1a viruses in our sample suggests that this sub-lineage is well established in the Accra region, however further virus sequences from nations between Ghana and Gabon are required to confirm the evolutionary history of this sub-lineage.
The presence of an Africa 1b sub-lineage RABV in our analysis is the first reported from West Africa. Analysis of the Africa 1 lineage viruses suggests that this virus shares an ancestry with viruses from East Africa, in particular, those from Kenya (Figure 4). The presence of this virus may be explained in one of two, not exclusive, ways. Firstly, sub-lineage 1b viruses may simply have been transmitted within the populations of dogs and other susceptible animals from eastern African countries to Ghana. Transmission from Kenya (with Nairobi approximately 4200 km from Accra) would require virus transmission at a rate of approximately 190–279 kilometers per year with the TMRCA estimated to be 18 years (15–22 years, 95% HPD). Given the distance infected dogs and potential wildlife hosts may travel, this is theoretically possible, but highly unlikely given that rabies spread in red foxes and raccoons in Europe and North America was estimated to be typically 30–60 kilometers a year [44], [45]. Therefore, we hypothesize that the more likely reason for this virus' presence in Ghana is that an infected animal was translocated from the east, thus introducing a new sub-lineage to the region. Indeed, we believe that this may be the first report of molecular evidence of a long distance translocation of a rabies sub-lineage in Africa.
Spatio-temporal models of rabies in eastern and southern Africa show large-scale synchrony of rabies epidemics across both regions [46]. The analysis by Hampson et al provided evidence that movement of infectious animals, or animals in the incubation period, and localized regional or national vaccination campaigns during epidemics, are likely to lead to rabies synchrony [46]. However, evidence provided by rabies control programmes in both Europe and the Americas show that large-scale control programmes can be successful [47], [48], [49], [50]. A study of rabies in Tanzania also suggested dog rabies control was feasible, but was hampered by perceived problems that were largely unfounded [7]. A subsequent analysis by Hampson et al suggested that regular regional pulsed vaccination programmes would be required to eliminate dog rabies [51]. Despite the analysis estimating the basic reproductive rate of domestic dog rabies throughout the world to be low (R0<2), the rapid turnover of dog populations led to enough susceptible hosts for rabies to be maintained [51]. Our molecular study suggests introductions of RABV from neighboring countries into Ghana are not infrequent, demonstrating that without substantial support for continuous vaccination or coherent regional cooperation, Ghana will be unable to eliminate rabies and maintain a rabies-free status. In addition to this, our analysis provides evidence of a virus that shares a recent common ancestry with viruses from East Africa, therefore providing further evidence that regional control programmes must be implemented and that once rabies is eliminated, vigilance and technical expertise must be maintained in order for new introductions to be controlled [46].
Currently rabies diagnostics within the Ghanaian veterinary services remain limited to non-Lyssavirus species specific staining techniques, including the Sellers' stain and, when FITC conjugate is available, FAT. Inadequate government and financial commitments and a resource limited veterinary infrastructure are restrictive factors that preclude a sustainable rabies diagnostic service in Ghana. Surveillance activities should be given a higher priority to maintain an effective diagnostic service with the co-operation of other national and international organizations. Each of the 76 brain samples used in this study was positive for RABV infection by RT-PCR at VLA Weybridge. Of the 76 samples full histories were available for 72 positive rabies cases. However, seven samples were negative when tested by Sellers' stain at the VSL. The VSL recorded 66 humans being bitten by those 72 dogs for which histories were recorded (data not shown), including six bites to humans by the seven RABV positive cases that tested negative in the VSL. Further training and the availability of FITC conjugate for the FAT or use of the direct rapid immunohistochemical test (dRIT) [12], [52], [53] may have overcome some of the diagnostic problems. However, given that low cost isothermal RT-LAMP assays have been developed for a number of viruses affecting livestock in Africa, including Rift Valley Fever virus [33] and African Swine Fever virus [54], we developed and tested the RT-LAMP for use in African laboratories. The RT-LAMP may be prone to some of the same problems as other molecular techniques, such as cross-contamination, however it is a cheap molecular technique that produces a product that is available for further analysis such as sequencing of the approximately 200 bp product. We developed the novel RT-LAMP on randomly selected RABV samples, including both Africa 1 (a cosmopolitan) and 2 lineages. This assay successfully amplified viral genetic material producing a measurable DNA product for both Africa 1 and 2 lineage viruses. This isothermal diagnostic assay negates the need for thermal-cyclers for molecular diagnosis of RABV. The assay reagents costs approximately $3 per assay and therefore may prove a useful alternative assay for those laboratories that already have molecular expertise and adds to the range of rapid cost-effective diagnostic assays that will be fundamental if developing countries wish to develop their own RABV diagnostic capabilities. Whilst “snap test” LFD tests have previously been reported [55] our adaptation of the RT-LAMP assay to use an LFD platform, instead of UV illumination, further reduces the technology required for RABV diagnosis in African laboratories. Additional validation of this method will require comparison with the gold standard assays, assessment of larger panels of samples from throughout Africa, as well as evaluation of its sensitivity in detecting RABV in brain samples from OIE reference laboratories. These preliminary findings, however, demonstrate proof-of-concept and suggest that this technique has the potential to provide African laboratories with a cheap and rapid molecular detection method.
We conclude that our analysis of rabies virus sequences derived from Ghana has furthered the understanding of RABV epidemiology in West Africa. In particular, our analyses suggest that both Africa 1 and Africa 2 RABV lineages are present in Ghana. Africa 1b sub-lineage had previously not been reported in West Africa, and its detection, along with evidence of an additional four further clades circulating in Ghana support previous analyses that suggest that only sustained regional level approaches to rabies control will be successful in rabies elimination. In addition, we have developed an African RABV RT-LAMP assay, which can be adapted for use with LFD platforms that we advocate will provide an additional diagnostic tool for African regional laboratories.
|
10.1371/journal.pcbi.1004679 | Capabilities and Limitations of Tissue Size Control through Passive Mechanical Forces | Embryogenesis is an extraordinarily robust process, exhibiting the ability to control tissue size and repair patterning defects in the face of environmental and genetic perturbations. The size and shape of a developing tissue is a function of the number and size of its constituent cells as well as their geometric packing. How these cellular properties are coordinated at the tissue level to ensure developmental robustness remains a mystery; understanding this process requires studying multiple concurrent processes that make up morphogenesis, including the spatial patterning of cell fates and apoptosis, as well as cell intercalations. In this work, we develop a computational model that aims to understand aspects of the robust pattern repair mechanisms of the Drosophila embryonic epidermal tissues. Size control in this system has previously been shown to rely on the regulation of apoptosis rather than proliferation; however, to date little work has been done to understand the role of cellular mechanics in this process. We employ a vertex model of an embryonic segment to test hypotheses about the emergence of this size control. Comparing the model to previously published data across wild type and genetic perturbations, we show that passive mechanical forces suffice to explain the observed size control in the posterior (P) compartment of a segment. However, observed asymmetries in cell death frequencies across the segment are demonstrated to require patterning of cellular properties in the model. Finally, we show that distinct forms of mechanical regulation in the model may be distinguished by differences in cell shapes in the P compartment, as quantified through experimentally accessible summary statistics, as well as by the tissue recoil after laser ablation experiments.
| Developing embryos are able to grow organs of the correct size even in the face of significant external perturbations. Such robust size control is achieved via tissue-level coordination of cell growth, proliferation, death and rearrangement, through mechanisms that are not well understood. Here, we employ computational modelling to test hypotheses of size control in the developing fruit fly. Segments in the surface tissues of the fruit fly embryo have been shown to achieve the same size even if the number of cells in each segment is perturbed genetically. We show that simple mechanical interactions between the cells of this tissue can recapitulate previously gathered data on tissue sizes and cell numbers. However, this simple model does not capture the experimentally observed spatial variation in cell death rates in this tissue, which may be explained through several adaptations to the model. These distinct adaptations may be distinguished through summary statistics of the tissue behaviour, such as statistics of cell shapes or tissue recoil after cutting. This work demonstrates how computational modelling can help investigate the complex mechanical interactions underlying tissue size and shape, which are important for understanding the underlying causes of birth defects and diseases driven by uncontrolled growth.
| The mechanisms underlying tissue size control during embryonic development are extremely robust. There are many cases where the rates of proliferation, growth, or death are perturbed significantly yet patterns are maintained or repaired during later stages of development. For example, even after 80% of the material in a mouse embryo is removed, accelerated growth can give rise to correctly proportioned, albeit non-viable offspring [1]. In fruit fly embryos, overexpressing the maternal effect gene bicoid leads to stark overgrowth in the head region, but the excess tissue is removed during later stages of development through apoptosis (programmed cell death), leading to viable adults [2]. Tetraploid salamanders of the species Amblystoma mexicanum have half the number of cells as their diploid counterparts, yet are the same size [3].
The robustness of tissue size control relies on tight coordination of cellular processes including growth, proliferation, apoptosis and movement at a tissue level. However, the fundamental mechanisms underlying such coordination remain largely unknown. In particular, the mechanical implementation of tissue size control is not well understood. The regulation of cellular mechanical properties is known to play a key role during morphogenetic events, such as tissue folding, elongation and cell sorting [4, 5]. For example, upregulation of myosin II generates tension that helps to straighten compartment boundaries in the Drosophila wing imaginal disc [6], while controlled cell death provides the tension required for invagination during Drosophila leg development [7]. It has been illustrated theoretically how mechanical feedback might facilitate uniform growth in epithelia in the face of morphogen gradients [8]. Could mechanical forces also play a significant role in robustly maintaining tissue size?
To explore questions of pattern repair, we develop a computational model of a patterned epithelium, with application to the segments of the Drosophila embryonic epidermis (Fig 1). These tissues define the body plan along the head-tail axis. They are first defined during stage 6 of embryonic development and are visible as stripes in the epidermis of the larva [9]. The segments are subdivided into anterior (A) and posterior (P) compartments, which are marked by distinct gene expression patterns. In particular, cells in the P compartment express the gene engrailed [10] (Fig 1D). While the initial specification and establishment of segments is relatively well studied [11], maintenance of segment identities have received much less attention. However, it is known that compartment dimensions can be robustly restored in the presence of genetic manipulations made during earlier developmental patterning events [2, 12–14]. Both the conserved epidermal growth factor receptor (EGFR) and Wnt/Wingless (WG) pathways have been implicated in regulating apoptosis to achieve pattern repair for perturbations made in each of the compartments and are known to antagonize each other [2, 14].
A major strength of Drosophila as a model organism is the availability of genetic tools that enable the ectopic expression of gene products or RNA interference (RNAi) constructs to manipulate cell growth, proliferation and signaling in a spatio-temporally controlled manner [15–17]. For example, the bipartite GAL4-UAS system can be used to drive expression of ectopic genes in embryos through a cross of one line containing a tissue-specific enhancer driving expression of the heterologous yeast transcription factor GAL4 with a second line that activates expression of a transgene upon binding of GAL4 to the UAS promoter region. Using this approach, Parker [14] investigated P compartment size using the GAL4 driver line as the control genotype engrailed-GAL4, UAS-GFP, in the following referred to as wt (wild type). This was compared to various perturbations (Fig 1B). In particular, these included crosses between the driver line and UAS-CyclinE (which we shall term en>CycE) and UAS-dacapo lines (further specified as en>dap), which perturbed the amount of final proliferation events towards the end of the normal range of proliferation in the epidermis (Fig 1A).
Parker [14] observed an increase in final cell number of more than 30% (Fig 1E, right bar) in the P compartments of en>CycE embryos, which exhibited an additional round of cell division. However, the size of the P compartment was much less affected by this perturbation (Fig 1F, right bar), as measured in first instar larvae [14]. Conversely, in en>dap embryos that exhibited a loss of one round of cell division, Parker [14] observed a reduction in cell number of 25% while, again, the compartment size was relatively unchanged (Fig 1E and 1F, middle bars).
Parker’s findings also suggest that epidermal growth factor receptor (EGFR) signaling, through the activating ligand Spitz, patterns apoptosis inside the P compartment. Spitz is released by a column of cells inside the anterior (A) compartment that is directly adjacent to the P compartment. Identifying cell death events through TUNEL staining [20], Parker [14] observed apoptosis much more frequently in the ‘front’ (more anterior) half of the P compartment, away from the Spitz source (Fig 1G), than the ‘back’ half. These numbers differed by a factor of nearly 40 in wt [14]. Counter-intuitively, inhibiting apoptosis by expressing the caspase inhibitor p35 inside the P compartment of en>CycE embryonic segments resulted in compartment shrinkage by nearly 10%.
The above findings shed light on tissue size control in the Drosophila embryonic epidermal tissues, suggesting a reliance on the regulation of apoptosis rather than proliferation. However, the cell-level interactions governing size control remain poorly understood. In particular, potential roles of cellular mechanics in augmenting or repairing growth defects in patterned tissues remain unexplored. To address this, we develop a vertex model of an embryonic segment to test hypotheses about the emergence of size control. Comparing the model to previously published data across wt and genetic perturbations, we investigate the extent to which passive mechanical forces might suffice to explain the observed size control and asymmetries in cell death frequencies across the P compartment. Our results suggest that the basis of size control can rely to a significant degree on the passive mechanical responses of cells. However, the observed spatial asymmetry in cell death frequencies requires patterning of mechanical properties by inter-cellular communication. These results also provide a basis for differentiating experimentally how extracellular signaling pathways like EGFR and WG might impact cellular decision making processes through predictions of observable cellular morphologies, and tissue behaviour after cell bond ablation.
We use a vertex model to simulate cell movement, intercalation, shape changes and apoptosis during the sixteenth round of divisions in a segment of the Drosophila embryonic epidermis. Vertex models were first introduced to study the structure of foams [21], and have since been applied to study a variety of epithelial tissues [6, 22–25]. For more information on vertex models and their application to epithelial morphogenesis, we refer the reader to two recent reviews [26, 27].
Vertex models approximate cells in epithelial sheets as polygons. The polygons represent the cells’ apical surfaces, where most inter-cellular forces originate [4]. The terms in the model account for the mechanical effect of the force-generating molecules that accumulate in the apical surface of the cells, such as actin, myosin, and E-cadherin. Vertices correspond to adherens junctions, and their positions are propagated over time using an overdamped force equation, reflecting that adherens junctions are not associated with a momentum. The force equation takes the form
μ d x i d t = - ∇ i E . (1)
Here, μ is the friction strength (which we assume to take the same constant value for all vertices), t is time, xi is the position vector of vertex i, and E denotes the energy of the whole system. The total number of vertices in the system may change over time due to cell division and apoptosis. The symbol ∇i denotes the gradient operator with respect to the coordinates of vertex i. The forces act to minimise a phenomenological energy function, based on the contributions thought to dominate epithelial mechanics [22]:
E = ∑ α K 2 ( A α - A 0 , α ) 2 + ∑ ⟨ i , j ⟩ Λ l i , j + ∑ α Γ 2 p α 2 . (2)
Here, the first sum runs over every cell in the sheet, Aα denotes the apical surface area of cell α and A0,α is its preferred area, or target area. This energy term penalises deviations from a target area for individual cells, thus imposing cellular bulk elasticity. The second sum runs over all edges 〈i,j〉 in the sheet and penalizes long edges (we choose Λ > 0), thus representing the combined effect of E-cadherin, myosin, and actin at the binding interface between two cells. The third sum also runs over all cells, and pα denotes the perimeter of cell α. This term models the effect of a contractile acto-myosin cable along the perimeter of each cell [22]. The parameters K, Λ, and Γ together govern the strength of the individual energy contributions. Although this description of cell mechanics is phenomenological, a variety of previous studies have demonstrated its ability to match observed junctional movements and cell shapes in epithelial sheets through validation against experimental measurements [6, 22, 25].
In contrast to many previous vertex model applications, we allow the mechanical parameters Λ, Γ, and A0 to vary between cells as a consequence of underlying tissue patterning. In particular, we consider A0 to be a function of cell generation and introduce the parameter
R A = A 0 , daughter / A 0 , mother (3)
as the ratio of target areas of daughter cells to mother cells. To ensure that the target areas of all cells add up to the total size of the spatial domain, which is assumed to be fixed, we choose the value RA = 0.5 unless stated otherwise. Throughout the study, variation of the parameter RA is used to account for cellular growth of daughter cells as well as changes in total target area upon division. In each simulation, the initial area of each cell, As, equals its initial target area, A 0 s, with A s = A 0 s = 121 μm2 (see discussion below for the choice of length scales in the model). In S1 Text and S4 Fig we analyse the extent to which deviations of cell target areas may affect the simulation results by increasing A 0 s. The simulated P compartment sizes and cell numbers are not strongly affected by such changes in initial condition, except for an increase in apoptosis for the en>CycE perturbation.
In contrast to several previous applications [22, 25] of the vertex model the spatial domain in this study is constrained due to the fact that there is no net organism growth during embryogenesis.
In addition to evolving vertex positions in accordance with Eq (1), we must allow for cell intercalation and cell removal through topological rearrangements. One such topological rearrangement is a T1 swap, which simulates cell neighbour exchanges. In a T1 swap an edge shared by two cells is removed and the cells are disconnected, while a new perpendicular edge is created that then connects the cells that were previously separated by the edge (see Fig 2B). In our implementation T1 swaps are executed whenever the length of a given edge decreases below a threshold lmin = 0.11 μm, which is 100 times smaller than the approximate length of a cell at the beginning of the simulation. The length of the new edge, lnew = 1.05lmin, is chosen to be slightly longer than this threshold in order to avoid an immediate reversion of the swap. A summary of the frequency of T1 swaps occurring in model simulations is provided in S1 Table. There are very few cell intercalation events in our simulations, with no T1 swaps observed for wt, in line with experimental observations of germ-band retraction [28].
A second topological rearrangement in vertex models is a T2 transition, during which a small triangular cell is removed from the tissue and replaced by a new vertex (see Fig 2B). In our implementation any triangular cell is removed if its area drops below a threshold Amin = 0.121 μm2, which is 100 times smaller than the initial area of each cell. The energy function Eq (2) in conjunction with T2 transitions can be understood as a model for cell removal: cells are extruded from the sheet by a T2 transition if they become mechanically unstable. Note that we do not discriminate between cell removal by cell death or by delamination, since this distinction is immaterial for our purposes. However, delamination has been shown to provide an alternative way of cell removal from epithelial sheets that is distinct from apoptosis [29]. Rates of cell removal predicted by previous vertex model applications have coincided with experimental measurements in the Drosophila wing imaginal disc [22] and notum [29].
All simulations start with N P s = 24 cells in the P compartment and N A s = 40 cells in the anterior compartment, to approximately match observed cell numbers [14] and to ensure that there are similar amounts of anterior tissue on either side of the P compartment.
In the case of a wt embryonic segment each cell divides once, with cell cycle times drawn independently from the uniform distribution on 0 to t ˜ w t = 600 time units. This corresponds to the duration of the sixteenth division cycle in the epidermis, which occurs during late stage 10 and early stage 11 and takes roughly 50 minutes [18]. After the round of divisions is complete, the system is allowed to relax for 200 more time units, corresponding to a total simulation time of twt = 800 time units.
For an en>CycE embryonic segment, the first round of divisions is implemented as for wt, but each cell in the P compartment then has a probability PCycE = 0.54 of dividing a second time once the first round of divisions is complete, with cell cycle times drawn independently from the uniform distribution from t ˜ C y c E = 600 to t ˜ C y c E = 1200 time units. This probability is inferred from published data on the en>CycE+p53 perturbation [14]; in this case apoptosis is blocked, allowing us to infer the average number of cell division events. The second period of 600 time units corresponds to the duration of the ectopic divisions in the en>CycE embryos, which occur during late stage 11 and early stage 12 [14]. After the second round of divisions is complete, the system is allowed to relax for 200 more time units, corresponding to a total simulation time of tCycE = 1400 time units.
For an en>dap embryonic segment, each cell in the P compartment has a fixed probability Pdap = 0.6 of not participating in the single round of divisions. This probability is inferred from published data on the en>dap perturbation [14]. As with wt, divisions occur during the first t ˜ w t = 600 time units, after which the system is allowed to relax for 200 more time units, corresponding to a total simulation time of twt = 800 time units.
These simulation times are chosen such that the system is at quasi-steady state at each time point. This quasi-steady state assumption is commonly used in vertex models [6, 22, 29, 30] and reflects the fact that the times associated with mechanical rearrangements (seconds to minutes) are an order of magnitude shorter than typical cell cycle times (hours) [22].
At each cell division event, a new edge is created that separates the newly created daughter cells. The new edge is drawn along the short axis of the polygon that represents the mother cell [31]. The short axis has been shown to approximate the division direction (cleavage plane) of cells in a variety of tissues [32], including the Drosophila wing imaginal disc [33]. The short axis of a polygon crosses the centre of mass of the polygon, and it is defined as the axis around which the moment of inertia of the polygon is maximised. Each daughter cell receives half the target area of the mother cell upon division unless stated otherwise.
In order to simulate the subsections of the P compartment we consider a spatial domain comprising two adjacent cell populations, the cells in the P compartment and parts of the adjacent tissue in the anterior compartment on either side of it. Sample simulation images are shown in Fig 2A and 2C. For simplicity, we assume that cells initially have regular hexagonal shapes. We analyse the sensitivity of P compartment sizes and cell numbers to this choice of initial condition in S1 Text and S3 Fig.
Motivated by the repeated pattern of A and P compartments along the anterior-posterior axis of the embryo, as well as by the fact that single P compartments stretch farther dorso-ventrally than the simulated region, doubly periodic boundary conditions are applied (Fig 2A). These boundary conditions keep the simulated region of interest at a fixed size. Compartment size changes are analysed as changes in the relative proportions of the anterior and posterior compartment within the simulated region.
An analysis of the sensitivity of P compartment sizes and cell numbers to this choice of boundary condition is provided in S1 Text and S1 Fig. The precise choice of boundary condition imposed in the model simulations does not significantly affect predicted compartment sizes and cell numbers.
To enable comparison of cell death rates in the front and back halves of the P compartment (see Fig 1G), a cell is defined to be in the front or back half if its centroid is located to the anterior (‘left’) or posterior (‘right’) side of the centre of the tissue, respectively. The tissue centre is defined to be the horizontal midpoint of the sheet at time t = 0 and is held fixed at all times.
When computing measures of cell shape in our analysis of simulation results, we define the area and perimeter of a cell to be those of the associated polygon in the vertex model, while ‘cell elongation’ is defined as the square root of the ratio of the largest to the smallest eigenvalues of the moment of inertia of that polygon. This latter measure provides a robust way to measure elongations of arbitrary shapes [31] and is comparable to the ratio of the lengths of the long and short axis of the best fit ellipse to a cell.
Unless stated otherwise, the line tension along the compartment boundaries is set to Λb = 2Λ, twice the value of the line tension in the remainder of the tissue. High tension along compartment boundaries is known to promote cell sorting and boundary straightness [6, 30], and the presence of myosin cables that can generate this tension between A and P compartments in the Drosophila embryonic epidermis has been reported [34]. S2 Fig shows that while the increase in line tension along compartment boundaries does affect the straightness of the boundary between A and P compartments in the model simulations, it does not significantly affect compartment sizes or cell numbers.
To investigate the consequences of asymmetries in cell mechanical properties on P compartment size control and patterning of apoptosis, we consider three distinct cases.
In the first case, we allow for asymmetry in cell target areas in the P compartment. This is implemented by modifying the target area of each cell in the P compartment to take the form
A 0 ′ = ( R A ) g ( 1 ± λ A ) , (4)
where RA = 0.5 as listed in Table 1, g ∈ {0, 1, 2} denotes the generation of the cell, and the − and + signs apply to cells located in the front and back halves of the compartment, respectively. We refer to the parameter λA as the area asymmetry.
In the second case, we allow for asymmetry in line tensions in the P compartment. This is implemented by modifying the line tension of each cell-cell interface (edge) inside the P compartment to take the form
Λ = Λ r ( 1 ± λ l ) , (5)
where Λr is the value of the line tension when no asymmetry is imposed. The + sign applies to all edges between P compartment edges whose midpoint is the front half of the compartment, while the − sign applies to all edges whose midpoint is in the back half of the compartment. We refer to the parameter λl as the line tension asymmetry.
In the third case, we allow for asymmetry in perimeter contractility in the P compartment. This is implemented by modifying the perimeter contractility of each cell in the P compartment to take the form
Γ = Γ r ( 1 ± λ p ) , (6)
where Γr is the value of the perimeter contractility when no asymmetry is imposed, and the + and − signs apply to cells in the front and the back halves of the P compartment, respectively. We refer to the parameter λp as the perimeter asymmetry.
The asymmetry parameters are all fixed at 0 in Figs 3, S1 and S2, and are varied in Figs 4, 5 and 6.
Prior to solving the model numerically, we non-dimensionalise it. Non-dimensionalising reduces the number of free parameters in the system and facilitates comparison of parameter values to previous implementations of the vertex model [22]. Rescaling space by the characteristic length scale L and time by the characteristic time scale T, Eqs (1) and (2) become
μ L T d x i ′ d t ′ = - 1 L ∇ i ′ E , (7) E = ∑ α K L 4 2 ( A α ′ - A 0 , α ′ ) 2 + ∑ ⟨ i , j ⟩ Λ L l i , j ′ + ∑ α Γ L 2 2 p α ′ 2 , (8)
where x′i, A α ′, A 0 , α ′, l i , j ′ and p α ′ denote the rescaled ith vertex positions, the rescaled cell area and cell target area, the rescaled edge length between vertices i and j, and the rescaled cell perimeter, respectively. The symbol ∇ i ′ denotes the gradient with respect to the rescaled ith vertex position. Multiplying the first equation by T/μL, we obtain
d x i ′ d t ′ = - ∇ i ′ T μ L 2 E , (9) E ′ = T μ L 2 E = ∑ α T K L 2 μ 1 2 ( A α ′ - A 0 , α ′ ) 2 + ∑ ⟨ i , j ⟩ Λ T L μ l i , j ′ + ∑ α Γ T μ 1 2 p α ′ 2 . (10)
Finally, by introducing the time scale T = μ/KL2, and the rescaled mechanical parameters Λ ¯ = Λ T / ( L μ ) = Λ / K L 3, Γ ¯ = Γ T / μ = Γ / ( K L 2 ) the non-dimensionalised equations read
d x i ′ d t ′ = - ∇ i ′ E ′ , (11) E ′ = ∑ α 1 2 ( A α ′ - A 0 , α ′ ) 2 + ∑ ⟨ i , j ⟩ Λ ¯ l i , j ′ + ∑ α Γ ¯ 2 p α ′ 2 . (12)
We choose the characteristic length scale L = 11 μm such that L2 is the mean cell area in the P compartment at the start of the simulation period, i.e. 121 μm2; the P compartment occupies a total area of 2.76×103 μm2 [14] and is initialized with 24 cells. The precise value of the characteristic time scale T depends on tissue properties (μ and K) and could be inferred from the duration of vertex recoil after laser ablation experiments, for example. In the non-dimensionalised model, cell shapes are governed by the rescaled target area of each cell and the rescaled mechanical parameters, Λ ¯ and Γ ¯. For these parameters we use previously proposed values [22], unless stated otherwise. A complete list of parameters used in this study is available in Table 1.
To solve Eqs (11) and (12) numerically we use an explicit forward Euler scheme:
x i ′ ( t ′ + Δ t ′ ) = x i ′ ( t ′ ) - ∇ i ′ E ′ ( t ′ ) Δ t ′ . (13)
The time step used in the forward Euler scheme is 0.01 rescaled time units and is manually chosen to ensure that the numerical scheme converges and that a further reduction in the time step does not change the results.
We implement the model within Chaste, an open source C++ library that provides a systematic framework for the simulation of vertex models [31, 35]. All code used to implement model simulations and to generate results presented in this work is provided (see S1 Software).
We first analyse the extent to which passive mechanical forces can lead to stable tissue size control as observed in [14]. We then investigate the effect of spatial regulation of cellular mechanical properties on P compartment sizes, cell numbers, and cell death locations.
As an initial study, we analyse simulations where compartment size is governed solely by passive mechanical properties of individual cells, and no further regulatory mechanism for size control is assumed. In particular, all cells in the tissue are specified to have the same mechanical properties, with the exception of interfaces shared by cells at the compartment boundary. As we shall show, such passive mechanical interactions are sufficient to explain the robustness of compartment size to hyperplastic manipulations.
Fig 3A shows snapshots of individual simulations of wt, en>dap and en>CycE embryonic segments. We observe cells that are larger but fewer in number in en>dap than in wt, while the en>CycE compartment contains more smaller cells. Generating statistical distributions by running 100 simulations in each case, we obtain the summary statistics visualized in Fig 3B and 3C. To allow for comparison with observed values we superimpose on each panel in Fig 3B and 3C either the upper and lower bounds in observed P compartment areas [14] across the three perturbations (shaded gray) or the upper and lower limits in cell numbers for each perturbation separately (blue, green, red for wt, en>CycE, and en>dap, respectively). We do not plot the distinct shaded regions in the case of P compartment areas since the regions for individual perturbations overlap. Fig 3B shows that, for wt and en>dap, the average P compartment sizes and cell numbers at the end of the final round of divisions predicted by the model closely match observed values. The difference in cell number between simulation and experiment for en>CycE is statistically significant (17%), indicating that the model underestimates the number of cell deaths in this perturbation.
These simulation results were achieved using literature values of the parameters Λ ¯ and Γ ¯ [22], and by assigning daughter cells to have half the target area of their mother cells (RA = 0.5). Although the model is a drastic simplification of epithelial compartment size homeostasis, the in silico results provide a close match to experimental values without any parameter tuning. The model thus provides a simple explanation for the emergence of P compartment size control [14]: size control can be achieved through passive mechanical forces without any further regulation of cellular properties through signaling gradients.
To explore how robust the observed size control is to the model parameters, we performed a single parameter sensitivity analysis while fixing the remaining parameters at their values listed in Table 1 (Fig 3C). For most parameter values considered, the simulation results fall within the bounds of experimentally observed values, except for values of the target area ratio RA smaller than 0.4 and larger than 0.9, and for values in Λ larger than 0.2.
Focusing on the results of en>CycE simulations, the model exhibits some counter-intuitive behaviour. In particular, uniformly low perimeter contractility, Λ ¯, or high line tension, Γ ¯, leads to mechanically induced P compartment shrinkage. In addition, an increase or decrease of RA away from 0.5 will increase compartment sizes for the en>CycE perturbation. We may interpret these results as follows.
Mechanically induced P compartment shrinkage can be understood as a result of the balance between the energy terms in Eq (2). The perimeter contractility and line tension terms act to minimise edge lengths and perimeters of cells. These force contributions can be counteracted by the area term, which acts to keep the cell close to its target area, or by stretching forces exerted by neighbouring cells. Upon division, a new edge is created, which adds an inward contractile force that any expansive forces must counteract. Therefore, daughter cells occupy a smaller area than their mother cell once they reached mechanical equilibrium. The observation that an increased rate of cell division leads to tissue shrinkage is counter-intuitive, yet not unrealistic; data from [14] for en>CycE and en>CycE+p53 embryonic compartments show a similar trend, in which the more cells are present, the smaller the compartment area. Inhibition of cell death in the en>CycE+p53 leads to more cells, but smaller compartments. Further, this counter-intuitive experimental result, which cannot be explained by a simpler hypothesis where EGFR signaling leads to size control through direct patterning of apoptosis and growth, may be explained by a simple mechanical argument.
A similar mechanism explains the dependence of the size of the en>CycE compartment on the target area ratio, RA. Mitosis induced shrinkage is a result of the perimeter contractility and line tension terms in the mechanical model. If we choose a value for RA that is not equal to 0.5, then the target areas of all cells will no longer add up to the total area of the tissue, and more cells have areas that are far away from their actual target areas. This increases the absolute value of the area elasticity term in the energy equation, and hence reduces the relative strength of the perimeter contractility and line tension terms. As the relative strength of these two terms decreases, the extent of mitosis-induced shrinkage is also reduced. In the case RA<0.5, the additional line tension and perimeter force due to the new edge during division are not strong enough to stretch the cells surrounding the division further away from their target area, and if RA>0.5 the forces originating from the new edge are not strong enough to further oppose the strength of the target area terms of the new cells. Hence, mitosis-induced shrinkage occurs only if RA ≈ 0.5. In our simulations, P compartment size is relatively robust to the value of RA, despite the fact that the areas of many cells differ widely from their target values. The bulk elasticity energy term in Eq (2) varies quadratically with deviations between cell area and cell target area. Thus, one might expect significant changes in P compartment areas or cell numbers when target areas are perturbed upon proliferation. Our simulation results suggest that P compartment areas or cell numbers are not affected by such changes in total tissue energy.
A further counter-intuitive result shown in Fig 3C is that increasing the line tension parameter Λ ¯ and increasing the perimeter contractility parameter Γ ¯ have opposing effects on P compartment size in the en>CycE perturbation. Increasing line tension results in a stronger contractile force on the cell, resulting in more T2 transitions and hence a smaller P compartment (Fig 3C, central panel). In contrast, although increasing perimeter contractility also results in a stronger contractile force for each cell, in this case the mechanical interactions between adjacent cells (a contracting cell acts to stretch its neighbours) result in fewer T2 transitions and hence a larger P compartment.
All the observed changes in P compartment sizes and cell numbers remain within experimentally measured values (Fig 3, shaded regions), the exception being the P compartment cell numbers for the en>CycE perturbation. The discrepancy between observed values and in silico results for the P compartment cell numbers in en>CycE is insensitive to parameter variation. The robustness of the simulation results in Fig 3B to parameter values provides further confirmation that size control is a natural outcome of passive mechanical cellular interactions in our model. Size control is preserved in the face of small amounts of cell growth or shrinkage (variations in RA) or perturbations of cellular mechanical properties (variations in Λ ¯ and Γ ¯).
However, this model fails to capture the observed asymmetry in cell death locations, as measured by the ratio of accumulated cell death occurrence between the front and the back half of the P compartment across multiple embryos. The third row of Fig 3C shows that the total number of cell deaths across 100 simulations is the same between the front half and the back half of the P compartment. Here we only plot the cell death occurrences of the en>CycE simulations, since no cell deaths were observed in any wt or en>dap simulations. This is in close agreement with experimental results [14], where only 0.7 (wt) or 0.2 (en>dap) cell deaths where identified by TUNEL staining per embryo.
We draw two main conclusions from the simulations presented in Fig 3: (i) mechanical interactions between identical cells can explain robust size control of all considered genetic perturbations (wt, en>CycE, en>dap), even if the parameters are varied significantly; (ii) passive mechanical interactions of cells with uniform mechanical properties cannot explain the observed asymmetry in cell death occurrence, nor completely recapitulate the changes in cell numbers for the en>CycE perturbation.
We next use the model to analyse how asymmetries in cellular mechanical properties across the P compartment may lead to the observed spatial patterning of apoptosis. We consider three cases (Fig 4A): (i) ‘area regulation’, which refers to patterning of the cell target areas A0,α through the parameter λA; (ii) ‘line tension regulation’, which refers to patterning of the line tension Λ ¯ through the parameter λl; and (iii) ‘perimeter regulation’, which refers to the patterning of the perimeter contractility Γ ¯ through the parameter λp. These parameters are defined in the Materials and Methods section. The ‘area regulation’ scenario can be interpreted as a patterned growth scenario, whereas the ‘line tension regulation’ and ‘perimeter regulation’ scenarios correspond to patterning of cellular mechanical properties. The biochemical process leading to such patterning could, for example, be Spitz-mediated EGFR-activation; this pathway has previously been identified to affect cell properties in the P compartment by Parker [14].
Fig 4B–4D shows the impact of small amounts of asymmetry on P compartment dynamics. In each of the cases (i)-(iii), we set the relevant asymmetry parameter to 0.2, while keeping the other two asymmetry parameters fixed at 0. Fig 4B shows snapshots of simulation outcomes for each case. A visual inspection suggests that these three cases give rise to P compartments with similar cell sizes and shapes as in Fig 3A.
Fig 4C shows that P compartment sizes and cell numbers are not affected by these low levels of asymmetry in the tissue. In each case, the in silico compartment area and cell number is as close to the observed values [14] as the passive mechanical model. Although cellular properties are now patterned, compartment size control still emerges within the model. Fig 4D shows the total number of cell deaths in the front and the back halves of the P compartment across 100 simulations in each asymmetry case. We find that each case can explain the observed spatial asymmetry in cell death locations.
To assess to which extent P compartment sizes and cell numbers are robust to spatial asymmetry in cell mechanical properties, we next vary each of the three asymmetry parameters in turn while keeping the others fixed at 0. Fig 5 shows that increases in asymmetry lead to decreases in P compartment sizes and cell numbers (top and middle row) and the degree of asymmetry in cell death across the front and back halves of the compartment (bottom row).
In the model, P compartment sizes and cell numbers are most sensitive to asymmetry in cell target areas; for example, a value of λA > 0.9 can result in loss of the entire P compartment. In contrast, P compartment sizes and cell numbers remain within experimentally measured regimes for values of λp or λl from 0 up to 0.4.
To identify experimentally observable signatures to differentiate between modes of regulating compartment homeostasis, we examined the distributions of four measures of cellular morphology for the scenarios described in Fig 4. We extract the distributions of cell areas, cell perimeters, lengths of edges between cells, and cell elongations within the P compartment at the end of each simulation. We observe distributions of these four measurements in the posterior compartment as a whole, and in the front and the back half of the compartment separately. The results of this investigation are summarized in Fig 6.
The top two rows of Fig 6 show that the distributions of cell areas and cell perimeters (row 1 and 2) for the area regulation scenario are distinct from the corresponding distributions for the line tension and the perimeter regulation scenario. In particular, the distribution of all areas is bimodal for the area regulation scenario, whereas it is not bimodal for the line tension and perimeter regulation scenarios. A similar distinction can be made for the perimeter distributions, which is bimodal for the ‘area regulation’ scenario and not bimodal for the ‘line tension regulation’ and ‘perimeter regulation’ scenarios. The bimodal distributions are marked by nearly non-overlapping distributions of cell areas and cell perimeters in the front and the back halves of the compartment for the area regulation scenario, whereas these distributions are overlapping in the line tension and perimeter regulation scenarios. Upon decomposing cell area distributions into contributions from the front and back halves of the P compartment, we see that the mean cell area is different between these two halves in the area and perimeter regulation scenarios, and the same holds for the cell perimeter distributions. Cell elongations and edge lengths have similar shapes and mean values for all three asymmetry scenarios (rows 3 and 4 of Fig 6).
The results in Fig 6 suggest that it is possible to distinguish between the ‘area regulation’ scenario (differential growth across the compartment) from the ‘line tension regulation’ scenario (regulation of apical mechanical properties) by measuring the distributions of cell areas or perimeters in the front and the back halves of the posterior compartment separately. The distribution of cell areas or perimeters across the P compartment may further allow one to distinguish the ‘area regulation’ scenario from the ‘perimeter regulation’ scenario, since this distribution is bimodal in the former scenario, but not clearly bi- or unimodal in the latter. However, multiple sources of noise in an experimental setup may make this distinction between the ‘area regulation’ and ‘perimeter regulation’ scenarios less clear. Measuring edge lengths or cell elongations will not reveal differences between the scenarios.
While cell area distributions in wt simulations may allow the different asymmetry scenarios considered to be distinguished from one another, these distributions in the en>dap and en>CycE cases provide model predictions that are preserved across all scenarios. In each case, we find that the cell area distribution is multimodal. In particular, the en>dap cell area distribution is trimodal in the ‘area regulation’ scenario, whereas it is bimodal in the other cases considered.
This multi-modality in areas arises from overlapping cell generations. Since we assume that cell target areas decrease upon division (RA < 1), each successive generation of cells will have a smaller target area. In simulations of the en>CycE perturbation, some cells divide twice while others only divide once, resulting in a bimodal cell area distribution. Similarly, for the en>dap perturbation, some cells divide once while others don’t divide at all; however, we also observe area differences between cells in the front and the back half of the P compartment (Fig 7). These effects combine to yield a trimodal cell area distribution.
In summary, the area distributions of the genetic perturbations en>dap and en>CycE may be used as a measure to validate the model assumptions, and provide a further tool to distinguish the ‘area regulation’ scenario from the ‘perimeter regulation’ and ‘line tension regulation’ scenarios.
As a further analysis of the model, we performed a laser ablation analysis on the final configuration of our wt, en>dap and en>CycE simulations. In 100 simulations for each perturbation, we ‘cut’ a randomly selected cell-cell interface (edge) in the P compartment. This was implemented by setting the line tension parameter Λ ¯ for this edge, as well as the perimeter tension parameter Γ ¯ for the cells adjacent to the edge, to zero. We then ran each simulation for 200 further time units and recorded the average initial vertex recoil velocity and total vertex recoil distance. Results for each of the three asymmetry scenarios are shown in Fig 8. We find that under the ‘perimeter regulation’ scenario, the average initial vertex recoil velocity and total vertex recoil distance are both smaller in each perturbation than in wt. In contrast, under the two other asymmetry scenarios there is no significant difference in these statistics across wt, en>dap and en>CycE simulations. These results offer a further experimentally testable prediction that, in conjunction with the cell area distribution results summarised in Fig 6, allows for discrimination between the three asymmetry scenarios considered.
We have employed a vertex model of a Drosophila embryonic segment to test hypotheses about the emergence of size control. A comparison of the in silico segment with extant literature values indicated that passive mechanical forces suffice to explain the observed size control. However, the observed spatial heterogeneity in cell death frequencies requires some form of patterning of mechanical properties across the tissue. Several conceptually distinct modifications of the model can explain size control while also recapitulating the spatially varying rates of cell death: first, individual cells could regulate their sizes through differential growth; and second, cells could regulate their apical mechanical properties through differential expression of tension regulating protein activities. It is possible to distinguish these two scenarios within the model by the spatial distribution of P compartment cell areas and perimeters, as well as by the speed of vertex recoil after laser ablations. These results hint at two possible mechanistic functions of trophic signaling pathways, such as EGFR or Wg [14, 36, 37]: they could either cause growth of individual cells, or else modulate cell shape through regulation of contractile cytoskeletal activity, either of which would explain the experimentally observed shrinkage or growth when the pathways are perturbed [14].
Understanding the mechanism of tissue size control is particularly challenging due to the interconnected and complex nature of cell signaling and the high degree of feedback between cell- and tissue-level processes. Computational models therefore offer an important tool for investigating and testing hypothesised mechanisms and to abstract the principles underlying developmental robustness [38–40].
The development of multicellular organisms requires control of total cell numbers and relative proportions of cell types with tissues. Size control can be divided into two steps: initial specification and maintenance [12]. Much work has focused on the regulation of the position of cellular fates during early embryonic development. Traditionally, tissue size specification has been associated with signaling gradients [41–43]. However, the mechanisms that ensure the maintenance of tissue size and of boundaries between tissues is less well understood. In particular, the physics of size homeostasis for patterned epithelia are not well understood, yet they are a recurring theme in development [44, 45] and it is increasingly recognised that mechanical feedback plays a role in controlled tissue behaviour [8, 46].
A gradient growth model has previously been proposed for the regulation of P compartment size in the Drosophila embryonic epidermis [14]. This conceptual model requires the correct maintenance of a morphogen gradient in the face of multiple genetic perturbations. The present study demonstrates that an alternative, passive mechanical model can partially explain robustness of P compartment sizes and cell numbers in the Drosophila embryonic epidermis, eliminating the need for a tightly controlled intermediary morphogen gradient. More detailed cell-level analysis and modelling is required in the future to fully understand how morphogen signals are established, maintained, and interpreted [47, 48], especially in the face of genetic or environmental perturbations.
Advancing our knowledge of how embryos achieve robustness to defects or damage to the initial patterning of tissue domains is important for understanding the underlying causes of birth defects, as well as diseases with an underlying basis of misregulated growth, such as cancers.
Although several studies have investigated the robustness of sizes of patterned epidermal segments of Drosophila, quantification has been somewhat sporadic and diffuse. This will in general require a thorough systems-level characterization of later stages of Drosophila morphogenesis for multiple experimental perturbations. The present study provides a basis for guiding future experiments that seek to identify possible modes of size control in late stages of epithelial development in Drosophila.
How could model predictions be validated against such experiments? Several previous vertex models of developing epithelia have been validated against key summary statistics. Such studies have focused primarily on the Drosophila wing imaginal disc, which undergoes up to 9 rounds of divisions to arrive at a distinct distribution of cell polygon numbers [22, 24]. In these studies, it is safe to assume that the initial distribution of cell polygon numbers will not affect simulation outcomes, due to the high levels of proliferation. Here, we considered one or two rounds of divisions; over such a short developmental timespan we expect the initial sheet topology to influence final polygon distributions. Hence, for a quantitative comparison of this summary statistic between model and data, experimentally informed cell shapes in late stage 10 segments may be required. Such summary statistics remain lacking for the Drosophila embryonic epidermis during its development, and poses an experimental challenge due to the small system size (20–60 cells). Large sample sizes will be required to obtain accurate distributions of cell polygon numbers. Figs 7 and 8 in this study suggest that distributions of cell areas, and characterization of vertex recoils following standard laser ablation experiments, for genetic perturbations of the P compartment maybe used to validate the underlying computational model. Thus, future iterations of the model may be further constrained through inference of mechanical parameters from laser ablation [6] or less invasive experimental protocols [49].
Embryogenesis is an extremely complex process. To make headway into the factors that influence robustness of tissue size maintenance, there needs to be conscious decoupling and abstraction through studies of simpler systems. This is also part of the rationale for studies in genetic model organisms from the worm and fly to mouse [40].
Due to the lack of kinematic data on cell shapes and compartment sizes during the latter stages of embryogenesis, we have not included an analysis for this initial study and have focused on more local mechanisms. In particular, we assumed that the overall tissue dimensions are constant during the considered time frame, since the epidermis forms at the outside of the embryo during stage five of Drosophila development and as a whole does not change dimensions for the remainder of development. However larger scale tissue morphogenetic movements, which are undoubtedly important for aspects of morphogenesis [50], may affect the exact size of a given subsection of the tissue. For example, dorsal closure occurs during the considered time frame, which leads to an extension of the tissue that we study [51]. The assumption that this extension should not affect the relative proportions in A and P compartment size requires future experimental validation. In addition, our finding that elevated tension along compartment boundaries does not affect compartment sizes may be contrasted with theoretical and experimental studies showing how differential line tension, either at compartment boundaries or across tissues, may drive convergent extension [52, 53]. A key conceptual difference between the present work and these studies is the assumption of a fixed, or free, boundary to the tissue.
In vertex models with a free boundary, contractile forces along cell perimeters may lead to deviations of cell areas from their respective target areas. The analysis of simulations with changed initial target areas presented in S4 Fig reveals that such deviations between cell target areas and their absolute areas may lead to increases in predicted apoptotic rates. Further investigation is required to understand the boundary conditions that best represent the effect of adjacent tissues in different epithelia, and the effects that forces along tissue boundaries can have on different summary statistics. It may be possible to gain insights to this question by quantifying tissue-level kinematics of germ-band retraction for the wt and developmental perturbations.
Our model relies on the quasi-steady state assumption that the tissue is at mechanical equilibrium at each time point. We justified this assumption on the basis that individual cell cycle times of the 16th division cycle in Drosophila development are around an hour [9]. However, if cell divisions occur highly synchronously, then, this assumption might not hold. In en>CycE embryonic segments, the numbers of cell divisions events in the model were inferred from data where apoptosis was blocked by expressing the protein p35 in the P compartments [14]. It has previously been reported that epithelial sheets can extrude cells that are not undergoing apoptosis [29]; if this occurs to a great extent in the Drosophila embryonic epidermis, then our inferred numbers of mitotic events would require adjustment. In this case, an in vivo cell tracking study would be necessary to measure the levels of cell division and extrusion events. Such data would also help to shed light, for example, on the possible impact of mitotic cell rounding on local cell shapes and possible short-range correlations between mitosis and apoptosis events. Since apoptosis in the vertex model is a passive process, we cannot extend our model analysis to p35 mutants in which apoptosis is blocked. How to adapt vertex models in such a way as to prevent the occurrence of T2 transitions while maintaining tissue integrity remains an open question.
Due to a current lack of data in the literature, our model does not include a description of upstream patterning of cell types. Instead, we infer the necessity of patterning of cell mechanics through simulations. This study is timely as it provides some guidance into important parameters and considerations that should be taken into account in future quantitative analyses of late stages of epidermal development including germband retraction and head involution. From the results presented here, further questions arise. If a passive mechanical model is sufficient to explain compartmental size control, then what is the functional role of Spitz-mediated EGFR regulation? It is known that EGFR signaling is required for dorsal closure during Drosophila development [54]. Hence, it is possible that the influence of EGFR signaling on larval compartment sizes reflects the role of EGFR signaling in convergent extension during dorsal closure. If the asymmetry in our model reflects patterning of mechanical properties through trophic signaling, then a more detailed experimental analysis of the spatio-temporal dynamics of cellular signaling will allow more detailed modelling of how these properties may be patterned.
Our study serves as an example of using computational models as an abstraction of the maintenance of tissue sizes with implications for a broad range of studies. Significant advances in stem cell engineering have resulted from understanding how to unlock the potential for multicellular aggregates to self-organize. Recent examples include the morphogenesis of optic eye cups in organ culture conditions [55] and the engineering of beating mini-hearts [56]. We posit that great success in developing tissue repair strategies will come through the reverse engineering of pattern repair mechanisms in situations where pattern repair is perturbed. Such reverse engineering will require guiding experimental efforts through modelling studies that identify the information needed to distinguish between mechanisms.
|
10.1371/journal.pgen.1008188 | A telomerase with novel non-canonical roles: TERT controls cellular aggregation and tissue size in Dictyostelium | Telomerase, particularly its main subunit, the reverse transcriptase, TERT, prevents DNA erosion during eukaryotic chromosomal replication, but also has poorly understood non-canonical functions. Here, in the model social amoeba Dictyostelium discoideum, we show that the protein encoded by tert has telomerase-like motifs, and regulates, non-canonically, important developmental processes. Expression levels of wild-type (WT) tert were biphasic, peaking at 8 and 12 h post-starvation, aligning with developmental events, such as the initiation of streaming (~7 h) and mound formation (~10 h). In tert KO mutants, however, aggregation was delayed until 16 h. Large, irregular streams formed, then broke up, forming small mounds. The mound-size defect was not induced when a KO mutant of countin (a master size-regulating gene) was treated with TERT inhibitors, but anti-countin antibodies did rescue size in the tert KO. Although, conditioned medium (CM) from countin mutants failed to rescue size in the tert KO, tert KO CM rescued the countin KO phenotype. These and additional observations indicate that TERT acts upstream of smlA/countin: (i) the observed expression levels of smlA and countin, being respectively lower and higher (than WT) in the tert KO; (ii) the levels of known size-regulation intermediates, glucose (low) and adenosine (high), in the tert mutant, and the size defect’s rescue by supplemented glucose or the adenosine-antagonist, caffeine; (iii) the induction of the size defect in the WT by tert KO CM and TERT inhibitors. The tert KO’s other defects (delayed aggregation, irregular streaming) were associated with changes to cAMP-regulated processes (e.g. chemotaxis, cAMP pulsing) and their regulatory factors (e.g. cAMP; acaA, carA expression). Overexpression of WT tert in the tert KO rescued these defects (and size), and restored a single cAMP signaling centre. Our results indicate that TERT acts in novel, non-canonical and upstream ways, regulating key developmental events in Dictyostelium.
| When cells divide, their chromosomes are prone to shrinkage. This risk is reduced by an enzyme that repairs protective caps on each chromosome after cell division. This enzyme, telomerase, also has several other important but unrelated roles in human health. Most importantly, via one or other of its functions, both high and low levels of the enzyme can contribute to cancer. We have studied, for the first time, the roles played by telomerase in the life-cycle of the cellular slime mould, Dictyostelium discoideum, a model system with a rich history of helping us understand human biology. While we did not find any evidence of telomerase having the features typically needed to repair a chromosome, telomerase was necessary for many aspects of development. The Dictyostelium telomerase mutant we generated shows delayed aggregation and forms irregular fruiting bodies. The tert mutant miscalculates, in effect, how big those fruiting bodies should be, and they end up being too small. These results are significant because they show, for the first time, that a telomerase can influence tissue size regulation, a process central to a wide range of cancers.
| Each time a chromosome replicates, it loses some DNA from each of its ends. This is not necessarily problematic, because the chromosome is initially capped at each end by a sacrificial strand of non-coding DNA, a telomere [1–3]. Further instances of replication, however, can expose the coding DNA, unless the cell can keep repairing the shortened telomeres, by the action of the enzyme complex, telomerase. Accordingly, telomerase, whose main subunits comprise a reverse transcriptase (TERT), and the telomerase RNA component (TERC) [4], has much significance in the biology and pathology of multicellular organisms. As somatic tissues age, for example, telomerase is downregulated, and the resulting telomeric dysfunction can lead to chromosomal instability and various pathologies, including disrupted pregnancies and cancer [5–7]. In other cases, the upregulation of telomerase is also associated with, and a biomarker of, some cancers, because it allows the unchecked proliferation of immortalised tumour cells [6, 8]. Telomerase also has many non-canonical roles, in which telomere maintenance, or even telomerase activity, is not required [9, 10]. For example, telomerase is known to have non-canonical roles in neuronal differentiation [11], RNA silencing [12], enhanced mitochondrial function [13], cell adhesion and migration [14, 15] and various cancers [9, 16].
Our understanding of telomeres and telomerase began, and has continued to develop, through the study of model organisms such as Drosophila, Zea mays, Tetrahymena, yeast and mice [2, 3, 17–21]. One model system in which the possible roles of telomerase have not yet been addressed is Dictyostelium discoideum. This system has proved its usefulness in many contexts, including the study of human diseases [22–26]. One of its advantages is that the processes of cell division (i.e. growth) and development are uncoupled [27], making the organism a highly tractable system for the study, in particular, of differentiation and tissue size regulation [28–35]. In culture, when its bacterial food source is abundant, D. discoideum multiplies as single-celled amoebae. This leads to denser colonies, and exhaustion of the food supply. The rising concentration of a secreted glycoprotein, CMF, triggers the organism to switch to a multicellular mode of development [34, 36]. With no resources for further cell proliferation, the amoebae move, in a radial pattern of streams, towards centres of aggregation. Rising levels of secreted proteins, of the counting factor (CF) complex [37, 38], trigger a series of changes that lead to breaking up of the streams, which therefore no longer contribute cells to the original aggregate. Each aggregate, which will typically contain 20, 000 to 100,000 cells [39], now rounds up into a mound, which then proceeds through several life-cycle stages, finally forming a spore-dispersing fruiting body about 1-2mm high [34, 40]. Mounds can also develop from the breaking-up of a large stream (or aggregate), a process similarly regulated by CF [29, 41]. The generic term, ‘group’, can be used to address the fact that mounds develop from clusters that arise in these slightly different ways, but in this paper we will refer to ‘mounds’. Some of the processes and regulators involved in our very abbreviated account of the life-cycle are shown in Fig 1, which focuses on those elements examined in this study.
In addition to being uncoupled from growth, development in D. discoideum has other features that make it potentially useful as a model system for the understanding of telomerase-based pathologies, in particular cancers that arise from disruption of non-canonical functions. First, as indicated in Fig 1, development in D. discoideum depends on properly regulated cell motility and cell adhesion, two processes fundamental to metastasis. Second, the switch to multicellular development, and the control of aggregate, mound and hence fruiting body size are influenced by various secreted factors that, respectively, promote aggregation and regulate tissue size, in ways analogous to the regulation of tumour size by chalones [42, 72]. Third, a putative TERT has been annotated in the D. discoideum genome. It is not known if the RNA component of telomerase (TERC) is present [73] and, in any case, extrachromosomal rDNA elements at the ends of each chromosome in D. discoideum suggest a novel telomere structure [74]. Thus, telomerase in this organism may have a separate mechanism for telomere addition or might have non-canonical roles. As yet, however, there have been no functional studies of TERT reported for D. discoideum.
In this study, we characterize the gene tert in D. discoideum, showing that it has both RT and RNA binding domains. We describe the pattern of tert’s expression levels during all stages of development, assay for any canonical telomerase function, and examine the effects of knocking out the gene’s function on development. The tert mutant exhibits a wide range of developmental defects, suggesting that wild-type TERT targets multiple elements of the regulatory network depicted in Fig 1. Most interestingly, these defects, and the results of experiments by which we attempt to rescue, or phenocopy, the tert KO phenotype, suggest that this telomerase influences the activity of smlA, and processes downstream of it. Tert thus emerges as one of the upstream genes of the cell-counting pathway, and its overall influence indicates that, despite having no obvious canonical activity, a telomerase can nevertheless play major regulatory roles by virtue of its non-canonical targets.
Extending previous predictions of tert encoding a protein with telomerase motifs [75], our use of the Simple Modular Architecture Research Tool (http://SMART.embl-heidelberg.de) and UniProt (Q54B44) revealed the presence of a highly conserved reverse transcriptase domain and a telomerase RNA binding domain (S1 Fig). These are characteristic of a telomerase reverse transcriptase [76], supporting the idea that the gene we characterized indeed encodes for TERT. The Dictyostelium TERT protein shares 23% and 18.7% identity with human and yeast TERT protein respectively (Pairwise sequence Alignment-Emboss Needle). The protein sequence identities between the TERT of D. discoideum and five other species are tabulated in S1 Table. In the case of the identity with the TERT of humans, the strongest homologies are seen in the reverse transcriptase domain. We did a phylogenetic analysis to examine the relatedness of DdTERT with that of other organisms. For this, TERT amino acid sequences from different organisms were obtained from the NCBI database or Dictybase (http://www.dictybase.org/) or SACGB database (http://sacgb.leibniz-fli.de) and compared with TERT of D. discoideum. Multiple sequence alignment of the TERT amino acid sequences of various organisms including other social amoebae were used to create the phylogenetic tree, employing the MUSCLE alignment feature of MEGAX software [77]. The phylogenetic analysis suggests that D. discoideum TERT falls in a separate clade and is likely to be a distant relative of vertebrate homologs (S2 Fig). The evolutionary history was inferred using the Neighbor-Joining method [78]. The evolutionary distances were computed using the p-distance method and the units shown are the number of amino acid differences per site.
Further, using the fold recognition technique on the I-TASSER server, the structure of D. discoideum TERT was predicted using Tribolium castaneum (telomerase in complex with the highly specific inhibitor BIBR1532; PDB-5cqgA) as a template (S3 Fig). The modeled structure of Dictyostelium TERT also suggests that D. discoideum has a structurally conserved TERT (S3 Fig).
Telomerase activity, if any, can be ascertained by performing a Telomeric Repeat Amplification Protocol (TRAP) assay, and activity has been successfully detected in organisms such as humans, C. elegans, yeast, Daphnia, and plants [79–84]. However, while human cell lines (HeLa, HEK) did show telomerase activity, we did not detect any telomerase activity in D. discoideum cell extracts (S4 Fig). This concurs with previous findings, namely that the telomeres of D. discoideum have a novel structure [85], and that, in other organisms, TERT has several non-canonical roles [11–13].
In humans, telomerase expression is reported to be low in somatic cells compared to germline and tumour cells [86]. To ascertain if tert expression is differentially regulated during growth and/or development, we performed qRT-PCR using RNA from different developmental stages (0, 4, 8, 10, 12, 16 and 24 h after starvation). Tert expression is higher in development than during growth, (8h and 12 h) (Fig 2), implying that tert plays a prominent role beyond the point at which D. discoideum is responding to starvation. Expression also shows a marked biphasic pattern, with the first peak at 8h (when streams are forming), a big dip during stream breaking (10h) and then rising gradually again to peak at about the time of mound formation (12h).
To understand the possible non-canonical roles of tert in development of D. discoideum, tert KO cells generated by homologous recombination were seeded at a density of 5x105 cells/cm2 on non-nutrient buffered agar plates and monitored throughout development. While aggregates appeared by 8 h in the wild-type, and streams began to break at 10 h, in the mutants there was a further 8 h delay before aggregates were seen, and stream breaking began at about 18 h. Because of these delays, ‘during aggregation’, in this study, refers to 8 h in WT and 16 h in the tert KO, and ‘during stream breakup’ refers to 10 h in WT and 18 h in the tert KO.
Wild-type cells formed long streams of polarized, elongated cells leading to aggregation, but tert KO cells did not form well-defined streams, failing to aggregate even at 5x104 cells/cm2 (wild-type cells aggregated even at a density of 2x104 cells/cm2), suggesting an inability to respond to aggregation-triggering conditions (S5 Fig). The mutant’s streams were also larger (Fig 3A). In contrast to streams moving continuously towards the aggregation centre in WT, tert KO streams break while they aggregate (S1 and S2 Videos). They did eventually form aggregates, largely by clumping. During the early stages of aggregate formation, the number of aggregation centres formed by the tert KO was only 10% of that formed by WT (Fig 3B, p<0.0001). Due to uneven fragmentation, the late aggregates were also of mixed sizes. The tert KO cells did eventually form all of the typical developmental structures, but by the mound stage, continued fragmentation had resulted in the mounds being more numerous, and smaller, on average, than in the WT. This was also the case for fruiting bodies.
Thus, with reference to Fig 1, tert appears to play roles in multiple aspects of Dictyostelium development: the timing of aggregation; streaming; and the regulation of the size of the mound and fruiting body (Table 1A and 1B).
Given the wide-ranging phenotypic defects seen in the tert KO, it seemed likely that tert is one of the key regulators of development in D. discoideum, affecting many of the processes and regulators depicted in Fig 1. We thus monitored the activity or levels of a number of those elements, comparing the wild-type and tert KO (summarised in Table 1A and 1B). As that summary shows, the tert KO showed significant changes from the wild-type in three broad areas: components of the mound-size regulation pathway; cAMP-related processes/regulators; and adhesion-related processes/regulators. As is clear from Fig 1, the factors that influence these features overlap considerably, both in terms of interacting with each other, and in regulating more than one of the various developmental stages disrupted in the tert KO. Nevertheless, we think it is useful to consider each of them in turn. As we do so below, we describe a series of experiments that largely fall into two broad categories, as shown in summary form in Tables 2 and 3: Those that attempt to rescue the normal phenotype in tert KO cells (Table 2); and those that attempt to phenocopy, or induce, the tert KO phenotype in wild-type cells (Table 3). First, however, we describe some experiments that support the direct involvement of tert in the effects already noted.
To support the idea that the changes observed in the tert KO are, in the first instance, due to changes involving tert itself, and not some other factor, we took two approaches: Overexpression of tert, and the use of TERT inhibitors. Most importantly, overexpression of wild-type TERT (act15/gfp::tert) in tert KO cells rescued all three of the phenotypic defects (Fig 4A, S3 Video; Table 2), suggesting that the tert KO phenotype is not due to any other mutation. Next, we treated wild-type cells with structurally unrelated TERT specific inhibitors, BIBR 1532 (100nM) and MST 312 (250nM). BIBR 1532 is a mixed type non-competitive inhibitor, whereas MST 312 is a reversible inhibitor of telomerase activity (see Methods). Both inhibitors strikingly phenocopied two features of the tert mutant, in that we observed large early aggregate streams that broke and eventually resulted in mounds (Fig 4B; Table 3) and fruiting bodies that were small. The developmental delay, however, was not induced. Since the two inhibitors phenocopied the tert KO to a remarkable degree, it is likely that the inhibitor binding sites of Dictyostelium TERT are conserved. Human TERT [87], which shares a 23% homology with Dictyostelium TERT, failed to rescue the tert KO phenotype (S6 Fig). Surprisingly, the morphologies of TERT-overexpressing lines in the wild-type did not show any significant difference to those of the untreated wild-type (Fig 4A).
Overall, these results strongly support the idea that the relevant changes in the tert KO involve tert itself. The fact that the TERT inhibitors induced only two of the three tert KO defects is interesting. Given the lack of any apparent interconnection between the pathway that regulates the switch to aggregation, and that regulating mound size, it seems likely that TERT acts on more than one molecular target. It could be that the inhibitors do not perturb that part of TERT that interacts with the target that regulates the switch to development.
Given the perturbations seen in the tert KO, one would predict some abnormalities associated with cAMP dynamics [44–46, 88–90]. The role of cAMP in streaming, in particular, has been much studied. Below we examine how various cAMP processes or factors, related to streaming and developmental delay, were affected in the tert KO.
Cell-substratum adhesion is also important for migration and proper streaming. By shaking cells at different speeds (0, 25, 50 and 75 rpm), it is possible to vary substratum dependent sheer force. Thus, by counting the fraction of floating cells at different speeds, it is possible to check substratum dependent adhesion. Although both wild-type and tert KO cells exhibited a sheer force-dependent decrease in cell-substratum adhesion, tert KO cells exhibited a significantly weaker cell-substratum adhesion (S12 Fig, p<0.0001), affecting cell motility in a way that might also contribute to stream breaking.
Cell-cell adhesion is also an important determinant of streaming and mound size in Dictyostelium [41]. To examine if adhesion is impaired in the mutant, we checked the expression of two major cell adhesion proteins, cadA, expressed post-starvation (2 h) and csaA expressed during early aggregation (6 h). cadA-mediated cell-cell adhesion is Ca2+-dependent and thus EDTA-sensitive, while csaA is Ca2+ independent and EDTA-resistant [67]. Both csaA and cadA expression were significantly down-regulated (Fig 14A and 14B).
Further, cell adhesion was monitored indirectly by counting the fraction of single cells not joining the aggregate. Aggregation results in the gradual disappearance of single cells and thus it is possible to measure aggregation by determining the ratio of single cells remaining. To examine Ca2+-dependent cell-cell adhesion, the assay was performed in the presence of 10 mM EDTA. Both EDTA-sensitive and resistant cell-cell adhesion were significantly defective in tert KO cells (Fig 14C, p = 0.0033 and 14D, p = 0.0015). The levels of csaA and cadA were also lower in the tert KO during aggregation when compared to the WT (Fig 14E, p = 0.0037 and 14F, p = 0.0508). Thus, the delay in tert KO development might be the basis for differences in gene expression.
These results imply that defective cell-substratum and cell-cell adhesion might play roles in the abnormal streaming and mound-size regulation of the tert KO.
One interesting observation was that the only treatment that fully rescued the tert KO cells was the overexpression of wild-type tert. Also, the only other treatment that rescued the developmental delay itself was mixing wild-type cells with the tert KO cells at a 1:1 ratio (Fig 15; Table 2). Even though caffeine and glucose rescued streaming and mound size, and apparently this was at least partly mediated via their impact on cAMP-regulated processes, neither of the compounds rescued the delay, even though abnormalities of cAMP-regulated processes are commonly reported causes of delay in other Dictyostelium studies [44–46].
Thus, we examined polyphosphate levels in the tert KO because of their known importance to developmental timing in Dictyostelium [43]. We stained the CM with DAPI for 5 minutes and checked the polyphosphate specific fluorescence using a spectrofluorometer. The CM of tert KO cells has reduced polyphosphate levels (49.55±2.02 μM) compared to wild-type (60.62±1.95 μM), implying that low polyphosphate levels might also contribute to the delay in initiating development in this system (Fig 16, p = 0.0009).
Our results reveal that TERT plays an important role in many aspects of Dictyostelium development. The tert KO exhibited a wide range of developmental defects. Despite suitable environmental conditions for multicellular development to begin, the start of the streaming phase is delayed by 8 h. Having once begun, development proceeds and ends abnormally, with large streams, uneven fragmentation, and, eventually, small mounds and fruiting bodies. The wide-ranging developmental defects are associated with changes to the levels, or expression, of genes and compounds that are known to be highly upstream regulators of the various stages of development, such as streaming and mound/fruiting body formation. Based on the perturbations in the tert KO, and our other experiments, Fig 17 depicts the possible extent of processes, and potential mediating factors, that might depend upon normal tert expression/TERT activity in the wild-type. Note that the arrows that connect tert/TERT to any element in the diagram are not meant to suggest that TERT directly regulates that element, only that TERT is important, perhaps in some indirect way, for the normal levels, or activity, of that element.
One of the most striking findings was that TERT appears to regulate, or is at least necessary for, the normal activity of what was previously known as the most upstream regulator of mound size, smlA [28, 30, 32]. Expression levels of smlA were reduced in the tert KO, and we also observed a wide variety of the expected downstream effects of lowered smlA levels. All of these, and a wide variety of treatments that rescued the size-defect of the mutant phenotype, support the idea that the reduction of mound size in the tert KO was indeed mediated via the abnormal functioning of the previously-identified elements of the mound-size regulation pathway.
In addition to the rescue approach, treatments that attempted to phenocopy the tert KO phenotype in the wild-type, also suggest TERT is one of the upstream regulators of mound size. In particular, given that size regulation in D. discoideum depends upon secreted factors of the CF complex, one would have predicted the effects we observed when tert KO CM was added to wild-type cells. Another strong indication that tert acts upstream, at least of CF, was that the inhibition of tert activity in countin mutants failed to phenocopy the tert KO phenotype.
A similarly rich range of results (involving the tert KO phenotype, and its rescue, and phenocopying) support the idea that TERT also plays a high-level role in the regulation of streaming. During the streaming phase, two genes associated with cAMP related-processes in D. discoideum (acaA, carA) were significantly downregulated (compared to the wild-type), and the levels of several other genes trended lower. This was also accompanied by lower cAMP levels. This might explain the defective chemotaxis and cell motility of the tert KO.
Of course, the regulation of streaming is not entirely isolated from that of size. Glucose, one of the central elements of the CF pathway, influences several cAMP-related processes [64]. Thus, it was not surprising that adding 1 mM glucose to the tert KO cells rescued both the size and streaming defects. This study, however, provided a new insight into how the rescue of streaming occurs, because added glucose also reduced adenosine levels. Thus, in the tert KO, the low glucose levels might lead to higher adenosine levels, allowing it to inhibit cAMP related processes (via pathway i, Fig 17). In normal development, given the known sequence of the telomere repeats of D. discoideum (A-G(1–8); [74]), and the fact that telomerase activity would therefore recruit cellular stores of adenosine, it is possible that normal TERT activity keeps adenosine levels low. As yet, however, whether TERT actually acts as a functional telomerase in D. discoideum is not known.
The tert gene we characterized includes the conserved domains and structure of a telomerase reverse transcriptase. Also, supplementing structurally unrelated but specific inhibitors of TERT to wild type cells phenocopies the mutant phenotype. The widely used method to test telomerase activity is the TRAP assay. However, this method failed to detect telomerase activity in D. discoideum and there may be both technical and innate limitations. For example, possible reasons for the lack of any observed activity are that: (i) the presence of rDNA palindrome elements in the chromosomal ends, suggesting a novel telomere structure and the possible role of TERT in maintaining both rDNA and chromosomal termini [74]. This could be an alternate pathway of telomere maintenance in D. discoideum; and (ii) polyasparagine repeats, present in the TERT protein of Dictyostelium, splitting the functional domain into two halves. For telomerase activity, a functional TERT is important in humans [100–102]. In yeast as well as humans, truncation of one of the TERT protein domains is known to abolish its function [103, 104]. While it is not yet clear whether the apparent absence of canonical TERT function in D. discoideum is due to the absence of normal eukaryote telomeres [105], other studies suggests that TERT is not always associated with telomerase activity. The silkworm genome contains a telomerase gene, but the telomerase itself displays little or no enzymatic activity [106, 107]. The telomeres of silkworm consist of the telomeric repeats typical of insects, but also harbor many types of non-LTR retrotransposons [106, 108, 109]. Also of interest is that species of Calcarea (sponges), Cnidaria (sea anemones and jellyfish) and Placozoa, all have metazoan telomeric sequences, but display little or no telomerase activity [110]. D. discoideum might employ an alternative mode of telomere addition, such as the recombination seen in yeast [111] or the retrotransposition of Drosophila [112, 113].
The discussion so far, while it establishes that TERT is needed for several developmental processes to take place, does not help to distinguish whether or not it acts more than once, or if it has more than one target. Could TERT for example act more like the much studied homeodomain proteins, master regulators of animal development, but which only act during very early embryological life [114, 115]? Likewise, in D. discoideum, CMF appears to act only once [34]. Two lines of argument suggest that TERT is different.
First, the biphasic nature of tert's expression pattern suggest that it could possibly act during two stages of development. In the wild-type, tert expression builds up to its first peak at 8 h, thus being a potential candidate for enabling streaming to begin, and to proceed correctly, around this time. It then dips markedly to a low point at 10 h, whereby it might help to enable stream break-up by its relative absence. Then, it begins its climb to its second peak at 12 h, when mound size is being finalised. However, it is also possible that the later-occurring defects seen in the tert KO correspond to pleiotropic effects of TERT being absent at a much earlier time-point.
Second, while it is well known that cAMP-related processes play important roles in allowing streaming to begin and to proceed properly, and while we have shown that TERT influences multiple cAMP related processes, the pathway by which TERT influences the initiation of streaming seems distinct from that used for maintaining it. Both glucose and caffeine, for example, rescued the streaming and size defects of the tert KO, but the delay was unaffected. Complementarily, when wild-type cells were mixed at 50% with tert KO cells, they rescued the delay defect only. In fact, the only treatment that fully rescued the tert KO was the overexpression of wild-type tert.
Interestingly, MAP kinase kinase (MEK1) disruption results in a stream-breaking phenotype similar to the tert KO [56], suggesting that MEK1 could be involved in either CF secretion or signal transduction. Also, signals transmitted through p38 mitogen-activated protein kinase (MAPK) regulate hTERT transcription in human sarcoma [116]. We speculate that MEK1 might regulate countin levels through TERT, thus helping to regulate tissue size in D. discoideum.
Also, it is known that MST 312 (a TERT inhibitor) treatment reduces tumour size by 70% in a mouse xenograft model and this inhibition preferentially targets aldehyde dehydrogenase-positive cancer stem cell-like cells in lung cancer [117]. In Dictyostelium, disruption of aldehyde reductase increases group size [118] and, since aldehyde dehydrogenase and aldehyde reductase have opposing activities (oxidation and reduction of aldehydes respectively), they might have opposite functions in group size regulation as well. TERT might possibly be regulating aldehyde reductase activity in determining mound size in D. discoideum.
Other genes are also known to play a significant role in aggregate size determination in Dictyostelium, such as dio3 [119] and pkc [120]. However, it is not known if they interact with TERT in determining mound size.
This study indicates for, the first time, that TERT acts in several non-canonical ways in D. discoideum, influencing when aggregation begins, the processes involved in streaming, and the eventual size of the fruiting body. TERT's influences appear to occur upstream of many other regulators of streaming and fruiting body size. Curiously, as yet we have no evidence that TERT acts as a canonical telomerase, nor is it known whether any other enzyme protects the unusually sequenced telomeres of this species. Given that telomere research is still in progress, we cannot even rule out that TERT’s apparently non-canonical roles in D. discoideum development are in fact mediated via some as-yet unidentified action on its unusual telomeres. In the most heavily studied stages of the organism’s life-cycle, that is, those that occur in response to starvation, replication has ceased, so further study of this particular point should focus on the amoeboid stage. More generally, this study has revealed a previously unreported non-canonical process influenced by a telomerase, tissue size regulation. This role of TERT, together with its influence on cell motility and adhesion, and the levels of chalone-like secreted factors, bear consideration by those engaged in cancer research.
Wild-type D. discoideum (AX2) cells were grown with Klebsiella aerogenes on SM5 plates, or axenically, in modified maltose-HL5 medium (28.4 g bacteriological peptone, 15 g yeast extract, 18 g maltose monohydrate, 0.641 g Na2HPO4 and 0.49 g KH2PO4 per litre, pH 6.4) containing 100 units penicillin and 100 mg/ml streptomycin-sulphate. Cells were also grown in Petri dishes as monolayers. Other dictyostelid species (D. minutum and D. purpureum) were grown with Klebsiella aerogenes on SM5 plates and cells were harvested when there was visible clearing of bacterial lawns.
To trigger development, cells were washed with KK2 buffer (2.25 g KH2PO4 and 0.67 g K2HPO4 per liter, pH 6.4) and plated on 1% non-nutrient KK2 agar plates at a density of 5x105 cells/cm2 in a dark, moist chamber [121]. To study streaming, cells were seeded in submerged condition (KK2 buffer) at a density of 5x105 cells/cm2.
BIBR 1532 is a specific non-competitive inhibitor of TERT with IC50 value of 93 nM for human telomerase [122]. To find the optimal dose response of BIBR 1532 in Dictyostelium, starved cells were plated in phosphate buffered agar with different concentrations of BIBR 1532 (10 nM, 25 nM, 50 nM, 100nM and 200 nM) and 100nM was found to be the minimal effective dose in inducing complete stream breaking. MST 312, which is structurally unrelated to BIBR 1532, is a reversible inhibitor of TERT with IC50 value of 0.67 μM for human telomerase [123]. The minimal effective dose in Dictyostelium was found to be 250 nM. Inhibitor treatments were carried out with freshly starved cells resuspended in KK2 buffer and plated on KK2 agar plates.
The TRAP assay takes advantage of the low substrate specificity of telomerase, and involves replacing the telomere sequence with a synthetic template. The telomerase first extends the synthetic substrate primer by adding telomere repeats and these primary products are further amplified by PCR. The primer must have certain modifications, such as an anchor sequence at the 5’ end and two mismatches within the telomerase repeats [124, 125]. For the TRAP assay in Dictyostelium, we have used different primer sets (S2 Table) according to the basic design principles [124].
The KO vector for tert disruption was designed following standard cloning procedures. A 5' fragment of 678 bp and a 3' fragment of 322 bp spanning the tert gene (DDB_G0293918) and intergenic regions were PCR amplified and cloned on either side of a bsR cassette in pLPBLP vector (S13 Fig). Restriction endonuclease digestion and DNA sequencing were carried out to confirm the integrity of the KO vector. The tert KO vector was transfected to D. discoideum cells by electroporation. Axenically grown AX2 cells were washed twice with ice-cold electroporation buffer and 1x107 cells were resuspended in 100 μl EP++ buffer containing 10 μg of linearized tert KO vector. The cell suspension mixed with linearized KO vector was transferred to pre-chilled cuvettes (2 mm gap, Bio-Rad) and electroporated (300 V, 2 ms, 5 square wave pulses with 5 s interval) using a BTX ECM830 electroporator (Harvard Apparatus). The cell suspension was then transferred to a Petri dish containing 10 ml of HL5 medium and incubated at 22°C. After 24 h, the cultures were replaced with fresh HL5 supplemented with 10 μg/ml blasticidin (MP Biomedicals). Blasticidin-resistant clones were screened after three days. Genomic DNA isolated from tert KO clones were subjected to PCR analysis to confirm tert disruption using different primer combinations (S3 Table).
Using genomic DNA as template, a 3.8kb tert sequence was PCR amplified using ExTaq polymerase (Takara) and ligated in pDXA-GFP2 vector by exploiting the HindIII and KpnI restriction sites. This vector was electroporated to tert KO and AX2 cells and G418 resistant (10 μg/ml) clones were selected and overexpression was confirmed by semi-quantitative PCR. Primer sequences used for generating the vectors are mentioned in S4 Table.
Conditioned medium was prepared as described previously with slight modifications [126]. Briefly, log phase cells of AX2 and tert KO were resuspended at a density of 1x107 cells/ml and kept under shaking conditions for 20 h. Cells were pelleted and the supernatant was further clarified by centrifugation. The clarified supernatant (CM) was used immediately. To check the effect of CM on aggregate size, cells were developed in the presence of CM on non-nutrient agar plates and development was monitored. KK2 buffer was used as control. To deplete extracellular CF with anti-countin antibodies, cells were starved in KK2 buffer. After 1 h, the cells were developed with anti-countin antisera (1:300 dilution) in KK2 buffer [65].
To examine countin protein expression levels during aggregation, a Western blot was performed with anti-countin antibody. Cells were resuspended in SDS Laemmli buffer, and boiled for 3 min. Subsequently, the samples were run in a 12% SDS-polyacrylamide gel and Western blots were developed using an ECL Western blotting kit (Bio-Rad). Rabbit anti-countin antibodies were used at 1: 3000 dilution.
Log phase cells were starved at a density of 1x107 cells/ml in KK2 buffer in shaking conditions at 22°C for 4 h. At the beginning of starvation, 4x107 cells were removed and resuspended in 2 ml Sorensen phosphate buffer, vortexed vigorously and 0.4 ml of cell suspension was pipetted immediately in vials containing 0.4 ml ice-cold Sorensen phosphate buffer or 0.4 ml of 20 mM EDTA solution. The cell suspension was then transferred to a shaker and incubated for 30 min and 0.2 ml of 10% glutaraldehyde was added to each sample at the end of incubation and stored for 10 min. Then, 7 ml Sorensen phosphate buffer was added to each vial. Cell adhesion was indirectly measured by counting the number of single cells left behind using a hemocytometer [127].
To measure cell-substratum adhesion, 5x105 cells were seeded in 60mm Petri dishes and incubated at 22°C for 12 h. The Petri dishes with the cell suspension was placed on an orbital shaker at different speeds (0, 25, 50, 75 rpm). After 1 h, adherent and non-adherent cells were harvested, counted using a hemocytometer and the fraction of adherent cells was plotted against the rotation speed [58].
To visualize cAMP wave propagation, 5x105 cells/cm2 were plated on 1% non-nutrient agar plates and developed in dark moist conditions at 22°C. On a real-time basis, the aggregates were filmed at an interval of 30 s/frame, using a Nikon CCD camera and documented with NIS-Elements D software (Nikon, Japan). For visualizing cAMP optical density waves, image pairs were subtracted [92] using Image J (NIH, Bethesda, MD).
The under agarose cAMP chemotaxis assay was performed as described previously [128]. Briefly, 100 μl of cell suspension starved at a density of 1x107 cells/ml in KK2 buffer was added to outer troughs and 10 μM cAMP was added in the middle trough of a 1% agarose plate. Cells migrating towards cAMP was recorded every 30 s for 15 min with an inverted Nikon Eclipse TE2000 microscope using NIS-Elements D software (Nikon, Japan). For calculating the average velocity, directionality and chemotactic index, each time 36 cells were analyzed. The cells were tracked using ImageJ. Velocity was calculated by dividing the total displacement of cells by time. Directionality was calculated as the ratio of absolute distance traveled to the total path length, where a maximum value of 1 represents a straight path without deviations. Chemotactic index was calculated as the ratio of the average velocity of a cell moving against a cAMP gradient to the average cell speed. It is a global measure of direction of cell motion.
Total RNA was isolated from AX2 and tert KO cells at the indicated time points (0–24 h) using TRIzol reagent (Life Technologies, USA) [129]. RNA samples were quantified with a spectrophotometer (Eppendorf) and were also analyzed on 1% TAE agarose gels. cDNA was synthesized from total RNA using cDNA synthesis kit (Verso, Thermo-scientific). 1 μg of total RNA was used as a template to synthesize cDNA using random primers provided by the manufacturer. 1 μl of cDNA was used for qRT-PCR, using SYBR Green Master Mix (Thermo-scientific). qRT-PCR was carried out to analyze the expression levels of tert, acaA, carA, pdsA, regA, pde4, 5’NT, countin and smlA using the QuantStudio Flex 7 (Thermo-Fischer). rnlA was used as mRNA amplification control. All the qRT-PCR data were analyzed as described [130]. The primer sequences are mentioned in S5 Table.
cAMP levels were quantitated using cAMP-XP assay kit as per the manufacturer’s protocol (Cell Signalling, USA). AX2 and tert KO cells developed on 1% KK2 agar, were lysed with 100 μl of 1X lysis buffer and incubated on ice for 10 min. 50 μl of the lysate and 50 μl HRP-linked cAMP solution were added to the assay plates, incubated at room temperature (RT) on a horizontal orbital shaker. The wells were emptied after 3 h, washed thrice with 200 μl of 1X wash buffer. 100 μl of tetramethylbenzidine (TMB) substrate was added and incubated at RT for 10 min. The reaction was terminated by adding 100 μl of stop solution and the absorbance was measured at an optical density of 450 nm. The cAMP standard curve was used to calculate absolute cAMP levels.
Glucose levels were quantified as per the manufacturer’s protocol (GAHK20; Sigma-Aldrich). Mid-log phase cells were harvested and resuspended at a density of 8x106 cells/ml in KK2 buffer and kept in shaking conditions at 22°C. Cells were collected again and lysed by freeze-thaw method. 35 μl of the supernatant was mixed with 200 μl of glucose assay reagent and incubated for 15 min. The absorbance was measured at an optical density of 540 nm. The glucose standard curve was used to calculate absolute glucose levels.
Adenosine quantification was performed as per the manufacturer’s protocol (MET5090; Cellbio Labs). Cells grown in HL5 media were washed and seeded at a density of 5x105 cells/cm2 on KK2 agar plates. The aggregates were harvested using the lysis buffer (62.5 mM Tris-HCl, pH 6.8, 2% SDS, 10% glycerol). 50 μl sample was mixed with control mix (without adenosine deaminase) or reaction mix (with adenosine deaminase) in separate wells and incubated for 15 min. The fluorescence was measured using a spectrofluorometer (Ex- 550 nm, Em- 595 nm). The adenosine fluorescence in the sample was calculated by subtracting fluorescence of control mixed sample from reaction mixed sample. The adenosine standard curve was used to calculate absolute adenosine levels.
The conditioned media was incubated with 25 μg/ml DAPI for 5 min and polyphosphate specific fluorescence was measured using a spectrofluorometer (Ex- 415 nm, Em- 550 nm) as previously described [131]. Conditioned medium samples were prepared in FM minimal media to reduce the amount of background fluorescence. Polyphosphate concentration, in terms of phosphate monomers were determined using polyphosphate standards.
ICP-OES was performed as described previously [99]. Cells were developed on KK2 agar, washed five times in Sorensen phosphate buffer and pelleted. Then, 1 ml of concentrated HNO3 (70%) was added to each sample, and these were further digested by microwave heating. After digestion, the volume of each sample was brought to 9 ml with ultrapure water, filtered with 0.45 mm filter and analysed by ICP-OES (Perkin Elmer Optima 5300 DV ICP-OES). Sample digestion and metal quantification were carried out at the SAIF facility (Sophisticated Analytical Instrument Facility, IIT Madras).
A Nikon SMZ-1000 stereo zoom microscope with epifluorescence optics, Nikon 80i Eclipse upright microscope or a Nikon Eclipse TE2000 inverted microscope equipped with a digital sight DS-5MC camera (Nikon) were used for microscopy. Images were processed with NIS-Elements D (Nikon) or Image J.
Microsoft Excel (2016) was used for data analyses. Unpaired Student's t-test and two-way ANOVA (GraphPad Prism, version 6) were used to determine the statistical significance.
|
10.1371/journal.pntd.0002418 | Human T-Cell Lymphotropic Virus Type 1 Subtype C Molecular Variants among Indigenous Australians: New Insights into the Molecular Epidemiology of HTLV-1 in Australo-Melanesia | HTLV-1 infection is endemic among people of Melanesian descent in Papua New Guinea, the Solomon Islands and Vanuatu. Molecular studies reveal that these Melanesian strains belong to the highly divergent HTLV-1c subtype. In Australia, HTLV-1 is also endemic among the Indigenous people of central Australia; however, the molecular epidemiology of HTLV-1 infection in this population remains poorly documented.
Studying a series of 23 HTLV-1 strains from Indigenous residents of central Australia, we analyzed coding (gag, pol, env, tax) and non-coding (LTR) genomic proviral regions. Four complete HTLV-1 proviral sequences were also characterized. Phylogenetic analyses implemented with both Neighbor-Joining and Maximum Likelihood methods revealed that all proviral strains belong to the HTLV-1c subtype with a high genetic diversity, which varied with the geographic origin of the infected individuals. Two distinct Australians clades were found, the first including strains derived from most patients whose origins are in the North, and the second comprising a majority of those from the South of central Australia. Time divergence estimation suggests that the speciation of these two Australian clades probably occurred 9,120 years ago (38,000–4,500).
The HTLV-1c subtype is endemic to central Australia where the Indigenous population is infected with diverse subtype c variants. At least two Australian clades exist, which cluster according to the geographic origin of the human hosts. These molecular variants are probably of very ancient origin. Further studies could provide new insights into the evolution and modes of dissemination of these retrovirus variants and the associated ancient migration events through which early human settlement of Australia and Melanesia was achieved.
| The Human T-lymphotropic virus type 1 (HTLV-1) infects at least 5–10 million persons worldwide. In Oceania, previous studies have shown that HTLV-1 is present in a few ancient populations from remote areas of Papua New Guinea, the Solomon Islands, the Vanuatu archipelago and central Australia. The latter comprise one of the most socio-economically disadvantaged groups within any developed country. Characterization of the few available HTLV-1 viruses from Oceania indicates that these belong to a specific HTLV-1 genotype, the Australo-Melanesian c-subtype. In this study, we provide details for 23 HTLV-1 viruses derived from the Indigenous population of central Australia, a vast remote area of 1,000,000 km2. We reveal considerable genetic diversity of HTLV-1c subtype viruses and the existence of two HTLV-1c clades within which a high degree of genetic diversity was also apparent. These newly described HTLV-1c clades clustered according to the geographic origin of their human hosts. Indigenous Australians from the North of central Australia harbor HTLV-1c subtype viruses that are distinct from those of individuals from regions to the South. These data suggest that HTLV-1 was probably introduced to Australia during ancient migration events and was then confined to isolated Indigenous communities in central Australia.
| The Human T-lymphotropic virus type 1 (HTLV-1) is the first described human oncoretrovirus [1]. HTLV-1 infection is associated with a universally fatal malignancy, adult T-cell leukemia/lymphoma (ATLL), and with inflammatory disorders, the prototype of which is HTLV-1 associated myelopathy/tropical spastic paraparesis (HAM/TSP) [2]. HTLV-1 infects at least 5 to 10 million people worldwide [3]. It is widely distributed, with substantial clusters of high endemicity in certain geographic areas and ethnic groups in Southwestern Japan, sub-Saharan Africa, South America, the Caribbean basin and smaller endemic foci in Iran and Australo-Melanesia [3]. Seven main molecular HTLV-1 subtypes are currently recognized, predominantly from nucleotide sequence analysis of the LTR region. These are the Cosmopolitan subtype (a) that has spread worldwide, five African subtypes (b, d-g) and an Australo/Melanesian subtype (c), which is found only in Oceania [4], [5], [6], [7], [8]. The limited horizontal transmission of HTLV-1 and its clustering in certain ethnic/geographic foci have encouraged the use of its very slow in vivo genetic drift as a mean of studying the origin and modes of dissemination of this retrovirus as well as the movements of ancient infected populations [4], [9], [10].
Characterization of HTLV-1c variants was initially performed by Yanagihara et al. among a small group of hunter-horticulturalists, the Hagahai, who live in the fringe highlands of Papua New Guinea (PNG) [11], [12], [13] and among people of Melanesian descent in the Solomon Islands [13], [14]. Subsequently, efforts have been made to characterize new HTLV-1c subtype isolates in the neighboring territories of Australia [15], [16] and the Vanuatu archipelago [9], [17]. Nevertheless, our understanding of the molecular virology of the HTLV-1c subtype remains largely based on partial genome sequences of the gp21 env gene and LTR regions [9], [10], [13]. Indeed, only a single complete HTLV-1c subtype nucleotide sequence has been published to date, the original MEL5 human isolate from the Solomon Islands [18]. Previous studies indicate that HTLV-1 is also endemic to central Australia where high HTLV-1 seropositivity rates have been documented among Indigenous adults admitted to the sole regional hospital [19], [20]. Indeed, cases of ATL, HAM/TSP and infective dermatitis have now been described and an association between HTLV-1 infection and bronchiectasis has also been reported among Indigenous Australians [15], [21], [22], [23], [24], [25]. Interestingly, many clinical cases arise from the same central Australian region suggesting that environmental and/or viral factors may contribute to the etiology of HTLV-1 related diseases in this population. Unfortunately, only partial HTLV-1 nucleotide sequences are available for a single HTLV-1 strain from an Indigenous Australian [15], precluding any understanding of genetic variability in this population. Establishing a large HTLV-1 sequence database is thus essential for any study of the epidemiology and pathogenicity of HTLV-1c subtype.
The aim of the present study was therefore to describe the HTLV-1 genotypes infecting Indigenous central Australian residents and to correlate the results of the HTLV-1 nucleotide sequence variability with the geographic origin of the individuals living within this vast region of approximately 1,000,000 km2.
Our work was performed using HTLV-1 isolates obtained from a large series of patients who were initially enrolled to HTLV-1 pathogenesis studies between October 2007 and August 2010 [22]. Plasma and peripheral blood buffy coat (PBBC) samples were obtained from 23 HTLV-1 infected patients who presented to Alice Springs Hospital, predominantly with bronchiectasis. Presumed place of origin was determined from language group and/or place of residence in infancy (figure 1). Numerous Indigenous languages are spoken in this region; however, for the purpose of the present study, these were divided into two groups according to the predominant geographic areas in which they are spoken; i) Northern (comprising the Ngarrkic language groups) and ii) Southern (comprising both Arandic and Western Desert language groups) (table 1). Also included in the present study were PBBC samples from four Natives of the Vanuatu archipelago. These were collected by us during work in Vanuatu between 2003 and 2005, as has been described previously [9], [17].
Written informed consent was given by all patients for their blood to be used for the pathogenicity studies, which included the molecular characterization of the HTLV-1 viral strains. The Central Australian Human Research Ethics Committee (CAHREC) approved this study (CAHREC Ref: 2011.11.01).
The plasma and PBBC samples were transferred to Institut Pasteur, Paris, and stored at −80°C until HTLV-1 analysis. Plasma HTLV-1 antibodies were tested by a particle agglutination (PA) technique (Serodia HTLV-1, Fujirebio, Tokyo, Japan) and by an indirect immunofluorescence assay (IFA) using the HTLV-1-transformed human T cell lines MT2. All samples were also tested by Western blot assay (WB) (HTLV Blot 2.4, MP Biomedicals Asia Pacific Pte. Ltd., Singapore).
High-molecular weight DNA was extracted from PBBC using the QIAamp DNA Blood Mini Kit (Qiagen, Hilden, Germany). Samples were first subjected to PCR using human β-globin specific primers, to ensure that DNA was amplifiable [26]. All 27 samples were then submitted to two series of PCR using LTR-gag and Px-LTR primers which were designed using highly conserved regions that are common to the major HTLV-1 subtypes (figures 2A and 2B). The LTR-gag primers are the following: Enh280: 5′-TGACGACAACCCCTCACCTCAA-3′ and R2380: 5′-GTCCGGAAAGGGAGGCGTATTAG-3′ corresponding to nucleotides 258 to 279 and 2,377 to 2,399 respectively of the prototype ATK-1 sequence (Genbank: J02029). The Px-LTR primers are F6501: 5′-CTTAACTGGGACCTTGGCCTCTCAC-3′ (nt 6,476 to 6,500) and 3VLTRext: 5′-CGCAGTTCAGGAGGCACCRMA-3′ (nt 8,741 to 8,761) (figure 2A).
To obtain the 522-bp fragment of the gp21 env gene, 1 µg of DNA from the Aus-GN strain was subjected to 2 series of PCR as previously described [27].
Two additional series of PCR using pro-pol and pol-env primers were necessary to obtain the entire genome amplification of the four Australian proviral sequences (Aus-CS, Aus-NR, Aus-DF and Aus-GM). The pro-pol primers are F2279: 5′-GGAGCAGACATGACAGTCCTTCC-3′ (nt 2,254 to 2,276) and R5005: 5′-GGCGGCTATTAAGACCAGGAAGC-3′ (nt 5,002 to 5,024). The pol-env primers are the following F4583: 5′-CAGGAGCCATCTCAGCTACCC-3′ (nt 4,560 to 4,580) and env2c: 5′-TTTATAAGAGAGTAATGGGGGTATCTG-3′ (nt 6,613 to 6,639) (figure 2A). The size of the different generated amplicons are the following: LTR-gag, 2,136-bp; pro-pol, 2,769-bp; pol-env, 2,078-bp and Px-LTR, 2,273-bp (figure 2A).
A PIKO thermocycler (Ozyme, Saint Quentin-en-Yvelines, France) was used with the following amplification conditions. LTR-gag: 98°C, 1 mn; 40×(98°C, 10 s; 69°C, 10 s; 72°C, 1 mn); 72°C, 2 mn and Px-LTR: 98°C, 1 mn; 40×(98°C, 10 s; 72°C, 1 mn); 72°C, 2 mn. For pro-pol: 98°C, 1 mn; 40×(98°C, 10 s; 68°C, 10 s; 72°C, 1 mn); 72°C, 2 mn and pol-env: 98°C, 1 mn; 40×(98°C, 10 s; 62.5°C, 10 s; 72°C, 1 mn); 72°C, 2 mn. Reaction tubes were prepared in a dedicated room outside the laboratory, with a final volume of 50 µl (DNA matrix, 500 ng; dNTP mix (Roche, Basel, Switzerland), 40 µM; 5×Phire reaction buffer which contains 1.5 mM MgCl2 at final reaction concentration (Ozyme, Saint Quentin-en-Yvelines, France), 5 µl; Phire hot start DNA polymerase (Ozyme, Saint Quentin-en-Yvelines, France), 2 U and 0.5 µM of each oligonucleotide primer (Eurofins MWG, Ebersberg, Germany). Five µl of amplified DNA was size fractionated by 1.5% agarose gel electrophoresis, and the PCR products (45 µl) were sent for purification and sequencing reactions to the MilleGen Company (MilleGen, Labège, France) (table 2).
Both strands of each PCR product were sequenced, and the ClustalW algorithm (MacVector 6.5 software, Oxford Molecular) was implemented to align forward and reverse sequences of each segment to derive a consensus sequence of the full LTR (758-bp) region, a fragment of the gp21 env gene (522-bp), colinearized gag-tax (2,346-bp) and gag-pol-env-tax (7567-bp) genes (figure 2). Phylogenetic trees were generated from multiple alignments of the LTR region and gp21 env together with the colinearized gag-tax and gag-pol-env-tax genes. Included in the phylogenetic analyses were the 23 new proviral sequences from Australia and the four novel sequences from Vanuatu (ESH18, ESW44, EM5, PE376) that were characterized in the present study together with appropriate sequences of previously characterized strains from PNG (MEL1, MEL2 and MEL7), the Solomon Islands (MEL3 to MEL6 and MEL 8 to MEL10), Vanuatu (PE376, VAN54, VAN136, VAN251, VAN335) and Australia (MSHR-1). Additional representative sequences of the HTLV-1 a, b, d-g subtypes available in Genbank were also included.
The sequences were aligned using the DAMBE program (version 4.2.13) [28]. Absence of saturation of the alignment was confirmed by 2 methods: likelihood mapping (model TN93; non uniform substitution) with Tree-Puzzle software (version 5.2) and the test of Xia and Xie [28] with the DAMBE program. The final alignment was submitted to the Modeltest program (version 3.6) and the best model was selected according to the Akaike information criterion. This was then applied to phylogenetic analyses using the PAUP program (version 4.0b10) to infer trees according to both Neighbor-Joining (NJ) and Maximum Likelihood (ML) methods. To test the robustness of the tree topologies, 1,000 bootstrap replicates were performed. Numbers applied to the nodes of the tree (bootstrap values) indicate frequencies of occurrence for 100 trees. The quartet puzzling algorithm included in the Tree-Puzzle software was applied for the maximum likelihood method [29].
In order to estimate the divergence time between the different clades, we initially performed a typical molecular clock analysis [9]. This method was not conclusive. Indeed, the sequences seem to be too short (when considering the low mutation rate for HTLV-1) to be informative under the different models. We therefore estimated the divergence time using the previously reported mutation rates for HTLV-1 [30]. A theoretical ancestral sequence was initially determined for each monophyletic clade, and the number of mutation events required to generate the reported current sequences was then calculated. Finally, this average number of mutation events was multiplied by the known HTLV-1 mutation rate [30]. Although the technique is rough, the estimated date for the Vanuatu/Solomon node is 7,440 years BP (31,000–3,700), which is consistent with the date previously proposed (i.e. 10,000 years ago) [9].
The 23 HTLV-1 infected individuals from Australia included ten women (mean age 49.7 years, range 27–70) and 13 men (mean age 42.2 years, range 16–67) (table 1). The four samples from Vanuatu were obtained from 3 women (mean age 59, range 40–76) and a 61-year old man. All plasma samples exhibited a complete HTLV-1 pattern in WB and the presence of HTLV-1 provirus was investigated in the DNA of these 27 individuals.
The primary purpose of our work was to study the molecular relationship between the new Australian HTLV-1 proviral strains and those from HTLV-1 infected individuals in Australia and the neighboring islands whose sequences have been previously published. For most strains, the only available sequences in the Genbank database are the 522-bp fragment of the gp21 env gene. We therefore compared the gp21 env gene fragments of seven proviral sequences from Australia, including five new proviral strains (Aus-Cs, Aus-DF, Aus-NR, Aus-GN and Aus-GM) and two previously characterized sequences (MSHR-1 and Aus-RDJ) (Genbank: M92818 and JX891480, respectively) [15], [25], with HTLV-1 proviral strains from PNG (MEL1, MEL2 and MEL7), Vanuatu (EM5, VAN54, VAN136, VAN251 and PE376) and the Solomon Islands (MEL3 to MEL6 and MEL8 to MEL10) [9], [13], [18], [31]. Phylogenetic analyses performed with both NJ and ML methods clearly demonstrate the existence of three subgroups: “Papua New Guinean”, “Solomon/Vanuatu” and “Australian”, within the HTLV-1c subtype. Furthermore, inside the Australian subgroup, which comprises all the 7 Australian proviral strains, two distinct clades are now observed. The first clade includes strains derived from 3 patients (Aus-CS, Aus-DF and Aus-NR) whose origins are in the North of central Australia plus the two published sequences (MSHR-1 and Aus-RDJ), and the second comprises 2 patients (Aus-GN and Aus-GM) from the South of central Australia (figures 3A and 3B).
Complete LTR proviral sequences were obtained for all 27 samples by PCR amplification of both LTR-gag and Px-LTR fragments. Alignment of the 746-bp LTR fragments for these 27 strains revealed no significant deletion or insertion in comparison to the HTLV-1 ATK-1 reference strain. Within group comparisons of the 23 new Australian HTLV-1 strains indicate that they are closely related to each other (range of nucleotide similarity, 99.5%–100%), though quite divergent from the LTR strains from Vanuatu (nucleotide similarity range, 94.5% – 95.3%) and from the known HTLV-1c subtype prototype strain from the Solomon Islands (MEL5) (nucleotide similarity range, 95.3%–95.5%).
Using both NJ and ML methods, the phylogenetic analyses of the LTR region revealed two distinct subgroups within the Australo-Melanesian HTLV-1c subtype. The first group includes strains from the Solomon Islands and Vanuatu (Sol/Van) while the second comprises all the HTLV-1 Australian proviral sequences (Aus) (figures 4A and 4B). Unfortunately, no complete LTR proviral sequence from PNG is available in the sequences databases. However, phylogenetic analysis based on a partial 627-bp fragment of the LTR region, including the PNG-1 strain from Papua New Guinea (Genbank: M85207), clearly confirmed the existence of three distinct clades within the HTLV-1c subtype (data not shown) [32].
Phylogenetic analyses using both gp21 env and LTR fragments were consistent with the existence of two Australian HTLV-1c clades. The nature of this relationship was further clarified by comparing larger genomic fragments using both ML and NJ methods.
A comparison of the concatenated and aligned 2,346-bp fragment of the gag-tax genes revealed a high degree of nucleotide homology among the Australian strains (range, 98.9%–100%) and a comparable degree of divergence of the Australian strains relative to those from both Vanuatu (range, 94.1%–97.5%) and the Solomon Islands (range, 94.4%–96.9%). Additional phylogenetic analyses using both NJ and ML methods of the colinearized gag-tax (2,346-bp) genomic fragment, confirmed the tree topology derived from the gp21 env and LTR analyses and demonstrated the existence of an Australian subgroup that was highly supported phylogenetically (bootstrap value ≥ 99%) (figures 5A and 5B). Interestingly, the Australian subgroup can be further subdivided into two clades, for which bootstrap values are also statistically significant (≥ 91%). The first includes HTLV-1 strains derived from most patients of Northern origin (6/7) and the second comprises a majority of individuals (14/16) from the South. Furthermore, a high genetic diversity exists within both Australian clades with sub-clades also supported by high bootstrap values (≥ 90%).
These analyses were performed using the complete proviral sequences obtained from four Australian samples: Aus-CS, Aus-NR, Aus-DF and Aus-GM (Genbank: KF242506, JX891479, KF242505 and JX891478 respectively). In addition, two of these complete proviral sequences were selected as the representative prototypes of each Australian clade (“Northern” clade, Aus-NR; “Southern” clade, Aus-GM. The general genomic organization of these two prototypic sequences is similar to that of HTLV-1 prototypes ATK-1 and MEL5 strains (Genbank: J02029 and L02534, respectively). The overall range of nucleotide divergence of the first complete Australian strains from the prototypes ATK-1 and MEL5 was 7.8–8% and 3.3–3.4% respectively. The nucleotide homology between these two Australian prototypic sequences was 98.9% (102 differences over 9,046-bp).
Phylogenetic analyses using both NJ and ML methods of the colinearized gag-pol-env-tax (6,000-bp) genomic fragments, including HTLV-1, HTLV-2 and HTLV-3 representative sequences available in Genbank, confirmed the existence of an Australian subgroup that was highly supported phylogenetically (figures 6A and 6B).
Finally, we estimated the time of divergence for the various Australian strains using the evolution rate of the HTLV-1 LTR region, which has previously been determined by Lemey et al. (5.6×10−7 substitutions/site/year; 90% confidence interval, 1.2×10−7 to 1.1×10−6) [30]. These calculations suggest that divergence between the Vanuatu/Solomon and Australian subgroups occurred 20,400 years ago (85,000–10,200) and that speciation of the two Australian clades followed 9,120 years ago (38,000–4,500). The estimated date for the Vanuatu/Solomon node in the present study was 7,440 years ago (31,000– 3,700), which is consistent with our previous estimate (10,000 years ago) [9].
The origin of most HTLV-1 subtypes appears to be linked to ancient and multiple episodes of interspecies transmission between STLV-1-infected non-human primates (NHPs) and humans [5], [33], . Indeed, Old-World NHPs constitute a large reservoir for different lineages of STLV-1, and the virus is considered as transmissible to humans through body fluid contacts [27], [36], [37], [38], [39]. The very high homology between some STLV-1 and HTLV-1 strains, particularly the b and d-f subtypes, suggests that interspecies transmission to humans is probably ongoing in some areas of West and central Africa and results from close contacts during the hunting or butchering of NHPs [8], [37], [38], [40], [41], [42].
Despite the presence of STLV-1-infected NHP species in Asia, there is no evidence of recent interspecies transmission in this area [43]. Furthermore, monkeys have never been endemic to the Australo-Melanesian region, indicating that interspecies transmission of STLV-1 to humans could not have occurred in these islands [44]. Therefore, HTLV-1c is likely to have been acquired by the ancestors of the Indigenous peoples of Australo-Melanesia as a result of interspecies transmission from NHPs during their migration through South-East Asia and prior to reaching the highlands of Papua New Guinea [5], [10], [43], [45]. The subsequent migratory movements of this ancestral population then resulted in the radiation of HTLV-1c throughout the Australo-Melanesian region.
A first wave of migration led to the progressive colonization of the Solomon Islands, followed by the Vanuatu archipelago and finally, New Caledonia and the neighboring Melanesian islands. Consistent with the common origin of these Melanesian populations are our analyses performed on gp21 env, the LTR and the colinearized gag-tax and gag-pol-env-tax genes, which confirm that HTLV-1 strains from the Solomon Islands and Vanuatu belong to the same subgroup. Based on a combination of paleo-anthropological data and genomic DNA analyses, it is believed that the initial human settlement of the Solomon archipelago dates from the Paleolithic period, ca. 30,000 years ago [46], while Vanuatu was settled much latter, during the Neolithic period, ca. 10,000 years ago [47]. In previous phylogenetic and molecular-clock analyses, we suggested that the HTLV-1c proviral strains from the Indigenous people of Vanuatu and the Solomon Islands emerged from a common ancestor ∼10,000 years ago [9], which is consistent with data presented here (ca. 7,440 yrs ago).
A second wave of migration is likely to have occurred from PNG to Australia. Indeed, it is thought that the occupation of Sahul, the continent formed when glacio-eustatically lowered sea levels exposed dry land connections between Australia and Papua New Guinea, may have commenced 45,000 years ago and continued until the end of the Pleistocene period 12,000 years ago [48], [49], [50]. Recently, Rasmussen and colleagues presented evidence derived from the gene flow between populations, which indicates that present-day Indigenous Australians are descendants of the earliest humans to occupy Australia and that they represent one of the oldest continuous populations outside Africa [51].
In the present study, we reveal a high degree of genetic diversity among the HTLV-1c subtype proviral strains that infect the Indigenous people of central Australia. At least two different HTLV-1c genetic clades exist in this Indigenous population and these cluster according to the geographic origin of their human hosts. Thus, it is possible to propose that the common ancestor of the modern Australian HTLV-1 strains arrived in Australia when a group originating from the ancestral HTLV-1 infected population migrated from PNG and settled in Australia. Subsequently, the Australian population split (ca. 9,000 yrs ago) leading to continued viral evolution among small, isolated clan groups of Indigenous people dwelling in the remote desert regions of central Australia and this resulted in the speciation of the two Australian clades. The broad ethno-geographic distinction between the Indigenous human hosts of these Australian clades is particularly interesting given that considerable movement of Indigenous people has resulted from a century of European dominance in this region. Thus, Northern Ngarrkic speaking clan groups were moved to the South while Western Desert groups moved to the East, in each case toward ration supply centers that were established nearer the major regional center of Alice Springs [52]. Prior to colonization contact between these groups is likely to have been minimal. The data presented here therefore describes probably the molecular epidemiological expression of both long-term evolution and more recent human movements that were driven by colonization.
Further studies, which characterize the HTLV-1c proviral strains that infect other Indigenous populations elsewhere in Australia and Oceania, will provide new insights into the origin of these retroviruses, potentially enhancing our understanding of the pathogenicity, evolution and modes of dissemination of these HTLV-1c variants and their human hosts.
|
10.1371/journal.ppat.1007185 | Changes in temperature alter the potential outcomes of virus host shifts | Host shifts–where a pathogen jumps between different host species–are an important source of emerging infectious disease. With on-going climate change there is an increasing need to understand the effect changes in temperature may have on emerging infectious disease. We investigated whether species’ susceptibilities change with temperature and ask if susceptibility is greatest at different temperatures in different species. We infected 45 species of Drosophilidae with an RNA virus and measured how viral load changes with temperature. We found the host phylogeny explained a large proportion of the variation in viral load at each temperature, with strong phylogenetic correlations between viral loads across temperature. The variance in viral load increased with temperature, while the mean viral load did not. This suggests that as temperature increases the most susceptible species become more susceptible, and the least susceptible less so. We found no significant relationship between a species’ susceptibility across temperatures, and proxies for thermal optima (critical thermal maximum and minimum or basal metabolic rate). These results suggest that whilst the rank order of species susceptibilities may remain the same with changes in temperature, some species may become more susceptible to a novel pathogen, and others less so.
| Emerging infectious diseases are often the result of a host shift, where a pathogen jumps from one host species into another. Understanding the factors underlying host shifts is a major goal for infectious disease research. This effort has been further complicated by the fact that host-parasite interactions are now taking place in a period of unprecedented global climatic warming. Here, we ask how host shifts are affected by temperature by carrying out experimental infections using an RNA virus across a wide range of related species, at three different temperatures. We find that as temperature increases the most susceptible species become more susceptible, and the least susceptible less so. This has important consequences for our understanding of host shift events in a changing climate as it suggests that temperature changes may affect the likelihood of a host shift into certain species.
| Temperature is arguably the most important abiotic factor that affects all organisms, having both indirect and direct effects on physiology and life history traits [1–3]. There is much to be learned about the impact of climate change on infectious diseases [1,4,5]. Changes in temperature can impact both host and parasite biology, leading to complex and difficult to predict outcomes [2,6].
Host shifts, where a parasite from one host species invades and establishes in a novel host species, are an important source of emerging infectious disease [7]. A successful host shift relies on a number of stages occurring [8]. Firstly, exposure of the host to the new pathogen species must occur in such a way that transmission is successful. Secondly, the pathogen must be able to replicate sufficiently to infect the novel host. Finally, there must be sufficient onwards transmission for the pathogen to become established in the new host species [7,9,10]. Some of the most deadly outbreaks of infectious diseases in humans including Ebola virus, HIV and SARS coronavirus have been linked to a host switch event [11–14] and many others have direct animal vectors or reservoirs (e.g. Dengue and Chikungunya viruses) [15,16]. The potential for novel host shifts may increase with changing temperatures due to, fluctuations in host and/or parasite fitness, or changes in species distributions and abundances [17,18]. Distribution changes may lead to new species assemblages, causing novel contacts between parasites and potential hosts [19–21].
Susceptibility to infection is known to vary with temperature, due to within individual physiological changes in factors such as the host immune response, metabolic rate or behavioural adaptations [22–25]. Thermally stressed hosts may face a trade-off between the resource investment needed to launch an immune response versus that needed for thermoregulation, or behavioural adaptations to withstand sub-optimal temperatures [26–29]. Temperature shifts could also cause asymmetrical or divergent effects on host and parasite traits [30]. For example, changes in temperature may allow differential production and survival of parasite transmission stages, and changes in replication rates, generation times, infectivity and virulence [31–33]. Temperature is also known to impact vector-borne disease transmission through multiple effects on both vector life cycles and transmission behaviours [20,34–37].
Host shifts have been shown to be more likely to occur between closely related species [38–40], but independently of this distance effect, clades of closely related hosts show similar levels of susceptibility [9,41]. Thermal tolerances − like virus susceptibility − are known to vary across species, with groups of closely related species having similar thermal limits, with a large proportion of the variation in these traits being explained by the phylogeny [42–45]. Previous studies on host shifts have assayed the susceptibility of species at a single temperature [9,39,41,46]. However, if the host phylogeny also explains much of the variation in thermal tolerance, then phylogenetic patterns in virus susceptibility could be due to differences between species’ natural thermal optima and the chosen assay temperatures. Therefore, for experiments carried out at a single temperature, phylogenetic signal in thermal tolerance may translate into phylogenetic signal in thermal stress. Any apparent phylogenetic signal in susceptibility could potentially be due to the effects of thermal stress, and may not hold true if each species was to be assayed at its optimal temperature. If this was indeed the case this would have implications for species distribution models that aim to use estimates of environmental conditions to predict host and pathogen ranges [5,47,48].
Here, we have asked how species’ susceptibilities change at different temperatures and whether susceptibility is greatest at different temperatures in different species. We infected 45 species of Drosophilidae with Drosophila C Virus (DCV; Dicistroviridae) at three different temperatures and measured how viral load changes with temperature. Viral load is used here as a measure of DCV’s ability to persist and replicate in a host, which has previously been shown to be tightly correlated to host mortality [41]. We are therefore examining one of the steps (“ability to infect a novel host”) needed for a host shift to successfully occur [7,9,10]. We also examine how proxies for thermal optima and cellular function (thermal tolerances and basal metabolic rate) relate to virus susceptibility across temperatures, as increasing temperatures may have broad effects on both host and parasite [43–45]. DCV is a positive sense RNA virus in the family Discistroviridae that was originally isolated from Drosophila melanogaster and in the wild has been found in D. melanogaster and D. simulans [49–51]. DCV infected flies show reduced metabolic rate and activity levels, develop an intestinal obstruction, reduced hemolymph pH and decreased survival [52–55]. This work examines how temperature can influence the probability of host shifts, and looks at some of the potential underlying causes.
We used Drosophila C virus (DCV) clone B6A, which is derived from an isolate collected from D. melanogaster in Charolles, France [56]. The virus was prepared as described previously [57]; briefly DCV was grown in Schneider’s Drosophila line 2 cells and the Tissue Culture Infective Dose 50 (TCID50) per ml was calculated using the Reed-Muench end-point method [58].
Flies were obtained from laboratory stocks of 45 different species. All stocks were maintained in multi generation populations, in Drosophila stock bottles (Dutscher Scientific) on 50ml of their respective food medium at 22°C and 70% relative humidity with a 12 hour light-dark cycle (Table A in S1 Text). Each day, two vials of 0–1 day old male flies were randomly assigned to one of three potential temperature regimes; low, medium or high (17°C, 22°C and 27 °C respectively) at 70% relative humidity. Flies were tipped onto fresh vials of food after 3 days, and after 5 days of acclimatisation at the experimental temperature were infected with DCV. Flies were anesthetized on CO2 and inoculated using a 0.0125 mm diameter stainless steel needle that was bent to a right angle ~0.25mm from the end (Fine Science Tools, CA, USA)[9,41,57]. The bent tip of the needle was dipped into the DCV solution (TCID50 = 6.32×109) and pricked into the pleural suture on the thorax of the flies. We selected this route of infection as oral inoculation has been shown to lead to stochastic infection outcomes in D. melanogaster [55]. However, once the virus passes through the gut barrier, both oral and pin-pricked infections follow a similar course, with both resulting in the same tissues becoming infected with DCV [55]. One vial of inoculated flies was immediately snap frozen in liquid nitrogen to provide a time point zero sample as a reference to control for relative viral dose. The second vial of flies were placed onto a new vial of fresh cornmeal food and returned to their experimental temperature. After 2 days (+/- 1 hour) flies were snap frozen in liquid nitrogen. This time point was chosen based on pilot data as infected flies showed little mortality at 2 days post infection, and viral load plateaus from day 2 at 22°C. Temperatures were rotated across incubators in each block to control for incubator effects. All frozen flies were homogenised in a bead homogeniser for 30 seconds (Bead Ruptor 24; Omni international, Georgia, USA) in Trizol reagent (Invitrogen) and stored at -80°C for later RNA extractions.
These collections and inoculations were carried out over three replicate blocks, with each block being completed over consecutive days. The order that the fly species were infected was randomized each day. We aimed for each block to contain a day 0 and day 2 replicate for each species, at each temperature treatment (45 species × 3 temperatures × 3 experimental blocks). In total we quantified viral load in 12,827 flies over 396 biological replicates (a biological replicate = change in viral load from day 0 to day 2 post-infection), with a mean of 17.1 flies per replicate (range across species = 4–27). Of the 45 species, 42 had 3 biological replicates and three species had 2 biological replicates.
The change in RNA viral load was measured using quantitative Reverse Transcription PCR (qRT-PCR). Total RNA was extracted from the Trizol homogenised flies, reverse-transcribed with Promega GoScript reverse transcriptase (Promega) and random hexamer primers. Viral RNA load was expressed relative to the endogenous control housekeeping gene RpL32 (RP49). RpL32 primers were designed to match the homologous sequence in each species and crossed an intron-exon boundary so will only amplify mRNA [9]. The primers in D. melanogaster were RpL32 qRT-PCR F (5’-TGCTAAGCTGTCGCACAAATGG -3’) and RpL32 qRT-PCR R (5’- TGCGCTTGTTCGATCCGTAAC -3’). DCV primers were 599F (5’-GACACTGCCTTTGATTAG-3’) and 733R (5’CCCTCTGGGAACTAAATG-3’) as previously described [41]. Two qRT-PCR reactions (technical replicates) were carried out per sample with both the viral and endogenous control primers, with replicates distributed across plates in a randomised block design.
qRT-PCR was performed on an Applied Biosystems StepOnePlus system using Sensifast Hi-Rox Sybr kit (Bioline) with the following PCR cycle: 95°C for 2min followed by 40 cycles of: 95°C for 5 sec followed by 60°C for 30 sec. Each qRT-PCR plate contained four standard samples. A linear model was used to correct the cycle threshold (Ct) values for differences between qRT-PCR plates. Any samples where the two technical replicates had cycle threshold (Ct) values more than 2 cycles apart after the plate correction were repeated. To estimate the change in viral load, we first calculated ΔCt as the difference between the cycle thresholds of the DCV qRT-PCR and the RpL32 endogenous control. For each species the viral load of day 2 flies relative to day 0 flies was calculated as 2-ΔΔCt; where ΔΔCt = ΔCtday0 –ΔCtday2. The ΔCtday0 and ΔCtday2 are a pair of ΔCt values from a day 0 biological replicate and a day 2 biological replicate. Calculating the change in viral load without the use of the endogenous control gene (RpL32) gave equivalent results (Spearman’s correlation between viral load calculated with and without endogenous control: ρ = 0.97, P< 0.005)
We carried out two assays to measure the thermal tolerances of species; a cold resistance measure to determine critical thermal minimum (CTmin) under gradual cooling, and a heat resistance measure through gradual heating to determine critical thermal maximum (CTmax). 0–1 day old males were collected and placed onto fresh un-yeasted cornmeal food vials. Flies were kept for 5 days at 22°C and 70% relative humidity and tipped onto fresh food every 2 days. In both assays individual flies were placed in 4 ml glass vials (ST5012, Ampulla, UK) and exposed to temperature change through submersion in a liquid filled glass tank (see Fig A in S1 Text). For CTmax the tank was filled with water and for CTmin a mixture of water and ethylene glycol (50:50 by volume) was used to prevent freezing and maintain a constant cooling gradient. Five biological replicates were carried out for each species for both CTmax and CTmin. Temperature was controlled using a heated/cooled circulator (TXF200, Grant Instruments, Cambridgeshire, UK) submerged in the tank and set to change temperatures at a rate of 0.1 °C/min, always starting from 22°C (the rearing temperature for stock populations). Flies were monitored continually throughout the assay and the temperature of knock down was ascertained by a disturbance method, whereby a fly was scored as completely paralysed if on gentle tapping of the vial wall the fly did not move any of its body parts.
To examine how cellular function changes with temperature, we estimated the resting metabolic rate of each species at 17°C, 22°C and 27 °C to examine if changes in general cellular processes were related to changes in viral load. Following the same methods as the viral inoculation assay, groups of 10, 0–1 day old male flies from 44 species were acclimatised at the three experimental temperatures for 5 days (D. pseudoobscura was excluded as not enough individuals could be obtained from stocks for sufficient replication). Every 2 days flies were tipped onto fresh vials of cornmeal food. This was repeated in three blocks in order to get three repeat measures of metabolic rate for each of the species, at each of the three experimental temperatures. Flies were collected in a randomly assigned order across the three blocks.
Closed system respirometry was used to measure the rate of CO2 production (VCO2) as a proxy for metabolic rate [59]. Flies were held in 10ml-3 airtight plastic chambers constructed from Bev-A-Line V Tubing (Cole-Parmer Instrument Company, UK). All measures were carried out during the day inside a temperature controlled incubator, with constant light, that was set to each of the experimental temperatures that the flies had been acclimatised to. The set up followed that of Okada et al. (2011)[60]. Compressed air of a known concentration of oxygen and nitrogen (21% O2:79% N2) was scrubbed of any CO2 and water (with Ascarite II & Magnesium Perchlorate respectively) and pumped through a Sable Systems RM8 eight-channel multiplexer (Las Vegas, NV, USA) at 100 ml/min-1 (±1%) into the metabolic chambers housing the groups of 10 flies. The first chamber was left empty as a reference cell, to acquire a baseline reading for all subsequent chambers at the start and end of each set of runs, therefore seven groups of flies were assayed in each run. Air was flushed into each chamber for 2 minutes, before reading the previous chamber. Readings were taken every second for 10 minutes by feeding the exiting air through a LiCor LI-7000 infrared gas analyser (Lincoln, NE, USA). Carbon dioxide production was measured using a Sable Systems UI2 analog–digital interface for acquisition, connected to a computer running Sable Systems Expedata software (v1.8.2) [61]. The metabolic rate was calculated from the entire 10-minute recording period by taking the CO2 reading of the ex-current gas from the chamber containing the flies and subtracting the CO2 measure of the incurrent gas entering the chamber. These values were also corrected for drift away from the baseline reading of the empty chamber. Volume of CO2 was calculated as VCO2 = FR (Fe CO2 –Fi CO2) / (1-Fi CO2). Where FR is the flow rate into the system (100ml/min-1), Fe CO2 is the concentration of CO2 exiting and Fi CO2 is the concentration CO2 entering the respirometer. Species were randomly assigned across the respiration chambers and the order in which flies were assayed (chamber order) was corrected for statistically (see below).
To check for any potential effect of body size differences between species on viral load, wing length was measured as a proxy for body size [62]. A mean of 26 (range 20–30) males of each species were collected and immediately stored in ethanol during the collections for the viral load assay. Subsequently, wings were removed and photographed under a dissecting microscope. Using ImageJ software (version 1.48) the length of the IV longitudinal vein from the tip of the proximal segment to where the distal segment joins vein V was recorded, and the mean taken for each species.
The host phylogeny was inferred as described in Longdon et al (2015) [41], using the 28S, Adh, Amyrel, COI, COII, RpL32 and SOD genes. Briefly, any publicly available sequences were downloaded from Genbank, and any not available we attempted to Sanger sequence [9]. In total we had RpL32 sequences for all 45 species, 28s from 41 species, Adh from 43 species, Amyrel from 29 species, COI from 38 species, COII from 43 species and SOD from 25 species (see www.doi.org/10.6084/m9.figshare.6653192 full details). The sequences of each gene were aligned in Geneious (version 9.1.8, [63]) using the global alignment setting, with free end gaps and a cost matrix of 70% similarity. The phylogeny was constructed using the BEAST program (version 1.8.4,[64]). Genes were partitioned into three groups each with their own molecular clock models. The three partitions were: mitochondrial (COI, COII); ribosomal (28S); and nuclear (Adh, SOD, Amyrel, RpL32). A random starting tree was used, with a relaxed uncorrelated lognormal molecular clock. Each of the partitions used a HKY substitution model with a gamma distribution of rate variation with 4 categories and estimated base frequencies. Additionally, the mitochondrial and nuclear data sets were partitioned into codon positions 1+2 and 3, with unlinked substitution rates and base frequencies across codon positions. The tree-shape prior was set to a birth-death process. The BEAST analysis was run twice to ensure convergence for 1000 million MCMC generations sampled every 10000 steps. The MCMC process was examined using the program Tracer (version 1.6, [65]) to ensure convergence and adequate sampling, and the constructed tree was then visualised using FigTree (version 1.4.3, [66]).
All data were analysed using phylogenetic mixed models to look at the effects of host relatedness on viral load across temperature. We fitted all models using a Bayesian approach in the R package MCMCglmm [67,68]. We ran trivariate models with viral load at each of the three temperatures as the response variable similar to that outlined in Longdon et al. (2011) [9]. The models took the form:
yhit=β1:t+bmrh∙β2+wingsizeh∙β3+CTminh∙β4+CTmaxh∙β5+up:ht+ehit
Where y is the change in viral load of the ith biological replicate of host species h, for temperature t (high, medium or low). β are the fixed effects, with β1 being the intercepts for each temperature, β2 being the effect of basal metabolic rate, β3 the effect of wing size, and β4 and β5 the effects of the critical thermal maximum (CTmax) and minimum (CTmin) respectively. up are the random phylogenetic species effects and e the model residuals. We also ran models that included a non-phylogenetic random species effect (unp:ht) to allow us to estimate the proportion of variation explained by the host phylogeny [9,41,69]. We do not use this term in the main model as we struggled to separate the phylogenetic and non-phylogenetic terms. Our main model therefore assumes a Brownian motion model of evolution [70]. The random effects and the residuals are assumed to be multivariate normal with a zero mean and a covariance structure Vp ⊗ A for the phylogenetic affects and Ve ⊗ I for the residuals (⊗ here is the Kronecker product). A is the phylogenetic relatedness matrix, I is an identity matrix and the V are 3×3 (co)variance matrices describing the (co)variances between viral titre at different temperatures. The phylogenetic covariance matrix, Vp, describes the inter-specific variances in each trait and the inter-specific covariances between them. The residual covariance matrix, Ve, describes the within-species variance that can be both due to real within-species effects and measurement or experimental errors. The off-diagonal elements of Ve (the covariances) can not be estimated because no vial has been subject to multiple temperatures and so were set to zero. We excluded D. pseudoobscura from the full model as data for BMR was not collected, but included it in models that did not include any fixed effects, which gave equivalent results.
Diffuse independent normal priors were placed on the fixed effects (means of zero and variances of 108). Parameter expanded priors were placed on the covariance matrices resulting in scaled multivariate F distributions, which have the property that the marginal distributions for the variances are scaled (by 1000) F 1,1. The exceptions were the residual variances for which an inverse-gamma prior was used with shape and scale equal to 0.001. The MCMC chain was run for 130 million iterations with a burn-in of 30 million iterations and a thinning interval of 100,000. We confirmed the results were not sensitive to the choice of prior by also fitting models with inverse-Wishart and flat priors for the variance covariance matrices (described in [9]), which gave qualitatively similar results (10.6084/m9.figshare.6177191). All confidence intervals (CI’s) reported are 95% highest posterior density intervals.
Using similar model structures we also ran a univariate model with BMR and a bivariate model with CTmin and CTmax as the response variables to calculate how much of the variation in these traits was explained by the host phylogeny. Both of these models were also run with wing length as a proxy for body size as this is known to influence thermal measures [59]. We observed significant levels of measurement error in the metabolic rate data; this was partially caused by respiratory chamber order during the assay. We corrected for this in two different ways. First, we fitted a linear model to the data to control for the effect of respiratory chamber number and then used this corrected data in all further models. We also used a measurement error model that controls for both respiratory chamber number effects and random error. Both of these models gave similar results although the measurement error model showed broad CIs suggesting the BMR data should be interpreted with caution. All datasets and R scripts with the model parameterisation are provided as supporting information (S1 Text).
To investigate the effect of temperature on virus host shifts we quantified viral load in 12,827 flies over 396 biological replicates, from 45 species of Drosophilidae at three temperatures (Fig 1). DCV replicated in all host species, but viral load differed between species and temperatures (Fig 1). Species with similar viral loads cluster together on the phylogeny (Fig 2). Measurements were highly repeatable (Table 1), with a large proportion of the variance being explained by the inter-specific phylogenetic component (vp), with little within species or measurement error (vr) (Repeatability = vp/(vp + vr): Low = 0.90 (95% CI: 0.84, 0.95), Medium = 0.96 (95% CI: 0.93, 0.98), and High = 0.95, (95% CI: 0.89, 0.98)). We also calculated the proportion of between species variance that can be explained by the phylogeny as vp/(vp+ vs) [71], which is equivalent to Pagel’s lambda or phylogenetic heritability [69,72]. We found the host phylogeny explains a large proportion of the inter-specific variation in viral load across all three temperatures, although these estimates have broad confidence intervals due to the model struggling to separate the phylogenetic and non-phylogenetic components (Low = 0.77, 95% CI: 0.28, 0.99; Medium = 0.53, 95% CI: 0.31×10−5, 0.85; High = 0.40, 95% CI: 0.99×10−5, 0.74)
To examine if species responded in the same or different way to changes in temperature we examined the relationships between susceptibilities across the different temperatures. We found strong positive phylogenetic correlations between viral loads across the three temperatures (Table 2). Our models showed that the variance in viral load increased with temperature, however the mean viral load showed no such upward trend (Table 1). This suggests that the changes in variance are not simply occurring due to an increase in the means, that is then driving an increase in variance.
The high correlations suggest the rank order of susceptibility of the species is not changing with increasing temperature. However, the change in variance suggests that although the reaction norms are not crossing they are diverging from each other as temperature increases i.e. the most susceptible species are becoming more susceptible with increasing temperature, and the least susceptible less so [73]. For example, D. obscura and D. affinis are the most susceptible species at all three temperatures. The responses of individual species show that some species have increasing viral load as temperature increases (Fig 1, e.g. Z. taronus, D. lummei), while others decease (e.g. D. littoralis, D. novamexicana).
The changes we observe could be explained by the increase in temperature effectively increasing the rate at which successful infection is progressing (i.e. altering where in the course of infection we have sampled). However, this seems unlikely as at 2 days post infection at the medium temperature (22°C), viral load peaks and then plateaus [41]. Therefore, in those species where viral load increases at higher temperatures the peak viral load itself must be increasing, rather than us effectively sampling the same growth curve but at a later time point. Likewise, in those species where viral load decreased at higher temperatures, viral load would need to first increase and then decrease, which we do not observe in a time course at 22°C [41]. To check whether this also holds at higher temperatures we carried out a time course of infection in a subset of six of the 45 original experimental species at 27°C, where we would expect the fastest transition between the rapid viral growth and the plateau phase of infection to occur (Fig B in S1 Text). This allowed us to confirm that the decreasing viral loads observed in some species at higher temperatures are not due to general trend for viral loads to decline over longer periods of (metabolic) time.
We quantified the lower and upper thermal tolerances (CTmin and CTmax) across all 45 species with 3 replicates per species. Neither CTmax nor CTmin were found to be significant predictors of viral load (CTmin -0.21, 95% CI: -0.79, 0.93, pMCMC = 0.95 and CTmax 0.31, 95% CI: -0.11, 0.74, pMCMC = 0.152). When treated as a response in models we found the host phylogeny explained a large proportion of the variation in thermal maximum (CTmax: 0.95, 95% CI: 0.84, 1) and thermal minima (CTmin: 0.98, 95% CI: 0.92, 0.99, see S1 Text Fig C).
We also measured the basal metabolic rate of 1320 flies from 44 species, across the three experimental temperatures, to examine how cellular function changes with temperature. BMR was not found to be a significant predictor of viral load when included as a fixed effect in our model (slope = 9.09, 95% CI = -10.13, 20.2689, pMCMC = 0.548).
BMR increased with temperature across all species (mean BMR and SE: Low 0.64 ± 0.02, Medium 1.00 ± 0.04, High 1.2 ± 0.04 CO2ml/min-1, see S1 Text Fig D).
When BMR was analysed as the response in models, the phylogeny explained a small amount of the between species variation (Low 0.19, 95% CI: 2 × 10−8, 0.55, Medium 0.10, 95% CI: 5 × 10−7, 0.27, High 0.03, 95% CI: 8 × 10−9–0.13, S1 Text Fig E) indicating high within species variation or large measurement error. Consequently the mean BMRs for each species, at each temperature, were used in the analysis of viral load will be poorly estimated and so the effects of BMR will be underestimated with too narrow credible intervals. To rectify this we ran a series of measurement error models, the most conservative of which gave a slope of -9.8 but with very wide credible intervals (-62.5, 42.6). Full details of these models are given in the Supporting Information (S1 Text).
We found that susceptibilities of different species responded in different ways to changes in temperature. The susceptibilities of different species showed differing responses as temperatures increased (Fig 1). There was a strong phylogenetic correlation in viral load across the three experimental temperatures (Table 2). However, the variance in viral load increased with temperature, whereas the mean viral load did not show the same trend. This suggests that the rank order of susceptibility of the species remains relatively constant across temperatures, but as temperature increases the most susceptible species become more susceptible, and the least susceptible less so.
Changes in global temperatures are widely predicted to alter host-parasite interactions and therefore the likelihood of host shifts occurring [5,21,47,74,75]. The outcome of these interactions may be difficult to predict if temperature causes a different effect in the host and pathogen species [18,37,76–78]. Our results show that changes in temperature may change the likelihood of pathogens successfully infecting certain species, although they suggest that it may not alter which species are the most susceptible to a novel pathogen.
The increase in phylogenetic variance with temperature is effectively a form of genotype-by-environment interaction [28,79–81]. However, it varies from the classically considered ecological crossing of reaction norms, as we do not see a change in the rank order of species susceptibly across the range of experimental temperatures. Instead, we find the species means diverge with increasing temperatures and so the between species differences increase [73,82]. It is also important to note that temperature may not simply be causing a change in effect size when considering the biological processes occurring during host-parasite interactions [22,83]. For example, virus replication may plateau at higher temperatures due to resource limitation. The observed level of susceptibility may be the combined outcome of both host and parasite traits, which may interact nonlinearly with temperature. We also note that by using a limited range of temperatures for practical reasons we may have not captured all unimodal relationships between viral load and temperature.
As temperature is an important abiotic factor in many cellular and physiological processes, we went on to examine the underlying basis of why viral load might change with temperature. Previous studies that found phylogenetic signal in host susceptibility were carried out at a single experimental temperature [9,41]. Therefore, the patterns observed could potentially be explained by some host clades being assayed at sub-optimal thermal conditions. We used CTmax and CTmin as proxies for thermal optima which, due to its multifaceted nature, is problematic to measure directly [84–86]. We also measured basal metabolic rate across three temperatures to see if the changes in viral load could be explained by general increases in enzymatic processes. We found that these measures were not significant predictors of the change in viral load with temperature. This may be driven by the fact that all temperature related traits are likely to be more complex than what any single measure can explore. Traits such as host susceptibility are a function of both the host and parasite thermal optima, as well as the shape of any temperature-trait relationship [37,78].
The host immune response and cellular components utilised by the virus are likely to function most efficiently at the thermal optima of a species, and several studies have demonstrated the outcomes of host-pathogen interactions can depend on temperature [26,28,76,81]. However, the mechanisms underlying the changes in susceptibility with temperature seen in this study are uncertain and a matter for speculation. Our results show that in the most susceptible species, viral load increases with temperature; this may be due to the virus being able to successfully infect and then freely proliferate, utilizing the host cells whist avoiding host immune defences. In less susceptible species viral load does not increase with temperature, and in some cases it actually appears to decreases. Here, temperature may be driving an increase in biological processes such as enhanced host immunity, or simply increasing the rate of degradation or clearance of virus particles that have failed to establish an infection of host cells.
We have investigated how an environmental variable can alter infection success following a novel viral challenge. However, temperature is just one of the potential environmental factors that will influence the different stages of a host shift event [8]. Using a controlled method of viral inoculation allows us to standardize inoculation dose so we can ask, given equal exposure, how does temperature affect the ability of a pathogen to persist and replicate in a given host? However, in nature hosts will be faced with variable levels of pathogen exposure, infected through various modes of transmission and often by multiple strains or genotypes [87]. Such variables may have consequences for the establishment and subsequent infection success of any potential host shift event. It is known that oral infection by DCV is stochastic and immune barriers such as the gut are important [55,88,89], therefore establishing the relevance of infection in the wild in this system would require further study using different potential routes of infection. The geographical distribution of a host will also influence factors such as diet and resource availability [28,90–93], and so further work on the role of nutrient and resource availability would therefore be needed to further explore the impact of these on potential host shifts.
In conclusion, we have found changes in temperature can both increase or decrease the likelihood of a host shift. Our results show the rank order of species’ susceptibilities remain the same across temperatures, suggesting that studies of host shifts at a single temperature can be informative in predicting which species are the most vulnerable to a novel pathogen. Changing global temperatures may influence pathogen host shifts; for example changes in distributions of both host and pathogen species may generate novel transmission opportunities. Our findings suggest that increases in global temperature could increase the likelihood of host shifts into the most susceptible species, and reduce it in others. Climate change may therefore lead to changing distributions of both host and pathogens, with pathogens potentially expanding or contracting their host range. Understanding how environmental factors might affect broader taxonomic groups of hosts and pathogens requires further study if we are to better understand host shifts in relation to climate change in nature.
|
10.1371/journal.pgen.1003448 | The Chromosomal Passenger Protein Birc5b Organizes Microfilaments and Germ Plasm in the Zebrafish Embryo | Microtubule-microfilament interactions are important for cytokinesis and subcellular localization of proteins and mRNAs. In the early zebrafish embryo, astral microtubule-microfilament interactions also facilitate a stereotypic segregation pattern of germ plasm ribonucleoparticles (GP RNPs), which is critical for their eventual selective inheritance by germ cells. The precise mechanisms and molecular mediators for both cytoskeletal interactions and GP RNPs segregation are the focus of intense research. Here, we report the molecular identification of a zebrafish maternal-effect mutation motley as Birc5b, a homolog of the mammalian Chromosomal Passenger Complex (CPC) component Survivin. The meiosis and mitosis defects in motley/birc5b mutant embryos are consistent with failed CPC function, and additional defects in astral microtubule remodeling contribute to failures in the initiation of cytokinesis furrow ingression. Unexpectedly, the motley/birc5b mutation also disrupts cortical microfilaments and GP RNP aggregation during early cell divisions. Birc5b localizes to the tips of astral microtubules along with polymerizing cortical F-actin and the GP RNPs. Mutant Birc5b co-localizes with cortical F-actin and GP RNPs, but fails to associate with astral microtubule tips, leading to disorganized microfilaments and GP RNP aggregation defects. Thus, maternal Birc5b localizes to astral microtubule tips and associates with cortical F-actin and GP RNPs, potentially linking the two cytoskeletons to mediate microtubule-microfilament reorganization and GP RNP aggregation during early embryonic cell cycles in zebrafish. In addition to the known mitotic function of CPC components, our analyses reveal a non-canonical role for an evolutionarily conserved CPC protein in microfilament reorganization and germ plasm aggregation.
| We address mechanisms by which germ cell precursors, a cell type that generates sperm and eggs for future generations, are specified in the zebrafish. Germ cell-specific genes are highly conserved across species, and in many animals germ cells are specified by the inheritance of germ plasm, a specialized cytoplasm containing specific proteins and RNAs corresponding to such conserved genes. Germ plasm is inherited as ribonucleoparticles, which are often present in the egg as singletons and which aggregate to generate larger masses that, when inherited by germ cell precursors, will initiate a germ cell-specific gene expression program. Here, we present the functional and molecular analysis of the zebrafish maternal gene, motley, which we show encodes a homologue of the Chromosomal Passenger Complex protein Survivin, or Birc5b. We found that, in addition to the expected role of this protein in cell division, characteristic of factors in this complex, Birc5b mediates germ plasm aggregation in the early zebrafish embryo through the coordination of dynamic changes in the cytoskeleton. Our studies provide a mechanistic basis to explain how germ cell determinants are transmitted from one generation to the next and reveal a non-conventional role for a Chromosomal Passenger Complex factor in this process.
| A fundamental feature of cell biology is cytoskeletal cross-talk between microtubule and microfilament networks. One key cellular process dependent on these interactions is the positioning of the contractile ring during cytokinesis. Two major groups of microtubules are involved in contractile ring positioning: the center of the mitotic spindle, which resolves into the antiparallel central spindle microtubules, and the poles of the mitotic spindle, which generates astral microtubules. Several studies indicate that central spindle and astral microtubules redundantly stimulate furrowing at the equatorial cortex [1]. Both sets of microtubules must ultimately communicate with cortical microfilaments that form the contractile ring, and the precise mechanism of this communication is an area of intense research. Candidate mediators include the Chromosomal Passenger Complex (CPC), which localizes with chromosomes during metaphase and transitions cortically to the prospective site of membrane ingression during telophase [2], [3]. Loss of CPC function affect two distinct yet related cellular events: chromosomes tend to lag during metaphase, resulting in chromosome segregation errors, and cleavage furrows fail to maintain ingression resulting in cytokinesis failures [4], [5], [6], [7], [8]. Lagging chromosomes can secondarily cause cytokinesis failures during telophase, but analysis of point mutations in CPC proteins reveal independent roles for components of this complex in the initiation of cytokinesis as well [9], [10], [11], in agreement with the localization of the CPC to the early equatorial cortex.
In addition to cytokinesis, a second major requirement of microtubule-microfilament cross-talk is for subcellular localization of proteins and/or mRNAs to either initiate developmental asymmetry during embryogenesis or to achieve a physiological output such as cell migration and axonogenesis. In animal eggs and early embryos, many ribonucleoparticles (RNPs) encode key cell-fate determinants, underscoring the importance of cytoskeletal function in pattern formation during embryogenesis. One such key molecular factor is the germ plasm, a specialized cytoplasm composed of a unique cohort of mRNAs and proteins. In several species including Drosophila, Xenopus, C. elegans, and zebrafish, the primordial germ cells (PGCs) form by selectively inheriting maternally derived germ plasm RNP (GP RNP) complexes [12]. Localization of GP RNPs is best characterized in Drosophila where they are transported on microtubules and anchored by microfilaments in a multi-step process that ensures localized germ cell specification [13], [14], [15]. Less is known about GP RNP localization in vertebrate species. Studies in zebrafish suggest that GP RNPs associate with cortical microfilaments, which organize in a microtubule-dependent manner into circumferential concentric rings that facilitate germ plasm aggregation [16]. However, the precise molecular mechanism(s) of cytoskeletal cross-talk that mediates this reorganization remain unknown.
Here, we describe a zebrafish maternal-effect mutant motley and identify it as birc5b, a zebrafish homolog of the mammalian CPC protein, Birc5/Survivin. Birc5b is subcellularly localized as a CPC protein and motley/birc5b mutants display meiotic and mitotic chromosome segregation errors and cell division phenotypes characteristic of failed CPC function. Additionally, motley/birc5b mutants fail to initiate cytokinesis furrow ingression as reflected by defects in astral microtubule reorganization at incipient furrows, confirming an early role for a CPC protein in furrow formation. Unexpectedly, motley/birc5b mutants also exhibit defects in microfilament reorganization in the embryo prior to initiation of the first cytokinesis furrow, and these defects are accompanied by a failure in GP RNP aggregation. In wild-type embryos, Birc5b protein localizes to the tips of astral microtubules contacting the cortex, where it also co-localizes with actin and GP RNPs. We propose a model in which Birc5b at astral microtubule tips mediates microtubule-microfilament interaction to achieve reorganization of cortical microfilaments and facilitate GP RNP aggregation prior to and during cytokinesis furrow initiation.
The mutations motley (motp1aiue) and p4anua, isolated in an ENU-induced mutagenesis screen for recessive zebrafish maternal-effect genes [17] display early cytokinesis defects in the embryo (Figure 1A–1F; Figure S1A–S1H). In this study, we present the identification and characterization of the motley mutation. Homozygous motley females mature into viable, fertile adults. However, embryos from such females (motley mutants herein) manifest a completely penetrant cell division defect, which results in lethality at ∼4 hours post fertilization (hpf). Live motley mutants were indistinguishable from wild-type embryos during the first 30 minutes post fertilization (mpf). However, shortly after, when the first cytokinesis furrow became visible in wild-type blastodiscs, motley mutant blastodiscs lacked a membrane indentation characteristic of furrow formation (Figure 1A, 1B). In early wild-type embryos at telophase, when furrow initiation occurs, immunolabeling for α-tubulin revealed arrays of microtubules at the incipient furrow during the first cell cycle (Figure 1C), which were absent in motley though karyokinesis appeared to have progressed (Figure 1D). In wild-type embryos, at this stage in furrow formation, a microtubule-free zone appears between abutting arrays of bundled microtubules at incipient furrows (shown for the second cell cycle in Figure 1E). In motley mutants at the same developmental stage, astral microtubules failed to bundle opposite each other resulting in a disorganized mesh (Figure 1F). By 2hpf, the cell adhesion molecule β-catenin accumulated at mature cleavage furrows in wild-type embryos (Figure S1B), a pattern that was absent in motley mutants (Figure S1D). Though DNA segregates in motley mutants initially (Figure 1D), the mitotic spindle itself was abnormal. In wild-type embryos, bipolar mitotic spindles always aligned sister chromatids at the metaphase plate (Figure 1G, 1I). In motley mutants, mitotic spindles were typically bent with chromosomal DNA aberrantly spread along its length (Figure 1H, 1J). DNA segregation defects in motley eventually manifested as unevenly distributed nuclear masses and chromosomal bridges by 2hpf (Figure S1D).
Linkage analysis mapped the motley locus to the zebrafish chromosome 23 where it was fully linked to the Simple Sequence Length Polymorphism (SSLP) marker z14967 in 565 meioses. We tested birc5b, a gene present in the vicinity of z14967, as the candidate locus affected in motley mutants. birc5b transcripts from wild-type eggs were of the expected ∼500 base pair (bp) size, however birc5b transcripts from motley eggs were ∼100 bp larger, indicating aberrant splicing (Figure 1K). Sequencing both transcripts revealed a single T to C transition at the highly conserved splice donor base pair GT in the second intron of the motley allele (Figure 1L–1N). This mutation results in alternative splicing at the GT base pair of a cryptic splice donor site 113 bps into the second intron (Figure 1N). While wild-type Birc5b protein is predicted to be a 144 amino acid product, the motley mutant allele is expected to result in a truncated Birc5b protein containing the first 79 wild-type residues followed by mis-translation from the point of intron insertion, yielding a mis-translated, truncated protein of 111 residues. The CX2CX16HX6C signature BIR domain of Birc5b spans the second and third exons and the mutation disrupts the BIR domain from residue S80 onwards, resulting in loss of the C-terminal part of the BIR domain and protein, including conserved H82 and C89 residues within the BIR domain (Figure 1N). Protein sequence analyses indicate that motley/Birc5b is a zebrafish homologue of the Baculoviral IAP protein, Birc5/Survivin (Figure S1L). The two paralogues in zebrafish, Birc5a and Birc5b are 51.4% and 46.8% similar to human BIRC5/Survivin, respectively (Figure S1L). Consistent with previous observations that both paralogues were maternally expressed [18] we isolated both birc5a and birc5b transcripts from wild-type embryos during early development at 1 and 4hpf (Figure S1I). RT-PCR analysis indicates that birc5a continues to be expressed during development at 24hpf and beyond, whereas birc5b transcripts were undetectable at these later stages (Figure S1I).
Survivin and its homologs are required for meiosis in both vertebrates and invertebrates [5], [19], [20], which prompted us to assay for meiosis defects in eggs from motley mutant females (Figure 2). Mature zebrafish eggs are arrested at metaphase of meiosis II, and egg activation, which occurs upon contact with water, results in meiosis resumption and release of the second polar body. We compared this process in unfertilized water-activated eggs from wild-type and homozygous motley females by a time-course analysis of meiosis II completion. In wild-type eggs at 5 minutes post activation (mpa), a meiotic spindle can be observed with the sister chromatids aligned at the spindle poles during anaphase (Figure 2A–2C). By 10mpa, the spindle apparatus bundles as the meiotic midbody between one set of condensing sister chromatids and a second set of decondensing sister chromatids (Figure 2D–2F). By 20mpa, a nuclear membrane forms around the decondensed DNA to generate the female pronucleus, while the condensed DNA remains tightly associated with the meiotic midbody and becomes the polar body (Figure 2G–2I). In motley eggs at the same time points, the meiotic spindle exhibits defects similar to those of its mitotic counterpart. At 5mpa, the meiotic spindle in motley eggs is bent and chromosomes spread along the spindle instead of being aligned at the poles (Figure 2J–2L). At 10mpa an incipient midbody-like microtubule bundle forms only occasionally between sister chromatid sets, both of which appear equally condensed in contrast to wild-type (Figure 2M–2O). At 20mpa, a meiotic midbody is not observed in motley eggs and both resulting nuclei appear highly condensed and connected by chromosomal bridges (Figure 2P–2R; Figure 3G–3I). Thus, in addition to its role in mitosis, maternal Birc5b is also required for meiotic spindle organization and successful meiosis completion. Despite defects in the completion of meiosis II and the aberrant appearance of the female pronucleus, the oocyte nucleus from motley mutants is able to fuse with the male pronucleus in fertilized embryos to form the zygotic nucleus (data not shown).
We validated that the molecular lesion in motley affected Birc5b function by providing motley eggs and embryos with exogenous wild-type birc5b mRNA (Figure 3; Figure S2). To assay for meiosis rescue in motley eggs, immature stage IV oocytes were isolated from homozygous motley females and microinjected with wild-type birc5b::eGFP mRNA. birc5b::eGFP mRNA injected motley oocytes express Birc5b::GFP fusion protein within 1 hour post injection (hpi) (Figure S2A–S2C). At 3hpi, motley oocytes had fully matured into translucent eggs that continued to exhibit strong Birc5b::GFP protein expression (Figure S2D–S2F), and activated normally upon contact with water (Figure 3A–3C). In these Birc5b::GFP-expressing motley eggs, we were able to unambiguously detect a distinct larger female pronucleus and a condensed polar body with its associated meiotic midbody (Figure 3D–3F), indicating that both female pronuclear decondensation and midbody formation defects were rescued. In contrast, motley eggs derived from an uninjected subset of oocytes exhibited two highly condensed nuclear bodies connected by chromosomal bridges and had no detectable meiotic midbody (Figure 3G–3I), similar to the phenotype observed in eggs that mature within homozygous motley females (Figure 2P–2R).
A subset of motley eggs injected with birc5b::eGFP mRNA during in vitro oogenesis were in vitro fertilized to assay for rescue of post-fertilization mitotic phenotypes. In such Birc5b::GFP-expressing motley embryos, normal cytokinesis furrows corresponding to the first two cell cycles were observed at ∼60mpf (Figure 3J–3L). These results indicate that exogenous wild-type birc5b::eGFP injected into immature oocytes rescues motley mutant phenotypes both during oogenesis and early embryogenesis.
We also observed late rescue of motley-associated phenotypes when birc5b::eGFP mRNA was injected into embryos at the 1-cell stage. At ∼2hpf several cytokinesis furrows were observed in motley mutants expressing Birc5b::GFP (Figure S2G–S2I), whereas uninjected sibling motley embryos did not exhibit cleavage furrows at any stages (Figure S2J). The timing of furrow formation in birc5b::GFP injected motley embryos, ∼80 minutes after the normal initiation of cell division in wild-type embryos, coincides with the appearance of Birc5b::GFP fluorescence in injected embryos and likely reflects a lag in translation of the injected mRNA to generate sufficient protein for rescue. We characterized this late-stage cell division rescue by comparing the spindle shape, linear spindle pole-to-pole distance (SPD, as a measure of spindle bending) and the presence of midbodies between Birc5b::GFP-expressing motley, uninjected siblings and wild-type embryos. At 2hpf, wild-type mitotic spindles were normal and SPD measured ∼27.8 µm (Figure 3M, 3P), while in motley embryos the spindles were bent with a significantly reduced SPD of ∼24.5 µm (Figure 3N, 3P). In motley mutants expressing Birc5b::GFP, mitotic spindle morphology was rescued to wild-type and the SPD measured ∼26.7 µm approaching the wild-type measurements (Figure 3O, 3P). Additionally, midbody formation was also rescued in motley mutants injected with birc5b::GFP at the 1-cell stage. During cleavage stages in wild-type embryos, midbodies were readily observed (Figure S2L), while midbodies were never observed in motley embryos due to failed cytokinesis (Figure S2M). However, in Birc5b::GFP-expressing motley embryos midbodies were detected, indicating that the cytokinesis furrows seen in live motley mutants at ∼2hpf could transition into mature cleavage furrows (Figure S2N, S2O). Thus, exogenous Birc5b can rescue all aspects of the mutant phenotypes observed in motley embryos during meiosis II and early embryonic cell divisions, confirming that the locus affected in motley is birc5b.
Since both birc5a and birc5b are maternally present in zebrafish embryos as mRNA (Figure S1I; [18]) and protein (Figure S3A; see below), we asked whether these duplicate genes might share a common function. In contrast to the case with Birc5b::GFP protein, expression of Birc5a::GFP during oogenesis is unable to rescue the motley/birc5b mutant phenotype (Figure S3B–S3J). This is in agreement with the observed fully-penetrant maternal-effect phenotype caused by the motley/birc5b mutation in spite of the fact that Birc5a expression is not affected in this background (Figure S1 and data not shown). Together, these data suggests subfunctionalization for the Birc5a and Birc5b paralogs with respect to cytokinesis in the early embryo.
Prior to ascertaining the subcellular localization of Birc5b protein by immunolabeling, we identified antibodies that specifically recognized this protein. We performed western blot analysis using two commercially available monoclonal antibodies: anti-Survivin, developed against full-length human Survivin, and anti-Survivin-BIR, developed against the BIR domain present in human Survivin. Our data indicate that the anti-Survivin and anti-Survivin-BIR antibodies recognize Birc5a and Birc5b, respectively, without detectable non-specific cross-reactivity (Figure 3SA).
We assayed the subcellular localization of Birc5b using the anti-Survivin-BIR antibody, which our western analysis suggests is specific to Birc5b. Immunolabeling during meiotic anaphase in wild-type eggs at 5mpa revealed that Birc5b protein localized to the central region of the meiotic spindle (Figure 4A–4E). At 20mpa, the meiotic midbody associates with the forming polar body (Figure 4F–4J). Interestingly, Birc5b protein always localized distinctly to the end of the meiotic midbody farthest from the polar body, presumably at the site of meiotic cytokinesis (Figure 4H–4J). During early mitosis in the embryo, co-immunolabeling with the CPC protein Aurora Kinase B (AurB) revealed that Birc5b co-localized with AurB at the mitotic spindle during metaphase (Figure 4K–4O). During cytokinesis, Birc5b translocated to the cortex, where it localized to bundled microtubule ends abutting at the incipient cleavage furrow (Figure 4P–4R). During these early cell divisions, AurB also localizes to the bundled tips of microtubules at the furrow [21], along with Birc5b (data not shown). Both Birc5b and AurB continued to colocalize in midbodies during mid-cleavage stages at 2hpf (Figure 4S–4V).
The subcellular expression of Birc5b protein is consistent with its inferred function as a CPC protein required for DNA segregation, spindle morphology, and cytokinesis during meiosis and mitosis in the early zebrafish embryo. Additionally, localization of Birc5b to the tips of bundled microtubules is consistent with an additional role for this CPC protein in facilitating microtubule remodeling for initiating furrow ingression, in agreement with previous studies [9], [10], [11].
In zebrafish embryos, prior to and during the first mitosis, the microfilament and microtubule cytoskeletons grow dynamically in a concerted manner, which is evident in the reorganization of cortical microfilaments before the end of the first zygotic cell cycle ([16]; Figure 5). Immediately upon fertilization, the sperm derived centrioles nucleate a sperm monoaster near the fusing pronuclei, microtubules from which grow towards the cortex [16]. 0.5 µm optical cross-sections through the wild-type blastodisc cortex reveal monoastral microtubule tips suggesting that the majority of microtubules grow and terminate at the cortex (Figure 5A). During the same time-frame, F-actin seed filaments at the center of the blastodisc cortex move towards the periphery creating an actin-free zone (Figure 5B, 5C). In motley/birc5b mutants, a 0.5 µm section through the blastodisc cortex reveals sperm monoaster microtubules, which are aberrantly located along the cortical plane, and an absence of microtubule tips at the cortex (Figure 5D). Analysis of the actin cytoskeleton during this time-frame show that F-actin seed filaments fail to clear from the center of the blastodisc cortex; instead the cortex remains mottled with F-actin seed filaments (Figure 5E, 5F). As the sperm monoaster disappears, the first embryonic mitosis proceeds in wild-type embryos and the mitotic spindle poles resolve into sets of astral microtubules, which like microtubules of the sperm monoaster, also radiate towards the blastodisc cortex (Figure 5G). Again, coincident with the cortical astral microtubule growth, cortical F-actin organizes into concentric rings at the blastodisc periphery (Figure 5H, 5I; [16]). Higher magnification views of the two cytoskeletons reveal microtubule tips at the cortex in a 0.5 µm optical section (Figure 5J, 5M, 5N), a subset of which are in contact with cortical microfilaments arranged in unbranched concentric rings at the periphery (Figure 5K, 5L, 5O). During the first zygotic mitosis in motley/birc5b mutant embryos, spindle pole astral microtubules reach the cortex, but as described earlier (Figure 1D), fail to resolve into the two sets (Figure 5P) characteristic of the first mitosis (Figure 1C, Figure 5G). Furthermore, similar to the sperm monoaster microtubules, in motley/birc5b mutants, spindle pole astral microtubule tips were also not detectable at the cortex (Figure 5S). Analysis of the cortical microfilaments during the first zygotic mitosis in motley/birc5b reveal that the microfilaments fail to organize into peripheral rings and are instead found ectopically in the center of the blastodisc cortex (Figure 5Q, 5R). Higher magnifications of the cortical cytoskeleton additionally revealed that the microfilaments are branched and as the microtubule ends are not seen at the cortex, do not colocalize with the tips (Figure 5T, 5U). These cortical cytoskeletal defects indicate a role for maternal Birc5b in microtubule dynamics at the blastodisc cortex, and an additional novel role for this CPC protein in cortical actin cytoskeleton rearrangements prior to the first zygotic mitosis.
Because of the postulated role for the cortical cytoskeleton on zebrafish germ plasm localization [16] and the function of motley/birc5b in cortical cytoskeletal reorganization, we tested whether germ plasm RNP segregation is affected in motley/birc5b mutants. We first corroborated that during the early embryonic cell divisions, germ plasm mRNAs such as nanos localize to cortical microfilaments at the blastodisc periphery (Figure S4), as had been previously posited [16]. We also discovered that an antibody against the human phosphorylated non-muscle myosin (NMII-p) labeled the distal furrow where GP RNPs are recruited in the 2- (Figure S5A–S5C) and 4-cell embryos. Double labeling experiments showed that the anti-NMII-p label co-localized with germ plasm mRNAs vasa, dead end (dnd) and nanos at the furrow as well as to non-furrow regions at the cortical periphery (Figure S5D–S5U). The functional relevance of the anti-NMII-p label to GP RNPs is currently under investigation, however, these observations indicated that the anti-NMII-p antibody serves as a convenient probe to detect GP RNPs in the early zebrafish embryo. We infer from the localization of NMII-p with all three tested germ plasm mRNAs that most GP RNPs may contain the same basic molecular components.
In wild-type embryos immediately upon fertilization (Figure S6A; [16]) and in unfertilized embryos (data not shown), GP RNPs are distributed as a broad cortical band surrounding a GP RNP-free zone at the center of the blastodisc. The underlying basis for this initial distribution is not known but may reflect intrinsic differences in the egg cortex established during oogenesis. Upon fertilization, the GP RNP-free zone is seen to expand outwardly, through a process we have previously proposed involves microtubules from the sperm monoaster and the spindle pole asters pushing growing microfilaments and associated GP RNPs away from the center of the blastodisc, generating an increasingly narrow and more peripherally located band of GP RNPs and microfilaments (Figure S6B, S6C; [16]). A failure in this proposed process is reflected in motley/birc5b mutants (Figure S6) and nocodazole-treated embryos (data not shown), where the initial broad cortical band of GP RNPs remain unaffected (Figure S6D), but the central GP RNP-free zone does not appear to expand, so that aggregates continue to exhibit a broad cortical distribution (Figure S6E, S6F) similar to that observed in the egg/embryo immediately after activation/fertilization. As expected from the furrow initiation defect, GP RNPs do not undergo furrow recruitment in motley/birc5b mutants (Figure S6E, S6F). In situ hybridization analysis to detect GP RNAs indicates similar defects in motley mutants (Figure S6J, S6K and data not shown). Together, these observations indicate that motley/birc5b mutants have defects in GP RNP segregation prior to and independent of their recruitment to the cleavage furrows.
As shown in Figure 5, cortical microtubule tips are in contact with peripheral microfilaments. We next tested the spatial relationship between cortical GP RNPs and microtubule ends. We found that prior to, and during the first 2–3 cell cycles, single and multimerized GP RNPs localized to tips of the monoastral and spindle pole astral microtubules at the cortex (Figure 6D–6F). We had previously postulated that cortical F-actin reorganization facilitates GP RNP multimerization prior to furrow formation [16]. Given the cortical microfilament rearrangement defects in motley/birc5b, and the localization of microfilaments and GP RNPs to microtubule tips, we asked whether GP RNP multimerization at the cortex would be affected in motley/birc5b. We tested this by comparing the degree of GP RNP aggregation (as determined by the number of GP RNPs that appear to be physically adjoined as multimerized aggregates), in motley/birc5b mutants to that in wild-type embryos (Figure 7). As described in the preceding section, in wild-type embryos GP RNPs are found in a band at the periphery of the cortex (Figure 6B; Figure S4E; Figure S6A–S6C) and at the tips of astral microtubules (Figure 6D–6F). Within this band, aggregation occurs such that multimerized GP RNPs are located at an apparent wave front (closer to the blastodisc center) and GP RNP singletons in more peripheral regions (closer to the blastodisc edge) (Figure 7A–7D; [16]). We analyzed GP RNP multimerization semi-quantitatively by dividing the embryo into four quadrants and imaging four random regions of interest (ROIs) within the aggregation wave in each quadrant at 300×. GP RNPs in physical contact with each other were considered as multimeric aggregates. Aggregation analysis of the GP RNPs in wild-type embryos revealed a quantal progression of multimerized GP RNP ranging from single GP RNP to multimeric aggregates of up to 17 GP RNPs (Figure 7C, 7D, 7M). In motley/birc5b mutants, the peripheral cortical band of GP RNPs exhibited a significant change in composition with an increase in the numbers of single GP RNP and a decrease in the numbers of multimeric GP RNP aggregates (Figure 7G, 7H, 7M). The largest multimerized GP RNPs in motley/birc5b mutants consisted of ∼7 GP RNPs compared to multimeric aggregates of ∼17 GP RNPs found in wild-type embryos (Figure 7M). Microtubule depolymerization by nocodazole treatment decreased GP RNP multimerization to an extent comparable to that seen in motley/birc5b mutants (Figure 7K–7M). Thus, Birc5b, and as expected, microtubules are required for multimerization of GP RNPs at the periphery of the blastodisc cortex, prior to their recruitment at the cleavage furrow.
The cortical cytoskeletal and GP RNP multimerization defects in motley/birc5b indicate an essential role for cortical microtubules and an additional specific role for Birc5b as a mediator of microtubule-dependent cortical microfilament rearrangements and GP RNP multimerization. This process occurs both prior to (when the sperm monoaster forms) and during early zygotic mitoses (when the spindle pole asters form), prompting us to assay for subcellular localization of Birc5b at the cortex during these early stages.
In wild-type embryos, Birc5b localized to the tips of cortical sperm monoaster microtubules (Figure 8A–8C, 8K–8M), and co-localized with F-actin seed filaments (Figure 8B, 8D, 8E), and GP RNPs (Figure 8L, 8N, 8O). In motley/birc5b mutants, Birc5b continued to co-localize with the F-actin seed filaments (Figure 8G, 8I, 8J) and GP RNPs at the cortex (Figure 8Q, 8S, 8T). Testing colocalization to microtubule tips was precluded by our inability to detect the microtubule tips themselves in motley/birc5b mutants (Figure 8F, 8P). However, the lack of peripherally-directed clearing of F-actin seed filaments (Figure 5) and associated GP RNPs (Figure S6) and the failure of GP RNP aggregation in motley/birc5b mutants (Figure 7) are consistent with a lack of interaction between Birc5b and associated factors to the tips of growing astral microtubules. Consistent with a role for Birc5b in GP RNP multimerization prior to and during furrow formation, colocalization of Birc5b to GP RNPs in wild-type embryos was maintained during furrow formation but was not observed after the germ plasm was fully compacted at the 4-cell stage (data not shown).
As cleavage furrows form, GP RNPs are recruited to the furrow, forming a rod-like structure at the distal end of the furrow [22], [23], [24]. Given our findings that GP RNPs localize to tips of peripheral astral microtubules prior to (Figure 8K–8O) and during (Figure 6D–6F) furrow formation, we also characterized the association of the GP RNPs to microtubules at the furrow region. Furrow induction occurs when astral microtubules from each side of the bipolar spindle reach the prospective site of cytokinesis at the equatorial cortex [21]. At this incipient furrow, GP RNP aggregates can be observed bound to furrow astral microtubule ends in two rows abutting the furrow center (Figure 6G–6I), likely representing aggregates from each opposing set of astral microtubules that meet at the furrow. This observed bilateral arrangement of GP RNPs support a previously proposed hypothesis that furrow recruitment of GP RNPs involves their gradual gathering near the furrow, enriched by the action of opposing sets of furrow-associated astral microtubules [25].
In summary, our analyses of the zebrafish maternal-effect mutation motley identifies it as Birc5b, one of two zebrafish paralogs of the mammalian CPC protein, Survivin/Birc5. Phenotypic and subcellular localization studies suggest a novel function for maternal Birc5b in establishing contact between tips of cortical astral microtubules and cortical F-actin seed filaments and GP RNPs in the early zebrafish embryo. This contact ensures the concerted growth of microtubules and polymerizing microfilaments at the cortex, which results in the reorganization of cortical cytoskeleton to facilitate GP RNP multimerization.
The Chromosomal Passenger Complex (CPC) consisting of AurB, INCENP (Inner Centromere Protein), Borealin/Dasra and Survivin/Bir1/BIRC5 has been ascribed a number of roles during cell division, including chromosome bi-orientation and cytokinesis [2], [3]. Survivin is a member of the Baculoviral Inhibitor of Apoptosis Repeat Containing (BIRC) protein family and contains a single CX2CX16HX6C BIR domain [26]. The BIR domain is an evolutionarily conserved Zn finger fold present in Inhibitor of Apoptosis (IAP) proteins from Baculoviruses to humans [27], [28], [29]. Survivin/BIRC5 is unique amongst CPC and BIRC proteins as it is thought to be directly involved in both cytokinesis and cell survival, though its exclusivity to each process is debated [30].
Here, we show that birc5b is largely expressed maternally in zebrafish and that a maternal-effect mutation motley, corresponds to birc5b. Our characterization of the motley mutation reveals maternal functions for motley/birc5b during meiosis II completion in oocytes and early embryonic cell divisions, due to its conserved cell cycle-associated CPC activity in zebrafish. Additionally, we uncover a novel role for motley/birc5b in the interaction between astral microtubules, F-actin and germ plasm RNPs, which is essential for cytoskeletal rearrangements and GP RNP aggregation immediately after fertilization and during furrow initiation.
Of the two zebrafish paralogs, birc5b is expressed exclusively maternally (this study), as opposed to birc5a, which is expressed both maternally and zygotically throughout development ([18], this study). These observations are consistent with the observed maternal-effect phenotype of motley/birc5b, as well as with morpholino knockdown analysis that show an essential zygotic role for birc5a, but not birc5b, during late embryogenesis [18], [31], [32]. The defects observed in motley/birc5b mutant embryos, despite the presence of endogenous maternal Birc5a protein and our inability to rescue motley/birc5b by exogenous maternal expression of Birc5a, indicate that Birc5b has unique maternally-derived functions required during early embryonic cell divisions for GP RNP aggregation. Our analysis is also consistent with previous reports for a requirement for CPC proteins in meiosis in oocytes and sperm of both vertebrates and invertebrates [5], [6], [19], [20], [33], [34], [35], and highlights a dedicated role for the maternal motley/birc5b paralog in this process in the zebrafish. Even though both birc5b and birc5a are expressed during spermatogenesis in the zebrafish (S.N. unpublished data), fertilization rates in crosses with homozygous motley mutant males appear unaffected, suggesting that that Birc5b may not have a role during spermatogenesis, or that it functions redundantly with Birc5a in this process. We can not rule out, however, that the mutation results in limited sperm production that is obscured by excess sperm during fertilization, and further studies will be required to address a potential role of this gene during spermatogenesis.
Our analysis in the early embryo relies on antibodies that in western analysis recognize Birc5a and Birc5b forms without cross-reactivity and therefore should be specific to each gene homolog. However, we can not rule out the possibility that these antibodies are cross-reactive in fixed embryos and therefore that labeling signals to detect Bir5b in wild-type and birc5b/motley mutants are not caused by the presence of Birc5a. In spite of this uncertainty, our experiments indicate that Birc5b, but not Birc5a, has a functional role in GP RNP segregation and early embryonic cell division. Recent studies have highlighted the functional overlap and gradual transition between maternal products involved in female meiosis and the early embryonic cell cycles [36], and birc5b/motley may constitute an example of intergenerational overlap in gene function, with birc5b/motley acting during maternal meiosis and the early embryonic cycles and birc5a acting at later embryonic stages. Future studies will determine the precise Birc5 forms present in GP RNPs and whether these are associated with other CPC components.
The CPC is present diffusely along chromatin and its centromeric concentration during mitotic metaphase is essential for its function in sister chromatid segregation [2]. The BIR domain of Survivin and several conserved residues within it are required for CPC centromeric concentration [37], [38], [39], [40], [41], [42]. The splice site mutation in the motley mutant allele leads to protein mis-translation of Birc5b, truncating its BIR domain and eliminating key conserved residues in the zinc-finger fold. This missing conserved protein domain likely results in the chromosomal segregation defects observed in this mutant both during meiosis and embryonic mitosis. The mitotic chromosome segregation errors in the zebrafish mutant motley reflect an evolutionarily conserved role for Birc5b and the CPC during mitosis in the zebrafish embryo.
During anaphase and telophase, CPC proteins localize to the central spindle and overlying equatorial cortex, where their expression precedes the actomyosin assembly required for cytokinesis [2]. A number of studies using separation-of-function alleles have revealed that, in addition to a well-described role in furrow completion [3], CPC function is important for furrow initiation [9], [10], [11]. Our studies confirm these conclusions, since cytokinesis furrows never initiate ingression in motley/birc5b mutants. Previous studies have shown that, even with defective DNA segregation, zebrafish embryos undergo normal cytokinesis through signals from asters formed from duplicating centrosomes [21], [43], [44], [45], [46], indicating that the cytokinesis defects observed in motley mutants are unlikely caused by prior defects in meiosis or mitosis.
It has been proposed that furrow ingression is triggered by low microtubule density at the cortex, achieved by local bundling of microtubules and/or by separation of astral microtubules [47]. Indeed, a clear boundary free of astral microtubule ends is normally established along the site of furrow initiation in the early zebrafish blastodisc. In motley/birc5b mutant embryos during anaphase and telophase, this microtubule-free boundary fails to be established as astral microtubules from each half of the spindle fail to separate distinctly, and are instead found as an interwoven mesh. Furthermore, in motley/birc5b, abutting microtubules fail to bundle at the equatorial cortex, as it normally occurs in wild-type embryos. Together with these findings, our observation that Birc5b protein localizes to the tips of bundled microtubules of the incipient furrow at the equatorial cortex suggest that Birc5b may play a direct role in initiating furrow ingression by facilitating low microtubule densities at incipient furrows.
Despite requirements of the CPC in actomyosin contractile ring formation and/or function, very little is known about potential interactions between the CPC components and microfilaments. The striking failure in cortical microfilament reorganization in motley/birc5b mutants is the first direct evidence that a member of the CPC is required for actin cytoskeleton rearrangements. In early zebrafish embryos, this cortical microfilament reorganization appears to be essential for GP RNP multimerization, which may facilitate efficient GP RNP recruitment into cytokinesis furrows. The resulting aggregation and subcellular localization of GP RNPs during early embryonic divisions is integral to the selective inheritance of this cell fate determinant at later stages of zebrafish development. Immunoprecipitation analysis has failed to detect a direct interaction between Survivin and F-actin (S.N. unpublished). However, GP RNPs are known to become anchored to the actin cytoskeleton in a variety of systems [48] and one possible scenario is that the association between Survivin and F-actin is mediated by other GP RNP components.
Animal embryos regulate transmission of germ plasm by restricting its localization to specific sites either in the mature egg or in the post-fertilized embryo. In Drosophila oocytes, oskar mRNA as well as RNPs containing vasa and nanos are transported during oogenesis along microtubules to the posterior pole of the oocyte, where they become anchored to the actin cortex [13] prior to their incorporation into primordial germ cells at the posterior pole of the embryo [14], [15]. In Xenopus the germ plasm is transported during oogenesis to the vegetal pole through association with a specialized cytoplasm called the mitochondrial cloud, resulting in the anchoring of the germ plasm at the vegetal pole cortex. After fertilization, Xenopus germ plasm undergoes aggregation at the vegetal pole in a process dependent on microtubules and the kinesin-like protein Xklp1 [49]. In zebrafish oocytes, germ plasm components such as vasa, nanos, dazl also localize to the mitochondrial cloud (Balbiani body), to become associated with the cortex [50]. While some mRNAs, such as dazl, maintain their association with the vegetal pole, others acquire a more dispersed pattern and become redistributed to the blastodisc cortex in mature eggs [50]. The mechanism for this early pattern of localization of germ plasm components in zebrafish eggs is currently uncharacterized.
In Drosophila, continued localization of oskar and nanos mRNA, and Oskar and Vasa protein requires a microfilament-dependent anchor [13]. In early zebrafish embryos GP RNPs are initially distributed within a wide band at the periphery of the blastodisc cortex where the microfilaments are arranged in concentric, overlapping rings. Microtubule depolymerization disrupts the cortical microfilaments and GP RNP multimerization. Based on observations of the dynamic changes in the cortical cytoskeleton and germ plasm mRNAs upon pharmacological treatments, it was proposed that cortical astral microtubule ends push microfilaments towards the periphery, a rearrangement that facilitates GP RNP multimerization at the cortical periphery [16]. However, direct demonstration of the cortical arrangement of microtubules and microfilaments and the mechanism by which such cytoskeletal cross-talk facilitates GP RNP aggregation remained to be elucidated.
In this study we show that the ends of astral microtubules at the cortex contact cortical microfilaments, providing support for the hypothesis that expanding astral microtubules push polymerizing microfilaments away from the center of the blastodisc. In motley/birc5b mutants, cortical microfilaments are disorganized and GP RNP multimerization is severely reduced, reinforcing previous observations that an intact microfilament network is essential for this process [16]. The present study identifies maternal Birc5b as a molecular mediator of microtubule-microfilament interactions in the early zebrafish embryo. We propose a model wherein Birc5b is present at the blastodisc cortex possibly in a complex with GP RNPs and/or F-actin seed filaments prior to first embryonic mitosis (Figure 9A). As sperm monoaster microtubules reach the cortex they may make contact with Birc5b, which couples the polymerizing f-actin filaments to the tips of peripherally expanding astral microtubules (Figure 9B). This begins the re-positioning of the microfilaments to the cortical periphery where they are required to facilitate GP RNP multimerization (Figure 9B). Microfilament repositioning continues during the first mitosis and is now mediated by spindle pole astral microtubules, which facilitate the ongoing multimerization of GP RNPs during the first 2–3 cleavage cycles (Figure 9C). Multimerized GP RNPs then become enriched at the forming furrow by recruitment at the ends of abutting furrow microtubules (Figure 9C). In motley/birc5b mutants, the Birc5b complex with actin and GP RNPs still form and asters appear to grow normally (Figure 9D). However, we hypothesize that mutant Birc5b is unable to associate with the cortical microtubule ends (Figure 9E). In the mutants, this effectively uncouples astral microtubule ends from the polymerizing F-actin seed filaments at the cortex (Figure 9D–9F) and results in the observed defects in microfilament reorganization and RNP multimerization (Figure 9F).
In motley/birc5b mutants, we also find that astral microtubules extend along the cortex suggesting that in mutants they may be unable to respond to a cortical signal that would otherwise cause them to undergo dynamic instability and terminate their growth (Figure 9B, 9E). This inference is further supported by the failure of astral microtubules at the incipient cytokinesis furrow in motley/birc5b mutants to bundle and terminate growth past their partners from the contralateral side (Figure 9C, 9F). The observation that Birc5b protein localizes to the ends of both astral microtubule tips and bundled cleavage furrow microtubules is consistent with a role for Birc5b in the regulation of microtubule dynamics at both locations in the developing zebrafish embryo.
This study of zebrafish maternal Birc5b provides novel insights into the functions of a conserved CPC protein. Particularly, Birc5b appears to be a key mediator of microtubule-microfilament interactions, a cross-talk that is fundamental to the dynamic cytoskeletal rearrangements facilitating germ plasm subcellular localization prior to and during cytokinesis furrow initiation.
All animals were handled in strict accordance with good animal practice as defined by the relevant national and/or local animal welfare bodies, and all animal work was approved by the appropriate committee (University of Wisconsin – Madison assurance number A3368-01).
Wild-type AB, WIK and motley mutant fish lines were maintained under standard conditions at 28.5°C. All experiments other than the linkage mapping of motley were carried out using embryos from AB fish.
Homozygous motley mutant males were crossed to WIK females to raise F1 fish, which were then in-crossed to obtain the F2 mapping generation. Embryos from F2 females were screened for the syncytial phenotype at 2–4hpf. Genomic DNA from F2 females that produced syncytial mutant clutches was analyzed for segregation of a pan genomic panel of 244 SSLP markers to link and map the motley lesion to a genomic locus. The SSLP markers z13363 and z31657 flanked the motley locus and z14967 was the closest linked marker on chromosome 23. The motley mutation was maintained by crossing homozygous mutant males to heterozygous females.
Total RNA was isolated from wild-type and motley eggs using the TRIZOL reagent (Invitrogen). cDNA was synthesized using an oligodT primer and AMV reverse transcriptase (Promega). birc5b alleles from wild-type and motley were amplified from cDNA with primers 5′-cagcaatccacacgcaaccagg and 5′- gaagatcaaataagagctctcaaatttttgctagtggc using the Easy-A polymerase (Agilent Technologies). birc5b PCR products were ligated into the pGEM-Teasy vector (Promega) for further analyses. Anti-sense birc5b WT mRNA for whole mount in situ hybridizations was synthesized from the T7 promoter in pGEM-Teasy after linearizing with PstI. birc5b::eGFP fusion construct was created by subcloning the wild-type birc5b allele from pGEM-Teasy using primers 5′-GAATTCatgtcgaacacagacgttatcgc-3′ and 5′-CTCGAGaataagagctctcaaatttttgctagtgg-3′ that contained EcoRI and XhoI restriction sites respectively. The eGFP coding sequence was subcloned from pEGFP-N1 plasmid (GenBank U55762.1) originally from Clontech using primers 5′-CTCGAGatggtgagcaagggc-3′ and 5′-TCTAGAttacttgtacagctcgtccatgcc-3′ that contained XhoI and XbaI restriction sites respectively. Subcloned birc5b and eGFP were then sequentially cloned in frame into the expression vector pCS2p+. Sense mRNA for rescue experiments was synthesized from the SP6 promoter in pCS2p+ after linearizing with NotI, using the mMessage mMachine kit (Ambion).
All clones were sequenced using BigDye Terminator sequencing and analyzed using 4Peaks and DNAStar Lasergene program suites. Protein sequences were compared using ClustalX and the phylogenetic tree was visualized using TreeView. We cloned Birc5a and Birc5b as proteins of 190 and 144 amino acids respectively, each from contiguous maternal transcripts. This corresponds to the zebrafish sequence information available at http://vega.sanger.ac.uk/index.html. However, the sequence information at NCBI lacks the first exon for both Birc5a (GI:68085121) and Birc5b (GI:76780231), listing the proteins as containing 142 and 128 amino acids respectively.
The detailed experimental procedure for injecting and culturing stage IV zebrafish oocytes is available elsewhere [51], [52]. Briefly, homozygous motley and AB adult females were purged of mature eggs by natural pair matings. Day 8 or 9 post purging, immature stage IV oocytes were isolated and injected with ∼200 pg of birc5b::eGFP or birc5a::eGFP mRNA. GFP-expressing mature eggs were manually defolliculated and fixed for meiosis rescue or in vitro fertilized and imaged or fixed for post-fertilization rescue analysis. For zygotic rescues ∼200 pg of birc5b::eGFP mRNA was injected into 1-cell embryos and GFP-expressing embryos were imaged and fixed between 2–4hpf.
∼100–500 wild-type or motley/birc5b embryos or eggs were collected and lysed in RIPA buffer on ice using a small syringe after discarding all the embryo medium. Lysates were centrifuged at 13000 rpm for 3–4 mins at 4°C to settle debris and protein concentration was determined. 50–200 µg of protein was loaded onto precast 4–15% TGX gels (BioRad) and blotted onto PVDF membranes for 10 hours at 4°C. Membranes were blotted with 1∶100 anti-Survivin (sc-17779, Santa Cruz Biotechnology Inc) or 1∶500 anti-Survivin-BIR (Unconjugated, Cell Signaling Technology 2808). Membranes were developed using the Fast Western Blot Kit SuperSignal (ThermoScientific). Anti-Survivin-BIR antibodies were developed against conserved BIR aminoacid sequence centered on human Cys60 (Cell Signaling Technology), whose corresponding aminoacid in zebrafish Birc5b is located 15 aminoacids upstream of the site of the mutation in the Motley/Birc5b product.
Wild-type and motley/birc5b mutant embryos were obtained by in vitro fertilization to synchronize the clutches for all experiments. Embryos were fixed with a paraformaldehyde-glutaraldehyde fixative and immunolabeled as described previously [24]. Primary antibodies used were Mouse anti-α-Tubulin (1∶2500, Sigma T5168), Rabbit anti-β-catenin (1∶1000, Sigma C2206), Rabbit anti-phospho-Myosin Light Chain 2 (1∶50, Cell Signaling Technology 3671L), Rabbit anti-actin (1∶100, Sigma A2066), Rabbit anti-AurB (1∶100, [21]) and Rabbit anti-Survivin Alexa 488 (herein anti-Survivin-BIR, same as used for western analysis but is conjugated; 1∶25, Cell Signaling Technology 2810). Fluorescent secondary antibodies Goat anti-Mouse-Cy5 (Jackson ImmunoResearch Labs (JIL) 115-175-003), Goat anti-Rabbit-Alexa 488 (Molecular Probes A-11008), Goat anti-Rabbit-Cy3 (JIL 111-165-144) and Goat anti-Mouse-Cy3 (JIL 115-165-062) were used at 1∶100. For triple immunolabeling, anti-Survivin-BIR was added after the secondary antibody wash and incubated overnight at 4°C prior to DAPI staining. All immunolabeled embryos were semi-flat mounted for animal views of the blastodisc. Images were obtained using a Zeiss LSM510 confocal microscope and analyzed using ImageJ.
Embryos were fixed in 4% paraformaldehyde for 12 hrs at room temperature, dechorionated and transferred into 100% Methanol at −20°C overnight. Rehydrated embryos were hybridized with antisense birc5b overnight at 65°C. Whole mount in situs were developed using an anti-DIG alkaline phosphatase antibody, followed by NBT-BCIP color reaction. Fluorescent in situs were incubated with an anti-DIG-POD antibody (Roche Applied Science) and developed using the Tyramide signal amplification kit (Invitrogen). For immunolabeling after fluorescent in situ hybridization, embryos were washed in PBS-Triton after the Tyramide reaction, deyolked and blocked in antibody block prior to addition of the Rabbit-anti-phospho Myosin Light Chain 2.
Embryos immunolabeled at ∼40mpf were divided into 4 quadrants and a Region of Interest (ROI) away from the furrow was imaged in each quadrant at 40×. The location of 40× ROIs was chosen such that it encompassed the RNP aggregation wave front, where multimerization would be most evident. Within each 40× ROI, 4 random ROIs were imaged using a 100× objective and a 3× digital zoom (300× ROIs, 16 per embryo). RNPs from 5 embryos each of WT, motley/birc5b and nocodazole treated were pooled for the pie charts. The number of GP RNPs that were directly adjoined was determined in the 300× ROIs using the Cell Counter plugin from ImageJ. This number was used as a semi-quantitative measure of GP RNP multimerization.
|
10.1371/journal.pgen.1006499 | Cell Cycle Constraints and Environmental Control of Local DNA Hypomethylation in α-Proteobacteria | Heritable DNA methylation imprints are ubiquitous and underlie genetic variability from bacteria to humans. In microbial genomes, DNA methylation has been implicated in gene transcription, DNA replication and repair, nucleoid segregation, transposition and virulence of pathogenic strains. Despite the importance of local (hypo)methylation at specific loci, how and when these patterns are established during the cell cycle remains poorly characterized. Taking advantage of the small genomes and the synchronizability of α-proteobacteria, we discovered that conserved determinants of the cell cycle transcriptional circuitry establish specific hypomethylation patterns in the cell cycle model system Caulobacter crescentus. We used genome-wide methyl-N6-adenine (m6A-) analyses by restriction-enzyme-cleavage sequencing (REC-Seq) and single-molecule real-time (SMRT) sequencing to show that MucR, a transcriptional regulator that represses virulence and cell cycle genes in S-phase but no longer in G1-phase, occludes 5’-GANTC-3’ sequence motifs that are methylated by the DNA adenine methyltransferase CcrM. Constitutive expression of CcrM or heterologous methylases in at least two different α-proteobacteria homogenizes m6A patterns even when MucR is present and affects promoter activity. Environmental stress (phosphate limitation) can override and reconfigure local hypomethylation patterns imposed by the cell cycle circuitry that dictate when and where local hypomethylation is instated.
| DNA methylation is the post-replicative addition of a methyl group to a base by a methyltransferase that recognise a specific sequence, and represents an epigenetic regulatory mechanism in both eukaryotes and prokaryotes. In microbial genomes, DNA methylation has been implicated in gene transcription, DNA replication and repair, nucleoid segregation, transposition and virulence of pathogenic strains. CcrM is a conserved, cell cycle regulated adenine methyltransferase that methylates GANTC sites in α-proteobacteria. N6-methyl-adenine (m6A) patterns generated by CcrM can change the affinity of a given DNA-binding protein for its target sequence, and therefore affect gene expression. Here, we combine restriction enzyme cleavage-deep sequencing (REC-Seq) with SMRT sequencing to identify hypomethylated 5’-GANTC-3’ (GANTCs) in α-proteobacterial genomes instated by conserved cell cycle factors. By comparing SMRT and REC-Seq data with chromatin immunoprecipitation-deep sequencing data (ChIP-Seq) we show that a conserved transcriptional regulator, MucR, induces local hypomethylation patterns by occluding GANTCs from the CcrM methylase and we provide evidence that this competition occurs during S-phase, but not in G1-phase cells. Furthermore, we find that environmental signals (such as phosphate depletion) are superimposed to the cell cycle control mechanism and can override the specific hypomethylation pattern imposed by the cell cycle transcriptional circuitry.
| DNA methylation is a conserved epigenetic modification that occurs from bacteria to humans and is implicated in control of transcription, DNA replication/repair, innate immunity and pathogenesis [1, 2]. Originally described as a mechanism that protects bacteria from invading foreign (viral) DNA [3], methyl-N6-adenine (m6A) modifications are thought to direct infrequent and stochastic phenotypic heterogeneity in bacterial cells [4, 5] and were recently implicated in transcriptional control of lower eukaryotic genomes and silencing in mouse embryonic stem cells [6–8].
How local changes in methylation are instated during the cell cycle remains poorly explored, even in γ-proteobacteria such as Escherichia coli and Salmonella enterica, as cell cycle studies on cell populations are cumbersome and require genetic manipulation [9]. Moreover, the replication regulator SeqA that controls the methylation state by preferentially binding hemi-methylated sequences is only encoded in γ-proteobacteria, suggesting that other mechanisms are likely operational in other systems [9, 10]. Model systems in which cell populations can be synchronized without genetic intervention are best suited to illuminate the interplay between methylation and cell cycle [11, 12]. The fresh-water bacterium Caulobacter crescentus and more recently the plant symbiont Sinorhizobium meliloti that reside in distinct environmental niches are such cell cycle model systems [13]. Akin to other α-proteobacteria, C. crescentus and S. meliloti divide asymmetrically into a smaller G1-phase cell and a larger S-phase cell and use conserved transcriptional regulators arranged in modules to coordinate transcription with cell cycle progression [13–16] (Fig 1A). In C. crescentus, MucR1 and MucR2 were recently shown to negatively regulate numerous promoters that are activated by the cell cycle transcriptional regulator A (CtrA) in G1-phase. MucR orthologs control virulence functions in α-proteobacterial pathogens and symbionts, but can also control cell cycle-regulated promoters in C. crescentus [17–20]. MucR1/2 target promoters by way of an ancestral zinc finger-like fold and both proteins are present throughout the C. crescentus cell cycle [17, 21, 22] (Fig 1A). By contrast, the OmpR-like DNA-binding response regulator CtrA is activated by phosphorylation and is only present in G1 and late S-phase cells [23, 24], but not in early S-phase cells (Fig 1A). The promoter controlling expression of the conserved DNA methyltransferase CcrM is among the targets activated by phosphorylated CtrA (CtrA~P) in late S-phase [15, 17, 25–27]. CcrM introduces m6A marks at sites harbouring the recognition sequence 5’-GANTC-3’ (henceforth GANTCs) once passage of the DNA replication fork leaves GANTCs hemi-methylated (Fig 1B). CcrM is an unstable protein degraded by the ATP-dependent protease Lon throughout the cell cycle [28, 29]. Since the ccrM gene is expressed only in late S-phase cells, the time of expression dictates when the unstable CcrM protein is present during the cell cycle. CcrM no longer cycles when it is expressed from a constitutive promoter in otherwise WT cells or when Lon is inactivated [28, 30].
With the advent of SMRT (single-molecule real-time) sequencing it is now possible to obtain m6A-methylome information of bacterial genomes at single base pair resolution [31, 32]. A recent cell cycle methylome analysis of C. crescentus by SMRT-sequencing revealed the large majority of GANTCs switch from hemi-methylated to a full methylated state (m6A-marked GANTCs on both strands) at the onset of CcrM expression [12]. Interestingly, a few sites were consistently hypomethylated, indicating that site-specific mechanisms control local hypomethylation patterns. Local hypomethylation patterns may arise if specific DNA-binding proteins and/or restricted local chromosome topology block access of CcrM to such GANTCs. Here, we combine restriction enzyme cleavage-deep sequencing (REC-Seq) with SMRT sequencing to unearth hypomethylated GANTCs in the genomes of wild type (WT) and mutant C. crescentus and S. meliloti. We show that the conserved transcriptional regulator MucR induces local m6A-hypomethylation by preventing CcrM from accessing GANTCs during S-phase, but only when CcrM cycles. Since repression of MucR target promoters is normally overcome in G1-phase, our data suggest that MucR is unable to shield GANTCs when CcrM is artificially present in G1 cells. Lastly, we discovered that phosphate starvation promotes methylation of specific MucR-shielded GANTCs, revealing an environmental override of the control system that normally instates local hypomethylation patterns during the cell cycle.
Detection of hypomethylated sites by SMRT-sequencing requires sufficient sequencing depth and sophisticated bioinformatic analysis to differentiate unmethylated GANTCs from methylated ones. Since unmethylated GANTCs can be conveniently enriched for in C. crescentus by restriction enzyme cleavage using the HinfI restriction enzyme (which only cleaves unmethylated GANTCs) [33], we sought to apply HinfI-based cleavage followed by Illumina-based deep-sequencing (REC-Seq) to identify hypomethylated GANTCs, similar to a previous procedure used for analysis of hypomethylated m6A sites in the unrelated γ-proteobacterium Vibrio cholerae [34]. We tested REC-Seq on HinfI-treated genomic DNA (gDNA) from C. crescentus and, following bioinformatic filtering, obtained a list of unprotected GANTCs scaling with HinfI cleavage efficiency (“score” in S1 Table). Since nearly all GANTCs suggested to be consistently unmethylated by SMRT sequencing [12] are represented as high scoring GANTCs in the REC-Seq (note the growth conditions or limited SMRT sequencing depth may explain the differences), we concluded that REC-Seq captures hypomethylated GANTCs in scaling manner (see below where selected sites cleaved in WT are no longer cleaved in the ΔmucR1/2 mutant). Since CcrM also methylates GANTCs in other α-proteobacteria [35, 36], we also determined the hypomethylated GANTCs on the multipartite genome of S. meliloti [37] by HinfI REC-Seq and found such hypomethylated sites on the chromosome and both megaplasmids (S1 Table).
To validate the HinfI REC-Seq approach, we conducted REC-Seq (using the methylation-sensitive MboI restriction enzyme) on gDNA from Escherichia coli K12 and V. cholerae, as previously determined either by SMRT sequencing or REC-Seq [10, 34]. The Dam methylase introduces m6A marks at GATCs in many γ-proteobacterial genomes [4] that protect from cleavage by MboI. As known unmethylated sites in these control experiments indeed emerged with high score (S2 Table), we conclude that HinfI REC-Seq is an efficient method to detect and quantitate GANTCs that escape methylation by CcrM.
Having identified hypomethylated GANTCs in the C. crescentus genome by HinfI REC-Seq, we noted that many high scoring GANTCs lie in regions that are occupied by MucR1/2 as determined by previous chromatin-immunoprecipitation deep-sequencing (ChIP-Seq) analysis [17]. Of the hits with a score higher than 100, one third lie in MucR1/2 target sequences, and the proportion is even higher (50%) in the case of the 50 top hits (Table 1 and S1 Table). To test if MucR1/2 occludes these GANTCs from methylation by CcrM, we conducted HinfI-cleavage analysis of gDNA from WT (NA1000) and ΔmucR1/2 double mutant by qPCR (henceforth HinfI-qPCR assay) at six MucR1/2 target sites. The CCNA_00169 promoter (henceforth P169) contains four GANTCs; the CCNA_02901 promoter (P2901), the CCNA_01149 promoter (P1149) and the CCNA_01083 internal sequence contain two GANTCs each; the CCNA_02830 and CCNA_03248 promoters (P2830 and P3248) carry one GANTC each (Fig 2A). A high percentage (100%) of methylation in the HinfI-qPCR assay indicates that HinfI cannot cleave this site because of prior methylation by CcrM, whereas a low percentage reflects efficient cleavage of the non-methylated DNA by HinfI. In WT gDNA these six MucR1/2-target sequences are almost completely cleaved by HinfI, indicating that the GANTCs are hypomethylated in the presence of MucR1/2. However, these sites are methylated and therefore not cleaved by HinfI in ΔmucR1/2 cells (Fig 2B). As control for the specificity of the HinfI-qPCR assay we conducted the same analysis on sequences that are not MucR1/2 targets harbouring either i) a hypomethylated GANTC (PnagA), ii) several methylated GANTCs (PpodJ) or iii) a control sequence that does not contain GANTCs (PxylX). These controls revealed a level of amplification in the HinfI-qPCR assay as predicted (S1A Fig) and showed no difference between WT and ΔmucR1/2 cells. Thus, only hypomethylated sequences that are bound by MucR1/2 in vivo are converted to methylated GANTCs in the absence of MucR1/2.
SMRT sequencing of WT and ΔmucR1/2 gDNA supported the result that these GANTCs carry m6A marks as inferred by a high characteristic interpulse-duration (IPD) ratio observed in ΔmucR1/2 versus WT cells (S3 Table). Interestingly, this analysis also revealed eleven GANTCs with the inverse behaviour, i.e. a low IPD ratio in ΔmucR1/2 versus WT cells, suggesting that they no longer carry m6A marks in the absence of MucR1/2. To confirm this result we conducted HinfI-qPCR assays at two of these GANTCs: the CCNA_01248 promoter (P1248) and the CCNA_03426 promoter (P3426). As predicted by the methylome analysis, we observed that the methylation percentage of these GANTCs was reduced in ΔmucR1/2 versus WT (S1B Fig). On the basis of these experiments, we conclude that MucR1/2 prevents m6A-methylation by CcrM at several MucR1/2-target sequences, but can also facilitate methylation at other sites. This would likely occur by an indirect mechanism involving other MucR-dependent DNA-binding proteins that compete with CcrM at certain GANTCs.
To obtain a global picture of hypomethylated GANTCs in the absence of MucR1/2, we conducted REC-Seq analysis on gDNA extracted from the ΔmucR1/2 strain (Table 1 and S1 Table). Comparison of the REC-Seq data for WT and ΔmucR1/2 cells (S2 Fig) supported the conclusion that binding of MucR1/2 prevents methylation by CcrM, as the GANTCs tested by HinfI-qPCR (shown in Fig 2) have a high REC-Seq score in WT and a low REC-Seq score (or they are not detected) in the ΔmucR1/2 strain. Moreover, most of the GANTCs that show a strong decrease in score between WT and ΔmucR1/2 cells are also lying in regions directly bound by MucR1/2 (Table 1 and S1 Table), based on ChIP-Exo (S4 Table) and published ChIP-Seq data [17].
Since CcrM is restricted to late S-phase and MucR1/2-repression is overcome in G1-phase [17, 28], we tested if MucR1/2-bound GANTCs are still hypomethylated when CcrM no longer cycles. To this end we used two strains: the Δlon::Ω (henceforth lon) mutant, as the Lon protease is responsible for degradation of CcrM throughout the cell cycle and upon inactivation of Lon the CcrM protein accumulates also in G1-cells, although it is only synthesized in S-phase [28, 29], and a strain with a second copy of the ccrM gene under control of the constitutive Plac promoter (integrated at the ccrM locus, ccrM::Plac-ccrM) [30, 33]. Indeed, HinfI-qPCR analysis revealed that the fraction of methylated P169, P1149 and P2901 GANTCs increases in lon and ccrM::Plac-ccrM strain relative to WT cells (Fig 3A).
To exclude that constitutive presence of CcrM simply prevents MucR binding to DNA because CcrM outnumbers and therefore outcompetes MucR, we conducted several control experiments to demonstrate the specificity of the methylation control at these GANTCs. First, immunoblotting experiments revealed that MucR1/2 levels were maintained in the lon and ccrM::Plac-ccrM strains compared to WT (Fig 3C). Second, overexpression of either WT MucR1 or of an N-terminally extended (dominant-negative) MucR1 variant from Pvan on a high copy plasmid (pMT335) [17, 38] did not prevent methylation of P169, P1149 and P2901 GANTCs in lon mutant cells (Fig 3A) or alter CcrM steady-state levels (S1C Fig). Conversely, constitutive expression of CcrM from the same vector (pMT335) in WT cells recapitulated the effect on methylation of the P169, P1149 and P2901 GANTCs (Fig 3B). Similarly, methylases of Thermoplasma acidophilum (TA), Helicobacter pylori (HP) or Haemophilus influenzae (Hinf), which also specifically methylate GANTCs but are not related to α-proteobacterial CcrM, also lead to methylation of these hypomethylated GANTCs when expressed from pMT335 (Fig 3B). By contrast, the methylation state of GANTCs at the parS locus was not significantly altered by the expression of the methylases or by the lon mutation (S1D Fig). On the basis that CcrM and unrelated methylases are able to compete against MucR1/2 for methylation of P169, P1149 and P2901 GANTCs when expressed constitutively, we hypothesize that MucR1/2 no longer efficiently compete with CcrM in G1-phase when both proteins are present at this time (Fig 3A and 3B).
To test if MucR1 binds to its targets in G1-phase, we conducted chromatin-immunoprecipitation-followed by deep-sequencing of exonuclease treated fragments (ChIP-Exo), a technique with enhanced resolution compared to conventional ChIP-Seq [39]. We treated with the anti-MucR1 antibody chromatin prepared from synchronized cells at four different time points after synchronization [10 min (T10, G1 phase), 40 min (T40, G1-to-S transition), 70 min (T70, early S-phase), 100 min (T100, late S-phase) (Fig 1B)] and used a bioinformatic algorithm to define the binding sites at super-resolution (see Methods and [40]). Surprisingly, the binding profiles at the four time points appeared to be nearly congruent (Fig 3D) and quantification of the enrichment ratio failed to reveal major changes of MucR1 binding to its targets during the cell cycle (Fig 3E and S4 Table). On the other hand, conformational changes or altered dynamics of binding (i.e. dissociation constants, on- and off-rates) that are undetectable by our methods might allow transcription from the MucR-bound promoters in G1-phase. Transient release of DNA by MucR1/2 or changes in chromatin conformation could provide access to competing DNA binding proteins such as CcrM, RNA polymerase (RNAP) and other transcription factors (like CtrA) in G1-phase to induce methylation or firing of the MucR1/2 target promoters.
As MucR1/2 regulates the methylation state of the aforementioned GANTCs, we wondered if the MucR1/2-targets P169, P1149 and P2901 display promoter activity in a MucR1/2-dependent and/or methylation-dependent manner. To this end, we conducted LacZ (β-galactosidase)-based promoter probe assays of P169-, P1149- and P2901-lacZ transcriptional reporters (driving expression of a promoterless lacZ gene) in WT and ΔmucR1/2 cells and observed that LacZ activity of all reporters was elevated in ΔmucR1/2 cells versus WT (Fig 4A–4C). The increase was less dramatic for P169-lacZ (156 ± 5.8% relative to WT) than for P1149- and P2901-lacZ (439 ± 7.4 and 385 ± 40%, respectively). We then asked if promoter activity is augmented when cycling of CcrM is prevented. Indeed, the P169-, P1149- and P2901-lacZ reporters indicated an increase in promoter activity in the lon mutant and Plac-ccrM strains compared to WT (Fig 4D). Importantly, no increase in LacZ activity was observed in lon and Plac-ccrM strains with other promoters (PhvyA and PpilA, Fig 4D) that are bound by MucR1/2 and whose activity is increased in ΔmucR1/2 cells [17, 41] but contain no hypomethylated GANTCs. We further corroborated these results by showing that constitutive expression of C. crescentus CcrM or the T. acidophilum GANTC-methylase from Pvan on pMT335 led to an increase in P169-, P1149- and P2901-lacZ promoter activity (Fig 4E). Consistent with the fact that in ΔmucR1/2 cells these promoters are no longer hypomethylated, constitutive expression of CcrM from Plac-ccrM in ΔmucR1/2 cells had no significant effect on P169-, P1149- and P2901-lacZ promoter activity (Fig 4E).
To determine if changing the GANTC methylation state (by mutation to GTNTC) in P169- and P2901-lacZ (5 sites mutated for P169, P169*; 2 sites for P2901, P2901*) also affects promoter activity, we measured LacZ activity of the mutant promoters in WT and ΔmucR1/2 cells and found that they still exhibited MucR1/2-dependency, as the P169*- and P2901*-lacZ were still strongly de-repressed in the absence of MucR1/2 (Fig 4A and 4B). We also observed an increase (136% ± 6%) in activity of P169*-lacZ relative to P169-lacZ in WT, while the activity of P2901*-lacZ was decreased compared to P2901-lacZ. The mutations may alter the target sequence for other regulator(s) in addition to the methylation properties, thereby affecting transcription directly or indirectly in a positive or negative fashion [42, 43]. For example, P2901 is bound by the master cell cycle regulator CtrA in vivo and the ΔmucR1/2 mutation is known to affect CtrA expression [17], whereas the P169 promoter is affected by the phosphate starvation response (see below) [44, 45].
LacZ-based assays are a general and indirect measurement of promoter activity, but they do not pinpoint the transcription start sites (TSSs), thus cannot reveal the physical proximity of the TSS relative to the hypomethylated GANTCs. To correlate transcriptional regulation of MucR1/2 and hypomethylated GANTCs, we took advantage of the recently developed RNA-Seq-based strategy for exact mapping of transcriptome ends (EMOTE) [46] that can also be used to map the (unprocessed) 5’ends of nascent transcripts that harbour triphosphate 5’end (5’-ppp). To this end, total RNA is first treated with XRN1 (to remove transcripts with monophosphorylated 5’ends) and then 5’-ppp transcripts are treated with E. coli RppH, which converts the 5’ends to a monophosphorylated form that can be ligated to a bar-tagged RNA oligo using T4 RNA ligase [46] (S3 Fig). Comparative TSS-EMOTE analysis in total RNA extracted from WT and ΔmucR1/2 cells unearthed several TSSs that are activated when MucR1/2 is absent (arrows in Fig 2A, S5 Table). Importantly, several of these TSSs were detected in close proximity to the GANTCs within MucR1/2 target sequences, including P2901, P2830, P3248 and CCNA_01083. These results, therefore, validate the physical proximity and functional interplay between MucR1/2 and hypomethylated GANTCs. While for weak MucR1/2 target promoters the sequencing depth may have limited their detection by TSS-EMOTE, this analysis unexpectedly revealed several MucR-dependent antisense transcripts with potential regulatory roles (green arrows in Fig 2A). We validated the MucR1/2-dependency of two such antisense promoters (P2902_AS and P3247_AS) by LacZ-promoter probe assays and detected a substantial increase in activity of P2902_AS-lacZ and P3247_AS-lacZ in ΔmucR1/2 versus WT cells (S1E Fig), indicating that these promoters (and the GANTCs within) are clearly MucR1/2 regulated in C. crescentus.
Knowing that MucR is functionally interchangeable in α-proteobacteria [17, 18] and that hypomethylated GANTCs are also detected in the S. meliloti multipartite genome by HinfI REC-Seq (see above and S1 Table), we tested whether S. meliloti MucR also occludes GANTCs from methylation by CcrM in target promoters. We compared the methylation of WT and mucR::Tn S. meliloti gDNA by HinfI REC-Seq and SMRT-sequencing (S1 and S3 Tables). Guided by these data sets, we validated hypomethylation of GANTCs at or near the SMa1635 (SM2011_RS04470) and SMa2245 (SM2011_RS06125) genes by HinfI-restriction/qPCR analysis. We chose these GANTCs, located on the symbiotic megaplasmid pSymA, to take advantage of the S. meliloti multipartite genome and to explore if MucR-control of hypomethylation also applies to episomal elements such as a symbiotic megaplasmid. HinfI-restriction/qPCR analysis revealed that these GANTCs are largely hypomethylated in WT compared to mucR::Tn cells (Fig 5A and 5B). To confirm that these GANTCs are indeed direct targets of S. meliloti MucR, we conducted quantitative ChIP (qChIP) experiments (Fig 5D) with chromatin from S. meliloti WT and mucR::Tn cells precipitated using antibodies to C. crescentus MucR2 that recognize S. meliloti MucR on immunoblots (S4A Fig). The qChIP experiments revealed that S. meliloti MucR indeed binds at or near the hypomethylated SMa1635 and SMa2245 GANTCs of WT cells (Fig 5D), but not at a control site (SMc01552). Moreover, since CcrM is restricted to late S-phase also in S. meliloti [14], we tested whether constitutive expression of ccrMCc in S. meliloti WT cells affected the methylation of GANTCs at SMa1635 and SMa2245. Ectopic expression of ccrMCc from Plac on pSRK vector [47] significantly increased the methylation of SMa1635 and SMa2245, showing that S. meliloti MucR no longer occludes GANTCs in target promoters when cycling of CcrM is impaired (Fig 5C). Consistent with SMa1635 and SMa2245 being MucR targets, LacZ-based promoter probe experiments (using Pa1635-lacZ and Pa2245-lacZ) revealed that they are de-repressed in S. meliloti mucR::Tn cells compared to WT (Fig 5E) and that S. meliloti MucR represses Pa1635-lacZ and Pa2245-lacZ in C. crescentus WT or ΔmucR1/2 cells (Fig 5F). Importantly, when cycling of CcrM in C. crescentus was prevented by Plac-ccrM or the lon mutation Pa1635-lacZ and Pa2245-lacZ activity was increased compared to the WT strain (Fig 5G). Thus, MucR controls hypomethylation during α-proteobacterial cell cycle.
To determine if other systemic (cell cycle) signals can alter methylation patterns in α-proteobacteria, we tested if CtrA can also occlude GANTC sites from methylation by CcrM. First, we constructed a synthetic promoter in which three GANTCs overlapping two CtrA-boxes (one GANTC in each CtrA box and one in between) were placed downstream of an attenuated E. coli phage T5 promoter on the lacZ promoter probe plasmid (Fig 6A). Next, we determined the methylation percentage of the GANTCs in WT C. crescentus cells harbouring the resulting reporter plasmid by HinfI-qPCR analysis and found that the GANTCs are only partially methylated in WT cells, but efficiently methylated when CcrM is expressed ectopically (Fig 6A). Thus, methylation patterns can also emerge from competition between CcrM and other cell cycle regulators such as CtrA at appropriately positioned GANTCs.
To explore if environmental signals can also affect local hypomethylation patterns, we took advantage of the fact that expression of CCNA_00169 (also known as elpS) is induced upon phosphate starvation of C. crescentus cells [44, 45]. Accordingly, we compared the P169 methylation patterns by HinfI-qPCR analysis of gDNA from WT cells grown in standard medium (PYE) and phosphate-limiting conditions. This revealed a significant increase in P169 GANTC methylation in phosphate-limiting conditions compared to PYE (Fig 6B) and we observed a commensurate induction of P169-lacZ and P169*-lacZ that was MucR1/2 independent (Fig 6C). Both the increase in P169 GANTC methylation and P169-lacZ activity are dependent on the phosphate-responsive transcriptional regulator PhoB (S4B and S4C Fig), suggesting that PhoB can facilitate methylation of MucR-protected GANTCs at P169. The result that no significant increase of the P1149 methylation or P1149-lacZ activity was seen when WT cells were grown in phosphate-limiting conditions compared to standard PYE medium (or in ΔphoB::Ω cells compared to WT) argues against the possibility that changes in CcrM expression or activity underlie the modified methylation pattern of P169 (Fig 6B and 6D; S4B and S4C Fig). Thus, P169 provides an example of an environmental override for a promoter subject to local hypomethylation control by the cell cycle transcriptional circuitry.
The correlative changes between human genetic variability and local (hypo)methylation prompt the question if and how such patterns are regulated by the cell cycle and/or environmental cues. Taking advantage of bacterial genomes that are small enough for full-methylome analysis by cutting-edge REC-Seq and SMRT-sequencing, we show that local m6A-hypomethylation exists in two different α-proteobacterial lineages and that conserved cell cycle factors govern its establishment in both systems. While in γ-proteobacterial lineages transcriptional regulators are also known to compete with the Dam m6A methylase to occlude certain methylation sites, local hypomethylation patterning has not been explored in the context of the transcriptional circuitry controlling progression of the (α-proteo)bacterial cell cycle. In eukaryotes methylation heterogeneity involves 5-methyl-cytosines introduced at CpG dinucleotides [2], but recently m6A marks, instated by unknown mechanisms, have also been reported [6–8]. Reliable detection of methylation sites by SMRT-sequencing requires extensive (25-fold) coverage for adenine methylation and even higher coverage for cytosine methylation (250-fold coverage needed in some instances) [5]. Non-methylated sites are only reliably detected by elimination of sites on which methylation is detected, thus leaving an element of uncertainty for those sites classified as non-methylated based on the absence of the kinetic signature for methylation. By contrast, REC-Seq with a methylation sensitive restriction enzyme was used here to enrich for non-methylated sites in α-proteobacteria by HinfI cleavage. The continuum of scores we detected in these experiments points towards the use of REC-Seq in detecting loci whose methylation is variable within a culture, for example phase variable loci [48, 49]. HinfI REC-Seq revealed the occurrence of non-methylated GANTCs in at least two α-proteobacterial genomes. Subsequent genetic analyses showed that the determinants controlling hypomethylation are conserved in these bacteria, but they are not encoded in eukaryotic genomes. However, at least one component, MucR, possesses an ancestral zinc-finger-type DNA binding domain [22], a protein domain which is also wide-spread in developmental regulation of eukaryotes [50]. The fact that MucR regulates expression of virulence and cell cycle genes [17–20], has recently been shown to confine genetic exchange by generalized transduction to G1-phase in C. crescentus via transcriptional regulation [41] and is responsible for hypomethylation of specific loci on the chromosome or megaplasmids thus raises the possibility that zinc-finger proteins may control (epigenetic) DNA transactions including local hypomethylation during the eukaryotic cell cycle as well. In bacteria local methylation changes may correlate with altered virulence behaviour and may underlie cell cycle-control in pathogens, endosymbionts or other microbial systems. Methylation is known to influence virulence functions in γ-proteobacteria, often by imposing bistability from phase-variable virulence promoters in subpopulations via transcriptional regulators such as Lrp, Fur or OxyR [5, 9, 42, 48, 51–55]. Phenotypic heterogeneity in antibiotic drug tolerance in vivo (a state known as persistence), which is acquired in a low fraction of bacterial cells, may also underlie epigenetic changes induced stochastically by methylation, either deterministically (during the cell cycle) or environmentally. Although no phase-variable promoters are currently known for the α-proteobacteria, these bacteria offer the possibility to investigate the relationship of local hypomethylation with cell cycle control, as both C. crescentus and S. meliloti are synchronizable and exhibit comparable cell cycle control systems and transcription [13–15, 56]. However, as binding of MucR to DNA is not impaired by methylation, the mechanisms underlying the increase in transcription of target genes induced by methylation in α-proteobacteria (Fig 4D–4F; Fig 5G) are likely to be different from those described for γ-proteobacteria. Moreover, the α-proteobacteria lineage includes the Rickettsiales order encompassing obligate intracellular pathogens, endosymbionts and the extinct proto-mitochondrion from which the modern day mitochondria descended [57]. As MucR and CcrM orthologs are not encoded in most Rickettsiales genomes, the determinants of hypomethylation in this order must be different, if they do exist. Interestingly, endosymbionts from the genus Wolbachia might provide a possible exception. Their genomes encode an unusual putative DNA methylase in which a C-terminal pfam01555 methylase-domain is fused to a pfam02195 ParB-like nuclease domain found in DNA-binding proteins and plasmid replication factors [58]. The sheltered niche of obligate intracellular Rickettsia contrasts with that of free-living relatives that are exposed to major environmental fluctuations.
In summary, our work shows that environmental regulatory responses like that to phosphate limitation, which is particularly pertinent for bacteria living in aquatic ecosystems as C. crescentus, are superimposed on (direct or indirect) hypo- or hyper-methylation control cued by the cell cycle. As many hypomethylated sites occur upstream of genes encoding transcription factors (see S1 Table) and transcription factors are often auto-regulatory, it is conceivable that local hypomethylation is often induced by cis-encoded site-specific DNA-binding proteins that can compete with DNA methylases for overlapping target sites. The mechanism of DNA binding and temporal regulation of MucR remain to be elucidated in detail in order to reveal why MucR shields certain target sites from methylation by CcrM. Our work on MucR-dependent hypomethylation by HinfI REC-Seq along with the comprehensive analysis of hypomethylated sites in other α-proteobacterial genomes [10] indicates that the functions controlled by hypomethylated promoters are distinct and generally not conserved among different α-proteobacteria. This suggests that hypomethylation does not play a major role in the regulation of the α-proteobacterial cell cycle, even though conserved cell cycle transcriptional regulators govern hypomethylation patterns. If it is largely serendipitous which sites MucR shields from methylation, it seems plausible that such hypomethylation control systems mediate species-specific transcriptional adaptations in response to stresses via MucR, CcrM or other variables that influence their binding, either directly or indirectly. For example, cell cycle controlled changes in local chromosomal topologies mediated by DNA replication or nucleoid-associated factors [59, 60] could exclude DNA methylases from specific target sites.
Caulobacter crescentus NA1000 [61] and derivatives were grown at 30°C in PYE (peptone-yeast extract) or M2G (minimal glucose). For phosphate starvation, Caulobacter cells were grown in 1/5X PYE (5-fold diluted PYE except 1 mM MgSO4 and 1 mM CaCl2, supplemented with 0.2% glucose). Sinorhizobium meliloti Rm2011 and derivatives were grown at 30°C in Luria broth (LB) supplemented with CaCl2 2.5 mM and MgSO4 2.5 mM. Escherichia coli S17-1 λpir and EC100D were grown at 37°C in LB. Swarmer cell isolation, electroporations, bi-parental matings and bacteriophage φCr30-mediated generalized transductions were performed as previously described [62–65]. Nalidixic acid, kanamycin, gentamicin and tetracycline were used at 20 (8 for S. meliloti), 20, 1 (10 for E. coli and S. meliloti) and 1 (10 for E. coli and S. meliloti) μg/mL, respectively. Plasmids for β-galactosidase assays were introduced into S. meliloti by bi-parental mating and into C. crescentus by electroporation. Strains and plasmids constructions are detailed in the S1 Text file.
Genomic DNA was extracted from mid-log phase cells (10 ml). Aliquots of DNA (0.5–1 μg) were digested with HinfI restriction endonuclease and used to determine the methylation percentage by Real-Time PCR. Real-time PCR was performed using a Step-One Real-Time PCR system (Applied Biosystems, Foster City, CA) using 0.05% of each DNA sample (5 μl of a dilution 1:100) digested with HinfI, 12.5 μl of SYBR green PCR master mix (Quanta Biosciences, Gaithersburg, MD) and primers 10 μM each, in a total volume of 25 μl. A standard curve generated from the cycle threshold (Ct) value of the serially diluted non-digested genomic DNA was used to calculate the methylation percentage value for each sample. Average values are from triplicate measurements done per culture, and the final data was generated from three independent cultures per strain and condition. The primers used for Real-Time PCR are listed in Table B in the S1 Text file.
SMRT (single-molecule real-time) sequencing libraries were prepared from gDNA extracted from the four samples (C. crescentus and S. meliloti WT and mucR mutant strains) using the DNA Template Prep Kit 2.0 (250bp–3Kb, Pacific Biosciences p/n 001-540-726). Sequences generated by the Pacific Bioscience RSII were aligned to the C. crescentus NA1000 or S. meliloti Rm2011 genomes [37, 66, 67] using Blasr (https://github.com/PacificBiosciences/blasr) and the modification and associated motifs patterns were identified applying the RS_Modification_and_Motif_Analyisis protocol in SMRT Analysis (https://github.com/PacificBiosciences/SMRT-Analysis/wiki/SMRT-Analysis-Software-Installation-v2.2.0). For each aligned base, a statistics measured as interpulse duration (IPD) combined with a modification quality value (QV) would mark the methylation status. On the one hand, a minimum QV of 45 is required for a position to be marked as methylated; on the other hand, a maximum QV between 10 and 30 (depending on the observed kinetic detections background in the sample), coupled with the requirement that such a score is observed on both strands, would mean that a position, in an otherwise methylated motif, is unmethylated.
For REC-Seq (restriction enzyme cleavage–sequencing) 1 μg of genomic DNA from C. crescentus NA1000 and S. meliloti Rm2011 was cleaved with HinfI, a blocked (5’biotinylated) specific adaptor was ligated to the ends and the ligated fragments were then sheared to an average size of 150–400 bp (Fasteris SA, Geneva, CH). Illumina adaptors were then ligated to the sheared ends followed by deep-sequencing using a Hi-Seq Illumina sequencer, and the (50 bp single end) reads were quality controlled with FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). To remove contaminating sequences, the reads were split according to the HinfI consensus motif (5’-G^ANTC-3’) considered as a barcode sequence using fastx_toolkit (http://hannonlab.cshl.edu/fastx_toolkit/) (fastx_barcode_splitter.pl —bcfile barcodelist.txt —bol —exact). Most of the reads (more than 90%) were rejected, and the reads kept were remapped to the reference genomes [37, 66, 67] with bwa mem [68] and samtools [69] to generate a sorted bam file. The bam file was further filtered to remove low mapping quality reads (keeping AS > = 45) and split by orientation (alignmentFlag 0 or 16) with bamtools [70]. The reads were counted at 5' positions using Bedtools [71] (bedtools genomecov -d -5). Both orientation count files were combined into a bed file at each identified 5’-GANTC-3’ motif (where reverse counts > = 1 at position N+1 and forward counts > = 1 at position N-1) using a home-made PERL script. The HinfI positions in the bed file were associated with the closest gene using Bedtools closest [71] and the gff3 file of the reference genomes [72]. The final bed file was converted to an MS Excel sheet (S1 and S2 Tables) with a homemade script. For the MboI-based REC-Seq, the strategy was identical except that a different adaptor was used for ligation after cleavage and the MboI consensus motif (5’-^GATC-3’) was used as barcode for filtering of V. cholerae O1 biovar El Tor [73] and E. coli K12 Ec100D gDNA mapped onto the MG1655 genome [74].
C. crescentus WT cells for ChIP-Exo were taken at different time points after synchronization (10, 40, 70 and 100 minutes). After cross-linking, chromatin was prepared as previously described [17]. ChIP-Exo was performed with 2 μl of polyclonal antibodies to MucR1 at Peconic LCC (http://www.peconicgenomics.com) (State College, PA), which provided standard genomic position format files (BAM) as output using the SOLiD genome sequencer (Applied Biosystems). A custom Perl script was then used to calculate the sequencing (read) coverage per base (per-base coverage) for each ChIP-Exo sample. Next, we computed the enrichment ratio (ER) for each promoter region. To this end, the Perl script extracted the per-base coverage of a 600 bp region for each ORF (from -500 to +100 from the start codon for each ORF annotated in C. crescentus genome) and calculated the average coverage for each of these regions. The resulting value was then normalized with respect to the coverage of all the intergenic regions. This was done (by the Perl script) by selecting all the intergenic regions in the C. crescentus genome, merging them and extracting the per-base coverage for all these intergenic regions. The coverage was averaged for windows of 600 bp, shifting each window by 100 bp, and the mean of all resulting values was computed. The ER for each promoter region was therefore calculated as the ratio between the average coverage of the promoter region divided by the mean obtained for the intergenic regions.
The transcription start sites in the NA1000 WT and the ΔmucR1/2 mutant were determined by TSS-EMOTE (Transcription Start Specific Exact Mapping Of Transcriptome Ends), a global assay that reveals the sequence of the 20 first nucleotides of 5’-triphosphorylated RNA in a sample based on an XRN-1 digest of transcripts lacking the 5’ triphosphate ends [46]. The TSS-EMOTE protocol and analyses were performed according to the scheme in S3 Fig and the detailed protocol described in [75]. We used a Worst-Case (i.e.) smallest difference) model to compare the number of Unique Molecular Identifiers between the two pairs of biological replicates (i.e. mutant vs. wild-type) and provide additional information about relative expression for each of the detected TSSs. The full list of detected TSSs is shown in S5 Table and TSSs at the relevant genomic loci are indicated by black (sense) and green (antisense) arrows in Fig 2A.
β-galactosidase assays were performed at 30°C. Cells (50–200 μl) at OD660nm = 0.1–0.5 were lysed with chloroform and mixed with Z buffer (60 mM Na2HPO4, 40 mM NaH2PO4, 10 mM KCl and 1 mM MgSO4, pH 7) to a final volume of 800 μl. Two hundred μl of ONPG (o-nitrophenyl-β-D-galactopyranoside, stock solution 4 mg/ml in 0.1 M potassium phosphate, pH 7) were added and the reaction timed. When a medium-yellow colour developed, the reaction was stopped by adding 400 μl of 1M Na2CO3. The OD420nm of the supernatant was determined and the Miller units (U) were calculated as follows: U = (OD420nm*1000)/(OD660nm*time [in min] *volume of culture used [in ml]). Error was computed as standard deviation (SD) of at least three independent experiments.
Samples for qChIP assay were prepared from mid-log phase S. meliloti cells as previously described [17]. Two microliters of polyclonal antibodies to MucR2 were used for the immunoprecipitation.
Real-time PCR was performed as described for HinfI-restricted genomic DNA, using 0.5% of each ChIP sample (5 μl of a dilution 1:10). A standard curve generated from the cycle threshold (Ct) value of the serially diluted chromatin input was used to calculate the percentage input value for each sample. Average values are from triplicate measurements done per culture, and the final data was generated from three independent cultures per strain. The primers used for SMa1635 and SMa2245 loci were the same as for the determination of the methylation percentage of these loci (Table B in S1 Text file).
For immunoblots, protein samples were separated on SDS polyacrylamide gel, transferred to polyvinylidene difluoride (PVDF) Immobilon-P membranes (Merck Millipore) and blocked in PBS (phosphate saline buffer) 0.1% Tween20 and 5% dry milk. The anti-sera were used at the following dilutions: anti-CcrM (1:10’000) [26], anti-MucR1 [17] (1:10’000), anti-MucR2 [17] (1:10’000). Protein-primary antibody complexes were visualized using horseradish peroxidase-labelled anti-rabbit antibodies and ECL detection reagents (Merck Millipore).
Plasmids, primers, synthetic fragments and strains constructions are described in the S1 Text file.
Deep-sequencing data are deposited in Gene Expression Omnibus database (GEO: GSE79880).
|
10.1371/journal.pcbi.1004987 | Large Scale Chromosome Folding Is Stable against Local Changes in Chromatin Structure | Characterizing the link between small-scale chromatin structure and large-scale chromosome folding during interphase is a prerequisite for understanding transcription. Yet, this link remains poorly investigated. Here, we introduce a simple biophysical model where interphase chromosomes are described in terms of the folding of chromatin sequences composed of alternating blocks of fibers with different thicknesses and flexibilities, and we use it to study the influence of sequence disorder on chromosome behaviors in space and time. By employing extensive computer simulations, we thus demonstrate that chromosomes undergo noticeable conformational changes only on length-scales smaller than 105 basepairs and time-scales shorter than a few seconds, and we suggest there might exist effective upper bounds to the detection of chromosome reorganization in eukaryotes. We prove the relevance of our framework by modeling recent experimental FISH data on murine chromosomes.
| A key determining factor in many important cellular processes as DNA transcription, for instance, the specific composition of the chromatin fiber sequence has a major influence on chromosome folding during interphase. Yet, how this is achieved in detail remains largely elusive. In this work, we explore this link by means of a novel quantitative computational polymer model for interphase chromosomes where the associated chromatin filaments are composed of mixtures of fibers with heterogeneous physical properties. Our work suggests a scenario where chromosomes undergo only limited reorganization, namely on length-scales below 105 basepairs and time-scales shorter than a few seconds. Our conclusions are supported by recent FISH data on murine chromosomes.
| Understanding how genomes fold within the crowded environment of the nucleus [1] during interphase represents a necessary step for the comprehension of important cellular processes such as gene expression and regulation [2]. The combined results of high-resolution microscopy [3, 4] and mathematical and computer modelling [5, 6] seem to suggest that genomes are organized hierarchically [7, 8]. Each genome is partitioned into a set of single units, the chromosomes, and each chromosome is made of a single filament of DNA complexed around histone octamers to form a necklace-like fiber ≈ 10 nm thick known as the 10nm chromatin fiber. In in vitro conditions close to the physiological ones, this fiber is observed to fold into a thicker, more compact structure known as the 30nm fiber [1], whose role and existence in vivo are nonetheless still quite debated [9, 10]. On larger scales, chromosome conformation capture (3C) techniques [2] have shown that chromosomes appear organized in Topologically Associated Domains (TADs) of sizes ranging from ≈ 0.1 to ≈ 1 megabasepairs (Mbp). Chromosome loci within TADs interact frequently between themselves, but less frequently across different TADs. Finally, chromosomes do not spread inside the whole nucleus, rather they occupy well localized nuclear regions (the so-called “chromosome territories”) which play a crucial role in gene expression and regulation [11].
While much progress on the causal relationship between chromosome structure and function has now been made, many fundamental aspects remain still obscure. One of these crucial issues concerns the link between chromosome (re)organization at various length and time-scales and the spread of inhomogeneties in the sequence of the chromatin fiber which may arise from (1) selective epigenetic marks induced by chemical modifications of the histone tails [12], (2) nucleosomes arrangements in discrete nanodomains of different sizes [4], and (3) selective stimulation of particular kinds of genes [13] or entire gene families [14]. These events will produce modifications in the local polymer properties of the chromatin fiber (as its persistence length or the local compaction ratio) which might affect in turn the whole hierarchical folding of chromosome organization [9, 10].
Motivated by these considerations, we present here the results of Molecular Dynamics computer simulations of a minimal polymer model for interphase chromosomes in order to quantify to what extent the simultaneous presence of chromatin fibers of heterogeneous composition (different thicknesses and flexibilities) is able to generate observable effects on the small- and large-scale structures and motions of the associated chromosome. The proposed model complements previous work by one of the authors [15–18] concerning the description of chromosome folding in terms of fundamental polymer physics. Similarly to other recent works discussing the explicit role of sequence disorder in chromatin [19] and chromosome behaviors [20, 21], here we “deviate” from the description of the chromatin filament as a homopolymer and we discuss sequence effects in space and time through the introduction of controlled amounts of disorder in the chromatin sequence. In this way, we provide a quantitative description for many crucial aspects concerning the structure and dynamics of interphase chromosomes which are “spontaneously” driven by the physical properties of the underlying chromatin sequence with a definite copolymer structure.
In particular, by considering the two “extreme” cases of chromosomes made of: (1) short stretches of a thinner, more flexible fiber randomly interspersed in a “sea” of thicker fiber and (2) chromosomes partitioned into two distinct blocks of thinner and thicker fibers, we show that significative spatial and dynamical rearrangements of chromatin loci appear to be restricted to limited contour lengths (up to ≈ 105 basepairs (bp)) and times scales (up to few seconds). Interestingly, there exists a limited range (104−105 bp) of chromatin contour lengths where chromosome expansion is accompanied by an increase (rather, than a decrease) of the associated contacts between the fibers. We apply this framework to rationalize the outcome of recent experiments which employ fluorescent microscopy to monitor conformational changes of chromosomes that occur upon transcription activation or chromatin decondensing in mouse embryonic stem cells [13]. Finally, we argue that the effects discussed here are not the consequence of the details of the model, but involve more general aspects of polymer physics.
In this section, we present the main results of our model by addressing in particular the specific question of how modifications in the small-scale properties of the chromatin fiber turn to affect chromosome behaviour on much larger scales. In order to model these modifications, we represented the linear chromosome filament as a heterogeneous copolymer made of two types of fibers with physical properties matching those of the 10nm and the 30nm chromatin fibers: 10nm fibers are modelled as completely flexible while 30nm fibers have a persistence length = 150 nm as in previous works [15–18]. Then, we considered several situations: either short (3000 bp) fragments of 10nm fibers are distributed randomly along the whole chromatin filament, or we studied chromosome structures where a single continuous 10nm filament is centered around the chromatin portion closer to or farther from the chromosome center of mass. For more details and technical issues, we invite the reader to look into the “Materials and Methods” section.
With its true existence in vivo appearing more and more debatable [9, 10], employment of 30nm fibers might look inappropriate. However, our choice of physical parameters for the model was finally motivated by their previous use and proven consistency [15–18] with experimental data. We posit though that our results should remain qualitatively valid for more general models of non-homogeneous polymers.
In order to quantify chromosome changes, we have considered either spatial relationships between distal fragments along the chromatin sequence or the dynamic behaviour of chromatin loci. The former aspects were investigated by focusing mainly on the following two observables: (1) the mean-squared internal distance (〈R2(L)〉) and (2) the average contact frequency 〈pc(L)〉 between all possible pairs of genomic loci at any given genomic separation L. For 〈pc(L)〉, we have adopted the choice that two monomers are in contact whenever their spatial distance is smaller than a conventional cut-off distance = 60 nm, corresponding to twice the thickness of the 30nm fiber. These two quantities are of particular experimental interest: in fact, 〈R2(L)〉 can be measured through fluorescence in-situ hybridization (FISH) [11], while 〈pc(L)〉 is the result of chromosome conformation capture (3C) techniques [22, 23]. Moreover, 〈R2(L)〉 and 〈pc(L)〉 are useful to distinguish between complementary aspects of chromosome conformation, as was recently highlighted by Williamson et al. [24]. It is then interesting to monitor our systems by employing both tools. Dynamical aspects were instead discussed in terms of the mean-square displacement 〈δr2(τ)〉 of chromatin loci at lag time τ. This is also a quantity of notable experimental interest, as specific chromatin loci can now be followed in vivo by, e.g., fluorescent microscopy on GFP tagged chromosome sequences [25, 26].
All results are presented as functions of the basepair content of the 10nm fibers present in the model chromosomes vs. the total chromosome basepair content, expressed in percent (see section “Materials and Methods” for details). For comparison, model chromosomes with no (0%) or entirely made of (100%) 10nm fibers are also discussed.
The results of our three case studies are summarized in Fig 1. It is visible that only length-scales smaller than L ≈ 0.1 megabasepairs (Mbp) are affected with 〈R2(L)〉 expanding sensibly more than in the situation where chromosomes are composed only of 30nm fiber (panels A, C, E), while the behaviour at large scales remains unaffected. Moreover, in the case where the unfolded sequences are grouped into a single cluster (middle and bottom panels), there is no dependence on their spatial positioning with respect to the center of mass of the corresponding chromosome territory. Insensitivity of large scales to changes at small ones is also confirmed (panels B, D, F) by the analysis of contact frequencies, 〈pc(L)〉, whose trend remains, in particular, compatible with the experimentally observed power-law 〈pc(L)〉 ∼ L−1 [27]. Interestingly, instead of decreasing as expected from panels A, C and E, corresponding contact frequencies in the limited range [0.01 Mbp − 0.1 Mbp] also increase as a function of the 10nm fiber content (see in particular panel B). In order to understand this result, we use the mathematical relationship [16] between average contact frequencies 〈pc(L)〉 and distribution function pL(R) of internal distances R(L) given by:
〈 p c ( L ) 〉 = ∫ 0 r c p L ( R ) 4 π R 2 d R ∫ 0 ∞ p L ( R ) 4 π R 2 d R , (1)
where rc = 60 nm is the cut-off distance for two monomers to form a contact. In particular, eq 1 implies that pL(R) should also increase as a function of the 10nm fiber content at given L. The inset of panel B confirms this behavior for L = 0.015 Mbp. This specific trend is due to the decreasing of the chromatin “effective persistence length” (from ≈ 104 to ≈ 102 basepairs for model chromosomes entirely made of 30nm and 10nm fibers, respectively) following from increasing amounts of 10nm fibers, which allows genomic loci to contact each other with enhanced probability.
For a more quantitative view on chromosome reorganization at small scales, the same data for 〈R2(L)〉 and 〈pc(L)〉 were normalized to corresponding values for chromosome conformations made of 30nm fiber, see S1 Fig. We highlight in particular the pronounced peaks in the left panels corresponding to a maximum volume expansion of ≈ 30, and we pinpoint again the increasing of contact frequencies in the aforementioned interval [0.01 Mbp − 0.1 Mbp].
The latter, in particular, constitutes a result with non trivial experimental implications. First, in connection to some recently proposed protocols for the reconstruction of chromosome conformations based on 3C (reviewed in [28]) where a monotonous relationship between chromatin distances and contacts is often assumed, our finding demonstrates that some caution is needed in order to avoid systematic bias in the final reconstructed structure. Second, the recent puzzling result [24] where chromatin domains with a high propensity to form 3C contacts seem to undergo rather pronounced decondensation when monitored by using FISH can be understood by considering that the two experimental techniques sample very different intervals of the corresponding pL(R)’s: close to the average value or the median for FISH, around to the lower tail for 3C techniques, as shown in the insets of Fig 1B, 1D and 1F. Our work thus supports the important conclusion of references [24, 29], namely that one needs to take into account both kinds of data to reconstruct correctly the shape of chromatin domains.
We complete the discussion by considering the full distributions pL(R) for L = 0.003, 0.015 and 3 Mbp, see Fig 2. For L = 0.003 and L = 0.015 Mbp (panels A-F) there are quantitative differences between the cases where the 10nm fiber is located at sparse random positions along the chromosome and where they form a single chromatin cluster, while for L = 3 Mbp (panels G-I) all distributions show no noticeable difference. In panels A-F, the largest peak corresponds to the most probable value of spatial distances between loci on the 30nm fiber region. Additionally, for random locations of 10nm fiber (panels A and D) corresponding pL(R)’s show a broader population of R values, while in the other cases there exist smaller peaks corresponding to the most probable distance between loci on the 10nm fiber region. Interestingly, similar distribution functions for spatial distances between chromatin loci seem to have been reported in yeast (see panels B, C, D, and E of Fig 1 in reference [30]). Although a direct comparison between these experimental results and our data is not possible (our setup applies to large chromosomes, like mammalian ones), we are tempted to speculate that the results reported in reference [30] are a manifestation of the presence of chromatin fibers of different compositions. Alternatively, we may interpret the observed shape of the pL(R)’s in terms of a bias towards large spatial distances: in that respect, we report that a similar feature has been observed in human and mouse chromosomes [14, 24] (including the experimental data that will be discussed in this work).
We have then considered those chromosome configurations with single long sequences of 10nm fiber and we have calculated 〈R2(L)〉 and 〈pc(L)〉 on these sequences, see Figs 3 and 4. Our results demonstrate that 〈R2(L)〉 increases systematically with respect to the analogous quantity for model chromosomes made of 30nm fiber, with no dependence on the specific location of 10nm fiber along the chromosome. This is also shown by the decreased contact probability. This insensitivity to positioning follows from the proposed picture [31] that chromosomes resemble a uniformly dense, semi-dilute solution of branched polymer rings. We expect then chromatin filaments to be similarly constrained or accessible regardless of their position along the genomic sequence.
In order to prove that the reported volume increase is not a fortuitous coincidence of the chosen sequence, we have also repeated the analysis on the same genomic region for the model chromosome with no 10nm fiber (0%) and the model chromosome entirely made of 10nm fiber (100%). Both 〈R2(L)〉 and 〈pc(L)〉 differ quantitatively from the case in which just one part of the chromosome is allowed to swell. Similarly to the previous section, the same results can be recast in terms of ratios to the corresponding quantities calculated for chromosome conformations with no 10nm fiber, see S2 and S3 Figs, which allows to appreciate to what extent chromosomes may effectively reorganize.
Because of the copolymer structure of our model chromosomes, the two types of fibers have different stiffnesses and thicknesses which cause the thinner fiber to move away from the thicker one. Consequently, 〈pc(L)〉 calculated for the 10nm fiber sequence tends to decrease and shows a scaling behavior ∼L−1.18±0.05 for L in the interval 0.01 − 1 Mbp which, being slightly steeper than the one for the entire chromosome, suggests an increase in the volume spanned by the fiber. Qualitatively, volume differences between chromosome regions with different transcription activities have been reported in a recent study on Drosophila [12]. Here the authors have studied three regions of the Drosophila genome, which, according to the histone modifications that they were bearing, were classified as active, inactive and polycomb-repressed. They found that the polycomb-repressed region was the most compact, followed by the inactive region, while the active region was the least compact. These results encourage us to speculate that the difference between the active and inactive regions could be due to different polymer physical properties induced by the histone modifications.
Here we discuss the impact of chromatin unfolding on the dynamics of the corresponding genomic loci. Specifically, we have considered the mean-square displacement 〈 δ r 2 ( τ ) 〉 ≡ 〈 ( r → i ( t + τ ) - r → i ( t ) ) 2 〉 at lag time τ, where r → i ( t ) is the spatial position of monomer i at time t and we implicitly assume average over specific monomer positions along the chromatin chain. In fact, this is an observable widely employed in many experiments monitoring the dynamic activity of specific chromatin loci, being especially suitable for comparing genome behavior in response to changes of the environment [32], or when the cell is targeted with drugs which are able to activate selectively certain types of genes [14].
Fig 5 summarizes our results for 〈δr2(τ)〉/τ vs. τ for the two cases of random positioning of small filaments of 10nm fiber within the chromatin fiber (panel A) and for chromosomes made of two large separated domains with different fiber composition (panel B). In both cases, we notice a general increase of chromatin mobility as larger and larger portions of 30nm fiber unfold, and, at larger times, a trend which does not substantially depend on the small scale details of chromatin fiber. Not surprisingly, data at short times reflect in part the discussed results for chromatin structure: in particular, we notice that differences in chromatin mobility before and after chromatin unfolding in random locations are only visible below time-scales of about 5 seconds (panel A). Slightly larger discrepancies are observed in the other situation where chromosomes are organized as two separate domains (panel B). In this latter case, unfolded chromatin loci move on average more than folded ones, the latter displaying the same motion than in the case of the homogeneously folded chromosome (compare green vs. blue lines).
In a recent work [13], Therizols et al. used FISH to show that, in embryonic stem cells, chromosome decondensation is sufficient to alter nuclear organization. Two sets of experiments were done: first, a viral transactivator was used to activate transcription of three genes (Ptn1, Nrp1 and Sox6). Alternatively, repositioning of the same genes was also observed after the treatment with an artificial peptide (DELQPASIDP) which decondenses chromatin without inducing transcription. By comparing the measured spatial distances between the two ends of each selected sequence, the authors concluded that nuclear organization is driven mainly by chromatin remodeling rather than transcription.
We have tested the predictions of our model by using data from these experiments. We have simulated the unfolding of a chromatin region of genomic length corresponding to the size of the specific gene we wanted to mimic. This situation corresponds to the second case studied in this work, namely the unfolding of a chromatin cluster. For a fair comparison, we have processed the experimental data as follows: we have reconstructed first the probability density distribution function for spatial distances between the ends of each gene, for the control condition (denoted as eGFP) and cells in which DELQPASIDP is recruited to the chromosome loci (denoted as DEL). Since FISH distances are recorded as two-dimensional vectors projected on the confocal plane, for the purpose of comparison a (large) set of three-dimensional distances with equivalent 2d projections was generated numerically by assuming random orientations of the 3d vectors in relation to the axis orthogonal to the confocal plane. It can be observed in Fig 6 panels A, C and E that, upon chromatin decondensation, the peak and the shape of the distribution change dramatically, in particular the distributions become wider. The same effect is displayed by our simulations (panels B, D and F). This comparison thus validates our result that larger genes tend to expand when unfolded.
While our model describes reasonably well the trend of the experimental data, the median values and the part of the distribution covering the larger distances, it seems to perform worse for the part of the distribution describing the small distances. This could be due to several factors which, to maintain the model simple, have been neglected. These include either protein linkers that physically bridge two sites along the genes forming a loop [20, 33, 34], either chemical modifications which could change the charge on some histones, thus inducing an effective electrostatic attraction [21] between regions of the gene that would make it more compact. Moreover, for the experiments using DEL, not all cells receive the exact same quantity of plasmid. That is, the experimental distributions of distances could be biased towards small R values because, in some cells, there was not enough plasmid to induce decondensation.
The choice for the physical parameters employed in this work as fibers flexibilities, thicknesses and excluded volume interactions (see section “Materials and Methods”) was mainly motivated by reasons of simplicity. Understanding to what extent our results depend on these specific choices requires a deeper analysis. In order to isolate and quantify the effects of one single parameter, we have considered a simpler model chromosome where monomers have fixed nearest neighbor distance along the sequence equal to 30 nm. The polymer was then split in two complementary domains: the fiber in one domain has the same features of the 30nm fiber, while the physical properties of the fiber in the other domain are changed, in turns, to: (1) completely flexible fiber; (2) thickness equal to 10 nm; (3) larger strength for the corresponding excluded volume interaction. Each of these changes was applied, in turns, to a domain occupying 20%, 50% and respectively 100% of the polymer. S4 Fig (top panels) shows that, not surprisingly, the persistence length influences 〈R2(L)〉 and 〈pc(L)〉 only at small scales up to 0.1 Mbp, suggesting that one of the main effects of having a heterogeneous chromatin composition consists in a global change of the overall persistence length of the chromatin fiber. By varying instead the range of the excluded volume interactions without modifying the nominal persistence length we produce variations in 〈R2(L)〉 and 〈pc(L)〉 similar to the ones reported in Figs 3 and 4 (middle panels). Viceversa, when varying the strength of the interaction no change is observed (bottom panels). These observations suggest a leading role for the effective range of the excluded volume interaction: in our model, the thinner 10nm fiber tends to occupy a larger region because its own self-repulsion becomes less important when compared to the repulsion from the thicker 30nm fiber. Quite interestingly, a similar effect concerning the change in the excluded volume was observed in equilibrium simulations of generic block copolymers [35]. Finally, we stress that this effect is less visible in Fig 1 because of the base pair content assigned to the monomers: since the base pair content of the 10nm fiber region is in general low compared to 30nm fiber region, its effect is not visible when averaging over the entire chromosome.
In this article, we have presented results of extensive Molecular Dynamics computer simulations of a coarse-grained polymer model for interphase chromosomes that extends previous numerical work [15–18] by introducing a crucial ingredient which was neglected before: namely, the presence of two kinds of chromatin fibers of different thickness and stiffness mutually interacting inside their own chromosome territory. Starting from a chromosome configuration made of a single and homogeneous 30nm fiber like filament, we have monitored chromosome spatial and temporal behaviors when this conformation is altered by the introduction of increasing amounts of 10nm-like fiber through the controlled unfolding of the thicker 30nm fiber.
The work shows that there exists detectable chromosome (re)organization for spatial scales smaller than 0.1 Mbp (Fig 1) and time-scales shorter than just a few seconds (Fig 5). Quite interestingly, these findings appear systematic and do not depend on the size of the chromosome region affected by the phenomenon of local unfolding or by its location along the chromosome. Interestingly, these results tend to suggest that experimental methodologies like FISH or HiC might be of little or no help in distinguishing between fibers of different compaction, unless they investigate genomic distances smaller than 0.1 Mbp. This prediction can be tested, for instance, by employing the recently developed oligonucleotide based FISH probes which seem to provide the necessary fine resolution [36, 37]. An important conclusion of our work is that, although our model uses parameter values that can be associated to the traditional “10nm/30nm” chromatin fiber paradigm [9, 10], our results reflect a generic physical effect [35] which ought to be observable in more general systems of crumpled polymers constituted of fibers with different thickness and/or stiffness (see section “Polymer physics aspects” and S4 Fig).
Our simulation protocol compares qualitatively well (see Fig 6) with experimental results on chromosome reorganization in mouse embryos treated with a synthetic transcription factor [13] which produces selective activation of a specific gene. In particular, we predict the observed shifting for distribution functions of spatial distances. Of course, it is quite possible that other mechanisms may explain the experimental results equally well. In fact, intentionally our model tends to neglect other important aspects of chromosome folding which have been highlighted recently by other authors: sequence-specific attractive interactions [21], protein linkers between chromatin fibers [20], mechanisms of active regulation [38, 39], “loop extrusion” [33, 34] involved in the reorganization of small chromosome domains, or the anchoring to the nuclear envelope or other nuclear organelles [40–42]. In the lack of a more quantitative analysis, we can only speculate on the fact that the inclusion of these mechanisms into our model might alter significantly the conclusions sketched here. On the other hand, this should represent an important stimulus for investigating further and more quantitatively the delicate relationship between chromosome structure and function.
In this work, the chromatin fiber is modeled as a coarse-grained polymer chain, with monomer-monomer interactions described by analytical expressions similar to the ones used in our previous works [15–17] and suitably adapted to take into account the different sizes of 30nm and 10nm monomers.
The full Hamiltonian governing the system, H, consists of three terms:
H = ∑ i = 1 N [ U F E N E ( i , i + 1 ) + U b r ( i , i + 1 , i + 2 ) + ∑ j = i + 1 N U L J ( i , j ) ] (2)
where N (see Table 1) is the total number of monomers constituting the ring polymer modeling the chromosome (see Section “Construction of model chromosome conformation” for details on this point) and i and j run over the indexes of the monomers. The latter are assumed to be numbered consecutively along the ring from one chosen reference monomer. The modulo-N indexing is implicitly assumed because of the ring periodicity.
By taking the nominal monomer diameter of the 30nm chromatin fiber, σ = 30 nm = 3000 bp [16], as our unit of length, the vector position of the ith monomer, r → i, the pairwise vector distance between monomers i and j, d → i , j = r → j - r → i, and its norm, di, j, the energy terms in eq 2 are given by the following expressions:
As in our previous work [15–18], chromosome dynamics was studied by performing fixed-volume Molecular Dynamics (MD) simulations with periodic boundary conditions at near-physiological fixed chromatin density ρ = 0.012 bp/nm3. Note that periodic boundary conditions do not introduce confinement to the simulation box: using properly unfolded coordinates, the model chromatin fibers can extend over arbitrarily large distances [15].
The system dynamics was integrated by using LAMMPS [44] with Langevin thermostat in order to keep the temperature of the system fixed to 1.0kB T. Given the unit mass m30 = 1 of the 30nm-bead, we fixed the mass of the 10nm-bead to m 10 = 1 27. The integration time step was fixed to tint = 0.001τMD, where τ M D = σ ( m 30 ϵ ) 1 / 2 is the elementary Lennard-Jones time. γ = 0.5/τMD is the friction coefficient [43] which takes into account the corresponding interaction with the background implicit solvent. The total length of each MD simulation run is = 3 ⋅ 105 τMD, with an overall computational effort ranging from a minimum of ≈ 103 to a maximum of ≈ 9⋅104 hours of single CPU for “0%” and “100%” model chromosome conformations, respectively. Single chromosome conformations were sampled every 103 τMD, implying 300 configurations per each run.
We have verified that our rings are well equilibrated by considering the mean-square distances, 〈R2(L)〉, and mean contact probabilities, 〈pc(L)〉, calculated on different sets of configurations log-spaced in time. As shown in S5 Fig for a 20% amount of 10nm fiber, all curves give the same results. Similar results are obtained for all simulated systems.
Quantities discussed in this work as the mean-square distances, 〈R2(L)〉, and the average contact frequencies, 〈pc(L)〉, between chromosome loci are plotted as a function of the genomic distance, L. In order to avoid unphysical behavior arising from the ring closure condition, we considered contour lengths L ≤ 1/4 of the total contour length of the ring, or L ≤ 30 Mbp.
Possible numerical artifacts due to the presence of monomers with different degrees of resolution (“10nm fiber” monomers vs. “30nm fiber” monomers) were removed by averaging over all possible pairs of monomers at fixed genomic separations L and spatial resolution of the 10nm fiber: this was achieved by replacing all 30nm fiber monomers not already decondensed by the 27 equivalent 10nm monomers, as shown in Fig 7 (panels C and D). In this way, each chromosome conformation always contributes with the same number of monomers and genomic distances smaller than 3 kbp can be effectively sampled. It is important to stress that what is described here concerns only the final analysis of the data, and it is not implicated in the motion of the monomers during the MD runs which is always performed as described in previous sections.
Final values for 〈R2(L)〉 and 〈pc(L)〉 at large L’s were obtained by averaging further over log-spaced intervals centered at the corresponding L’s. This procedure improves the accuracy by reducing considerably statistical fluctuations. Error bars were calculated accordingly.
|
10.1371/journal.pgen.1003485 | The Genome Organization of Thermotoga maritima Reflects Its Lifestyle | The generation of genome-scale data is becoming more routine, yet the subsequent analysis of omics data remains a significant challenge. Here, an approach that integrates multiple omics datasets with bioinformatics tools was developed that produces a detailed annotation of several microbial genomic features. This methodology was used to characterize the genome of Thermotoga maritima—a phylogenetically deep-branching, hyperthermophilic bacterium. Experimental data were generated for whole-genome resequencing, transcription start site (TSS) determination, transcriptome profiling, and proteome profiling. These datasets, analyzed in combination with bioinformatics tools, served as a basis for the improvement of gene annotation, the elucidation of transcription units (TUs), the identification of putative non-coding RNAs (ncRNAs), and the determination of promoters and ribosome binding sites. This revealed many distinctive properties of the T. maritima genome organization relative to other bacteria. This genome has a high number of genes per TU (3.3), a paucity of putative ncRNAs (12), and few TUs with multiple TSSs (3.7%). Quantitative analysis of promoters and ribosome binding sites showed increased sequence conservation relative to other bacteria. The 5′UTRs follow an atypical bimodal length distribution comprised of “Short” 5′UTRs (11–17 nt) and “Common” 5′UTRs (26–32 nt). Transcriptional regulation is limited by a lack of intergenic space for the majority of TUs. Lastly, a high fraction of annotated genes are expressed independent of growth state and a linear correlation of mRNA/protein is observed (Pearson r = 0.63, p<2.2×10−16 t-test). These distinctive properties are hypothesized to be a reflection of this organism's hyperthermophilic lifestyle and could yield novel insights into the evolutionary trajectory of microbial life on earth.
| Genomic studies have greatly benefited from the advent of high-throughput technologies and bioinformatics tools. Here, a methodology integrating genome-scale data and bioinformatics tools is developed to characterize the genome organization of the hyperthermophilic, phylogenetically deep-branching bacterium Thermotoga maritima. This approach elucidates several features of the genome organization and enables comparative analysis of these features across diverse taxa. Our results suggest that the genome of T. maritima is reflective of its hyperthermophilic lifestyle. Ultimately, constraints imposed on the genome have negative impacts on regulatory complexity and phenotypic diversity. Investigating the genome organization of Thermotogae species will help resolve various causal factors contributing to the genome organization such as phylogeny and environment. Applying a similar analysis of the genome organization to numerous taxa will likely provide insights into microbial evolution.
| A fundamental step towards obtaining a systems-level understanding of organisms is to obtain an accurate inventory of cellular components and their interconnectivities [1]–[3]. The genome sequence and in silico predictions of gene annotation are the starting points for assembling a network. For prokaryotes, these in silico approaches detect open reading frames and structural RNAs with varying degrees of accuracy [4]. Recently, multi-omic data generation and analysis studies [5]–[11] have revealed an abundance of genomic features that are not detected computationally such as transcription start sites (TSSs), promoters, untranslated regions (UTRs), non-coding RNAs, ribosome binding sites (RBSs) and transcription termination sites [12]. However, the rate at which multi-omic datasets are being generated is substantially outpacing the development of analysis workflows for these inherently dissimilar data types [13]. Here, multi-omic experimental data is generated and analyzed in conjunction with bioinformatics tools to annotate numerous bacterial genomic features that cannot accurately be detected using in silico approaches alone. This methodology was applied to study the genome organization of Thermotoga maritima—a phylogenetically deep-branching, hyperthermophilic bacterium with a compact 1.86 Mb genome.
Originally isolated from geothermally heated marine sediment, T. maritima grows between 60–90°C with an optimal growth temperature of 80°C [14]. This species belongs to the order Thermotogales that have, until recently, been exclusively comprised of thermophilic or hyperthermophilic organisms. Compared to most bacteria, Thermotogales are capable of sustaining growth over a remarkably wide range of temperatures. For instance, Kosmotoga olearia can be cultivated between 20–80 °C [15]. Recently, the existence of mesophilic Thermotogales [16], [17] was confirmed with the description of Mesotoga prima, which grows from 20–50 °C with an optimum at 37 °C [18]. Sequencing of M. prima revealed that it has the largest genome of all the Thermotogales at 2.97 Mb with ∼15% noncoding DNA [19]. T. maritima, which grows at the upper-limit known for Thermotogales, has one of the smallest genomes in this order and maintains one of the most compact genomes among all sequenced bacterial species (<5% noncoding DNA) [20], [21]. The short intergenic regions in the T. maritima genome (5 bp median) resemble those in the genome of Pelagibacter ubique, a bacterium that has undergone genome streamlining and has the shortest median intergenic space (3 bp) among free-living bacteria [20]. Although it remains unclear whether T. maritima has also undergone streamlining, both organisms encode only a few global regulators (four sigma factors in T. maritima versus two in P. ubique) and carry just a single rRNA operon. In contrast with P. ubique, T. maritima displays more metabolic diversity through its ability to ferment numerous mono- and polysaccharides [14], [22].
Thermotogales have been the focus of many evolutionary studies [23]–[25]. Organisms in hydrothermal vent communities, where many Thermotogales have been isolated, are thought to harbor traits of early microorganisms [26]. Phylogenetic analysis of 16S rRNA sequences place the Thermotogae at the base of the bacterial phylogenetic tree [27], [28]; however, Zhaxybayeva et al. [25] determined through analysis of 16S rRNA and ribosomal protein genes that Thermotogae and Aquificales (a hyperthermophilic order) are sister taxa. The authors also determined that the majority of Thermotogae proteins align best with those found in the order Firmicutes [25]; therefore, the exact phylogenetic position of Thermotogae is still unresolved. Nevertheless, members of this phylum are among the deepest branching bacterial species and, as such, prime candidates for evolutionary studies.
Thermophiles such as T. maritima implement numerous strategies at both the protein and nucleic acid levels to support growth at high temperatures. For instance, intrinsic protein stabilization is achieved by utilizing more charged residues at the protein surface, encoding for a dense hydrophobic core, and increasing disulfide bond usage [29], [30]. DNA is typically kept from denaturing by introducing positive supercoils via reverse gyrase activity while phosphodiester bond degradation is prevented by stabilization through interaction with cations (e.g. K+, Mg2+) and polyamines [31], [32]. However, the impact of temperature on genome features essential to gene expression such as promoters and RBSs remains largely unexplored. Bacterial transcription initiation is governed by recognition of promoter sequences by sigma factors, which load the RNA polymerase holoenzyme upstream of the transcription start site (TSS). Translation initiation is predominantly reliant on base pairing between the anti-Shine-Dalgarno sequence found near the 3′-terminus of the 16S rRNA and the Shine-Dalgarno sequence (i.e. the RBS). Therefore, thermophilic macromolecular synthesis machinery must establish and retain contacts with nucleic acids while facing greater thermodynamic challenges.
The integrated approach described here enables an experimentally anchored annotation of several bacterial genomic features including protein-coding genes, functional RNAs, non-coding RNAs, transcription units (TUs), promoters, ribosome binding sites (RBSs) and regulatory sites such as transcription factor (TF) binding sites, 5′ and 3′ untranslated regions (UTRs) and intergenic regions. This is achieved through the simultaneous analysis of genomic, transcriptomic and proteomic experimental datasets with complementary bioinformatics approaches. In addition to providing a valuable resource to the research community, this analysis framework facilitates quantitative and comparative analysis of annotated features across microbial species. For the genome of T. maritima, several distinguishing characteristics were identified and their potential causal factors are discussed.
An integrative workflow was developed to re-annotate the genome of T. maritima. The re-annotated genome is the result of the simultaneous reconciliation of multiple omics data sources (Figure 1, upper left) with bioinformatics approaches (Figure 1, upper right). Omics data generated included: (1) genome resequencing, (2) transcription start site (TSS) identification using a modified 5'RACE (Rapid Amplification of cDNA Ends) protocol, (3) transcriptome profiling using both high-density tiling arrays and strand-specific RNA-seq, and (4) LC-MS/MS shotgun proteomics. Transcriptome data were generated from cultures grown in diverse conditions including log phase growth, late exponential phase, heat shock, and growth inhibition by hydrogen (See Materials and Methods). Proteomic datasets include log phase growth and late exponential phase growth conditions. In combination with various bioinformatics approaches, integration of these omics datasets allowed for the definition of gene and transcription units (TU) boundaries with single base-pair resolution. The updated and expanded annotation served as the basis for genome-wide identification of promoters, ribosome binding sites (RBSs), intrinsic transcriptional terminators and UTRs.
The genome-wide identification of promoter and RBS sites was facilitated by the annotated TU start loci and protein start codons (Figure 2A). Promoter and RBS sequences were then quantitatively analyzed using thermodynamic principles. These same quantitative measures were applied to numerous organisms for interspecies comparison.
Regulation in T. maritima was studied from the vantage point of an organism with extremely short intergenic regions. In both microbes [55] and higher organisms [56] it was shown that the regulatory complexity of an operon positively correlates with the amount of intergenic space found upstream of that operon. Promoter-containing intergenic regions (PIRs) served as well-defined genomic regions for this analysis (Figure 3A). PIRs contain target sites for transcriptional regulation (e.g. promoters and TF binding sites) as well as translational regulation (e.g. RBSs). Each PIR can be divided into two components in relation to the TSS: the sequence downstream of the TSS (the 5′UTR) and the sequence upstream of the TSS.
Transcriptome data indicate that the genome of T. maritima is exceptionally active irrespective of growth condition (Figure 4A) with 91–96% of genes expressed above an FPKM threshold of 8. This fraction of genes transcribed is uncharacteristically high compared to other free-living bacteria (see Table S7). Furthermore, translational evidence supporting the high gene expression activity of T. maritima is found in the proteomic datasets. In each condition tested, peptide evidence was detected for 74% of the annotated proteins. It is also found that mRNA and protein abundances are tightly linked (Pearson r = 0.63, p<2.2×10−16 t-test) (Figure 4B). This correlation is stronger and more significant than those reported in comparable studies for other bacteria [58], [59].
Genome-scale technologies have provided researchers unprecedented access to large volumes of data detailing the composition of a cell. However, approaches for data analysis and interpretation have lagged behind due to the scope and complexity of these data types. Here, we present a framework for multi-omic data analysis that annotates genomic features involved in transcription, translation and regulation. This methodology integrates genome-scale datasets with bioinformatics predictions to produce 1) an improvement of the gene annotation, 2) an experimentally validated TU architecture and 3) the identification of putative antisense, non-coding transcripts and alternative TSSs. Using these annotated genomic features enabled the genome-wide identification of promoters and RBSs, which are difficult to identify solely using in silico approaches [60], [61]. Furthermore, the relative binding strength of individual promoters and RBSs was quantitatively measured using thermodynamic principles enabling multi-species comparison of these sequence features. The annotated genome organization served as a scaffold for analyzing regulatory features. Transcription factor regulation was examined with respect to promoter containing intergenic regions while the translational impact of the 5′UTR distribution was considered. The multi-omic data generation and analysis demonstrated here is applicable to many microbial species.
Applying this methodology to study the genome organization of T. maritima revealed that it has many distinctive properties compared to other organisms. Genome-scale analysis of promoters showed that T. maritima encodes a highly conserved, robust architecture that ensures transcription initiation. Similarly, RBS sequence conservation was shown to be thermodynamically sufficient for translation initiation for almost all T. maritima genes at 80°C compared with only a fraction of E. coli genes. The distinctive properties of the T. maritima genome extend beyond sequence composition and are apparent at the organizational level. The high protein-coding density and minimal intergenic space found in this organism have resulted in a high number of genes per TU, a paucity of putative ncRNAs and few TUs with multiple start sites. Furthermore, transcriptional regulation appears to be limited to a few TUs due to a lack of genomic space in PIRs. Interestingly, the 5′UTR component of the PIR was found to be uncharacteristically bimodal and was comprised of an unusually short grouping of 5′UTRs. Lastly, the constrained genome organization of T. maritima is reflected in the physiological state of the cell. Transcription of the vast majority of genes is detected independent of culture condition and the correlation between protein and mRNA is stronger than previously observed in other bacteria.
We hypothesize that the hyperthermophilic lifestyle of T. maritima could potentially explain the distinctive characteristics of this organism's genome organization. For instance, the increased sequence conservation of promoter elements and RBSs throughout the T. maritima genome may be attributed to the need to maintain gene expression under extreme temperature conditions. Macromolecular interactions (e.g. protein/protein, protein/DNA and RNA/RNA) are intrinsically harder to maintain at higher temperatures. In the case of TF binding sites, it has been shown that each nucleotide deviation from consensus results in a ∼2kbT penalty to the maximum binding free energy for a given TF (where kb is Boltzmann's constant and T is temperature) [62]. Increasing the temperature amplifies the binding free energy penalty for every non-conserved base pair. Therefore at 80°C, mismatches between the Shine-Dalgarno and anti-Shine-Dalgarno sequence are especially severe. Thus, T. maritima must overcome the intrinsic challenge of recognizing and retaining contact at the initiation site for both transcription and translation. Our data suggests that high sequence conservation of promoter and RBS sequences is one of the mechanisms used by T. maritima to ensure sufficient gene expression. This sequence-level adaptation could be analogous to many others observed in thermophilic organisms such as the amino acid composition of proteins [29], [30] and the GC content of structural RNAs [63].
The minimal intergenic space found in the T. maritima genome is reminiscent of a streamlined genome, which could explain the limited regulatory capacity observed in this organism. Inflexibility of metabolic regulons has been previously alluded to for other Thermotogales [64]. Here it is demonstrated that, for most TUs, a lack of physical space exists for transcriptional regulation by TFs. Furthermore, the Short 5′UTR group carries the minimum number of nucleotides needed to recruit the ribosome based on Shine-Dalgarno/anti-Shine-Dalgarno interactions [54]. Further reduction in 5′UTR length would abolish translation. Short 5′UTRs also reduce the capacity to regulate by limiting 5′UTR interactions [65], [66].
Though thermodynamics and physical space are hypothesized to contribute to the characteristic features of the T. maritima genome, the phylogenetic contribution cannot be dismissed. These potential causal factors are difficult to decouple. For RBSs, we were able to determine the impact of phylogeny and optimal growth temperature on RBS binding strength. By analyzing RBSs from 109 bacterial species spanning many phyla and having a diverse range of optimal growth temperatures we were able to demonstrate that both phylogeny and optimal growth temperature were significant determinants of RBSs sequence composition. However, a recent analysis of genome size among species of the order Thermotogales could not resolve the impact of phylogeny from optimal growth temperature [19]. The authors found that a negative correlation between genome size and optimal growth temperature exists within this order but the correlation did not hold when phylogeny was accounted for in the analysis. Interestingly, this study also found that the number of predicted transcriptional regulators and intergenic space is higher in Mesotoga prima, a mesophilic member of the Thermotogales. Thus, the relationship between phylogeny and the genome organization is difficult to elucidate without the generation of more datasets similar to the one presented here.
Thermotogae are an ideal phylum for future investigations on the causal impact of factors such as temperature, intergenic space and phylogeny on genome organization. This phylum contains organisms that are found in many diverse environments with a wide range of optimal growth temperatures. Generating multi-omic datasets and analyzing them using an integrated, quantitative workflow for numerous Thermotogae species would enable assessment of various environmental factors in the context of phylogenetic distance. Furthermore, given their phylogenetic depth, characterization of the Thermotogae will also provide insights in the evolutionary trajectory of microbial life on earth.
T. maritima MSB8 ATCC derived cultures were grown at 80°C under anoxic conditions in a chemically defined, minimal medium [67]. Cultures were maintained in either serum bottles or pH-controlled (6.5) fermenters with continuous 80% N2, 20% CO2 sparging. Maltose and acetate concentrations were measured using an HPLC. HPLC parameters were previously described [68]. The following growth conditions were used for omics analysis: 1) log phase, 2) carbon-limited late exponential phase, 3) heat shock and 4) H2 inhibition. Log phase samples were collected from mid-exponential phase cultures grown in 125 mL serum bottles with 50 mL working volume of media and 10 mM maltose as the sole carbon source. Carbon-limited late exponential phase cultures were grown in pH controlled fermenters with pH control and continuous stripping of evolved hydrogen. Cultures were monitored for OD and maltose concentration and samples were collected upon depletion of maltose. The heat shock condition was achieved by rapidly heating mid-exponential phase cultures grown in serum bottles (similar to the log phase condition) to 90°C and sampled after 10 minutes for transcriptome analysis. This has been shown to result in the heat shock response [69]. H2 inhibition was achieved by allowing the native evolution of hydrogen to accumulate in serum bottles (similar to the log phase condition). Arrested growth was indicated by successive OD readings that showed no change measured every 30 minutes. Growth profiles for these conditions are shown in Figure S4.
The recent identification of a 9 kb gap in the T. maritima MSB8 genome [33] prompted genome resequencing. Genomic DNA was isolated using Promega's Wizard Genomic DNA Purification Kit. Paired-end resequencing libraries were generated following standard Illumina protocols and sequenced on an Illumina GAIIx platform. The updated genome sequence was assembled as follows: (1) Reads were aligned to the 8.9 kb region identified in the T. maritima MSB8 DSMZ genomovar (AGIJ00000000.1) [33] and the TIGR genomovar (AE000512.1) sequence using SHOREmap [70] and MosaikAligner (http://bioinformatics.bc.edu/marthlab/Mosaik). (2) Unaligned reads were de novo assembled using Velvet [71] to ensure no additional assemblies were present. (3) The sequence was corrected for SNPs and indels detected during read alignment.
An updated genome annotation was generated using the RAST pipeline with the default parameters [34]. Predicted gene sequences were mapped to the AE000512.1 annotation using a bidirectional Smith-Waterman alignment to identify the corresponding locus tags. Instances where ≥30 bp separated the predicted gene length between annotations were reconciled through manual inspection of gene expression data and bioinformatics predictions. Gene length differences <30 bp could not be reconciled (unless peptide data supported only one annotation). In these cases, the updated sequence annotation was retained.
Total RNA was isolated from log phase cultures using the hot SDS/phenol approach as previously described (http://www.bio.davidson.edu/projects/GCAT/protocols/ecoli/RNApurification.pdf). DNase-treated total RNA samples were recovered using Fisher SurePrep TrueTotal RNA columns. Two biological replicate TSS sequencing libraries were constructed as previously described [7]. Illumina reads were aligned to the updated T. maritima genome using the Mosaik Aligner. The number of sequenced reads and the number of aligned reads can be found in Table S10. Only uniquely mapped 5′ ends with ≥5 reads were retained as potential TSSs.
Tiling array and RNA-seq data were generated under log phase growth, carbon-limiting late exponential phase, heat shock and hydrogen inhibited conditions. Total RNA was isolated using the TRIzol (Invitrogen) extraction procedure followed by DNase treatment and purification using either the Qiagen RNeasy Mini Kit (Tiling Arrays) or the SurePrep TrueTotal RNA columns (RNA-seq).
Custom tiling arrays were synthesized based on the AE000512.1 genome sequence by Roche Nimblegen to carry 71,548 probes with a mean interval of 25 bp. Probe information was remapped to the updated genome sequence. Of the original 71,548 probes, only 125 did not map. Labeled cDNA was generated and processed as previously described [7]. The Transcription Detector algorithm [72] determined probes expressed above background at a FDR = 0.05.
Paired-end, strand-specific RNA-seq was performed using the dUTP method [73] with the following modifications. rRNA was removed with Epicentre's Ribo-Zero rRNA Removal Kit. Subtracted RNA was fragmented for 3 min using Ambion's RNA Fragmentation Reagents. cDNA was generated using Invitrogen's SuperScript III First-Strand Synthesis protocol with random hexamer priming. Illumina reads were aligned to the updated T. maritima genome using Bowtie [74] with up to 2 mismatches per read alignment. The number of sequenced reads and the number of aligned reads can be found in Table S10. FPKM values were calculated using Cufflinks [75]. Functional RNA transcripts were excluded from FPKM determination.
Proteomics samples and data were generally prepared as previously described [76]. In summary, triplicate samples of both log phase and late exponential phase culture were lysed by French press, and proteins were extracted into global, soluble, and insoluble fractions. The three protein fractions were digested with trypsin (Promega) for 4 h at 37°C and then cleaned-up using C18 or SCX SPE columns (Supelco), as appropriate. Resulting peptide samples were separated in the first dimension by high pH HPLC (Agilent) and then analyzed by LC-MS/MS using C18 resin (Phenomenex) with an expontial gradient on a custom built LC platform coupled to a linear ion trap (LTQ) or a Velos Orbitrap mass spectrometer (Thermo Scientific) operated in data dependent mode. Peptides were identified by SEQUEST (Thermo Scientific) against a six-frame translation of the T. maritima genome with no protease specified in the search. Xcorr values were refined to conform to generally accepted criteria and were applied to result in a false discovery rate of 0.16% at the peptide level. Non-quantitative peptide-level data can be found in Table S8.
Normalized protein abundances can be found in Table S9. Quantitative Peptide-level data was extracted from Lerman et al. [77] and mapped to the CP004077 genome annotation. The following criteria were used to filter proteins for quantitative analysis: 1) the protein has a total spectral count ≥2 across all conditions (minimum of two unique peptides or a single unique peptide with two observations), 2) the protein has ≥1 observed peptide under log phase since our data was correlated against log phase transcriptome data. Redundant peptides (i.e. peptides mapping to multiple protein entries) were excluded from the analysis to minimize potential ambiguity. For quantitative analysis, we normalized the observed spectral counts for each ORF by the number of possible fully tryptic peptides in the ORF. The number of possible fully tryptic peptides for each ORF was determined using the Protein Digestion Simulator (http://omics.pnl.gov/software/ProteinDigestionSimulator.php). Default settings were used, except the parameter “Max Missed Cleavages” was set to 0 and “Minimum Residue Count” was set to 6. These options require fully tryptic peptides of at least length 6. This program only considers peptides 400–2000 m/z up to a charge state (z) of 3, hence a maximum fragment mass of 6000.
The process of determining individual σ70 promoter elements upstream of each unique TU start in T. maritima was an iterative process, involving two software packages: BioProspector [78] and MEME [79]. BioProspector is able to identify gapped motif elements so it was used to initially identify T. maritima motifs. In BioProspector, sequences 75 bp upstream of TU starts were searched for bipartite elements (6 and 9 bp in width) with a 10–25 bp allowable gap and visualized through WebLogo [80]. MEME provides deterministic position-weight matrices appropriate for information content calculations. The −10 and extended −10 boxes were searched [−1 to −18] upstream of the TSS while the −35 box was searched [−20 to −44]. E. coli TUs annotated with σ70 promoters and experimentally validated TSSs in the EcoCyc Database (version 15.0) [49] were extracted for comparative analysis.
A similar approach was applied to identify promoter motifs for alternative sigma factors. T. maritima has three annotated alternative sigma factors: RpoE (Tmari_1606), SigH (Tmari_0531) and FliA (Tmari_0904). For RpoE and SigH, the upstream region of TUs having genes showing high differential expression under a given stress condition (heat shock, hydrogen inhibited and carbon-limited late exponential phase) were searched for motif elements. The upstream regions of flagellar gene encoding TUs were searched for a FliA motif. However, no sequence motif could be detected for any of the three alternate sigma factors.
Position weight matrices (PWMs) for each promoter element were converted to individual information weight matrices using the following formula established in the field of molecular information theory [43]: Riw(b, i) = 2−(−log2f(b, i)), where f(b, i) is taken to be the probability of observing base b at position i. The individual information of a sequence, Iseq, was calculated by summing the relevant entries of Riw. For any particular sequence, only one entry of Riw is relevant among 4 bases for each position i in the sequence. Iseq is measured throughout in bits since the log was base 2 in converting the PWM to Riw.
Iseq reflects sequence conservation for a single sequence, but natural promoters are often formed by multiple promoter elements, each with their own sequences and corresponding Iseq values. When multiple elements are present, variable length spacers are frequently found between the elements. We applied an approach previously described by Shultzaberger et al. [48] to properly account for all possible promoter elements and the variation in their spacing. This allowed us to assess total sequence conservation for an entire promoter. For each promoter, the information content for a particular binding mode was calculated based on the formulas: (1) Mode 1: Iseq_whole_promoter = Iseq(−10 element)+Iseq(−35 element)−GS(d); (2) Mode 2: Iseq_whole_promoter = Iseq(extended−10 element); (3) Mode 3: Iseq_whole_promoter = Iseq(extended−10 element)+Iseq(−35 element)−GS(d). GS(d) is ‘gap surprisal’ accounting for variable spacing (of length d) between the −10 and −35 elements. GS(d) penalizes for unexpected spacing given the major groove accessibility of B-form DNA and was defined as in equation (3) in Shultzaberger [48] with no small-sample correction factor as the analysis here is performed at genome scale. In accordance with the Shultzaberger model, the space between the −10 and −35 elements was restricted to 15–20 bp as measured from the 3′ end of the −35 element and the 5′ end of the −10 element. This limit on the spacer distance Iseq_whole_promoter is measured in bits.
The anti-RBS sequence 5′-UCACCUCCUU-3′ (3′ end of the 16S rRNA) was selected for this study. The hybrid-2s program in the UNAFold software package [81] was used to compute hybridization energies (ΔG) for all possible 10-mers over the temperature range 20–100°C. This dictionary was mined for three applications: (1) binding energy values for all 10-mer sequences in the updated T. maritima genome were computed to aid in annotation improvement, (2) the median positional ΔG for all CDSs ±100 bp from the start codon, and (3) the local minimum ΔG for all CDSs 30 bp upstream of the start codon. RBS binding energies across 109 organisms were calculated using this dictionary. Optimal growth temperatures for all non-Thermotogae bacteria were collected from Takemoto et al. [82] and the protein coding gene annotation for each bacterium was extracted from NCBI. CDS data for all Thermotogae with a complete genome sequence were extracted from NCBI with the exception of T. maritima for which the annotation generated in this study was used. For each organism, the median RBS ΔG was calculated from the set of minimum RBS ΔG's found for each CDS 30 bp upstream of the annotated start codon. Three distance matrices were constructed for analysis of the 109 bacterial species for which optimum growth temperatures were found. The matrices included are as follows: (1) the absolute difference of median RBS strength values, (2) the absolute difference of optimal growth temperatures and (3) the distance matrix generated by aligning full-length 16S rRNA gene sequences using ClustalW2 (slow mode) followed by the phylogenetic tree generation script (http://www.ebi.ac.uk/Tools/phylogeny/) with default settings. Next, the Mantel test, which tests the correlation between two distance matrices, was applied to compute the significance of various correlations. The ‘vegan’ package of R was used with its default settings.
Intrinsic terminators were predicted using the TransTermHP program [83]. To avoid bias introduced by annotation, no genome annotation was used in prediction of Rho-independent terminators. Only terminator structures predicted with a “100%” confidence score were included in the curation of TUs.
Small RNAs were predicted with Infernal [39] using cmsearch with default settings against the Rfam 10.0 Database [84] of small RNA families. sRNAs with an E-value<0.01 were manually curated to verify expression. These sRNAs were checked against the sRNA predictions from Rfam and fRNA-DB (http://www.ncrna.org) based on the AE000512.1 genome sequence.
TU assembly was accomplished through an iterative procedure beginning with tiling array expression data. Tiling array data was processed with two Bioconductor packages for transcript segmentation based on change point analysis: tilingArray (http://www.bioconductor.org/packages/2.2/bioc/html/tilingArray.html) and DNAcopy (http://www.bioconductor.org/packages/2.3/bioc/html/DNAcopy.html). Manual comparison of the output from both packages with array data was used to refine the automated set of transcriptional segments. Additional datasets and bioinformatics predictions were added and manually curated to fully characterize the TU assembly. TSS and RNA-seq data provided single-base pair resolution of segment boundaries. Intrinsic terminator predictions were also used for 3′ boundary definition. ncRNAs were identified using the transcript segments. Transcribed regions not associated with a TU and with length exceeding 68 nt (the combined length of the paired end reads with no insert separating them) were quantified using Cufflinks to generate FPKM values across all RNA-seq conditions. Regions with at least two conditions showing FPKM values >8 were retained as putative ncRNAs.
TF binding sites were extracted from RegPrecise [57] and coordinates were mapped to the updated genome. Table S6 has the TF binding sites used in Figure 3C.
The T. maritima MSB8 ATCC (genomovar) genome and annotation are found under Genbank Accession CP004077. RNA-seq, TSS, and tiling array datasets are available in the Gene Expression Omnibus under Accession GSE37483. Proteogenomic data are made available through PNNL (http://omics.pnl.gov) and in Table S8.
|
10.1371/journal.pntd.0005948 | Sanitation, hookworm, anemia, stunting, and wasting in primary school children in southern Ethiopia: Baseline results from a study in 30 schools | Inadequate nutrition; neglected topical diseases; and insufficient water, sanitation, and hygiene (WASH) are interrelated problems in schools in low-income countries, but are not routinely tackled together. A recent three-year longitudinal study investigated integrated school health and nutrition approaches in 30 government primary schools in southern Ethiopia. Here, we report on baseline associations between sanitation, hookworm infection, anemia, stunting, and wasting.
In each school, the Schistosoma mansoni, S. haematobium, and soil-transmitted helminth infection intensities; blood hemoglobin concentrations; heights; and weights of approximately 125 students were assessed. Of these 125 students, approximately 20 were randomly selected for student WASH surveys. Of these 20, approximately 15 were randomly selected for household sanitation observations. School WASH was also assessed through a combination of observations and questions to the headteacher. Mixed-effects logistic regression was used to compare household sanitation with hookworm infection (the other parasites being much less prevalent); and hookworm infection with anemia, stunting, and wasting.
Blood, stool, and urine samples were provided by 3,729 children, and student WASH and household WASH surveys were conducted with 596 and 448 of these students, respectively.
Hookworm, Ascaris lumbricoides, Trichuris trichiura, and S. mansoni infections had prevalences of 18%, 4.8%, 0.6%, and 0.3%, respectively, and no S. haematobium infections were found. Anemia, stunting, and wasting had prevalences of 23%, 28%, and 14%, respectively.
No statistically significant associations were found between latrine absence or evidence of open defecation at home, and hookworm infection (adjusted odds ratio, OR = 1.28, 95% confidence interval, CI: 0.476–3.44; and adjusted OR = 1.21, 95% CI: 0.468–3.12; respectively); or between hookworm infection and anemia, stunting, or wasting (adjusted OR = 1.24, 95% CI: 0.988–1.57; adjusted OR = 0.992, 95% CI: 0.789–1.25; and adjusted OR = 0.969, 95% CI: 0.722–1.30; respectively).
In this setting, no statistically significant associations were found between sanitation and hookworm; or between hookworm and anemia, stunting, or wasting. More evidence on best practices for integrated school health interventions will be gathered from the follow-up surveys in this study.
| It is thought that inadequate sanitation may exacerbate hookworm transmission, and that hookworm infection may give rise to health problems including anemia, stunting, and wasting. Integrating monitoring of, and interventions against, these problems may yield significant cost savings. Such integrated interventions should be guided by both evidence of the relationships between the various health problems, and examples of optimally effective integration. Here, we present baseline findings from a three-year longitudinal study investigating combined school feeding; school water, sanitation, and hygiene (WASH); and deworming interventions in 30 primary schools in southern Ethiopia. In particular, we compare household sanitation, hookworm infection, and anemia, stunting, and wasting, in the schoolchildren. None of these associations were found to be statistically significant (although hookworm infection was associated with borderline statistically significantly higher odds of anemia; P = 0.06). The lack of significant associations may be due to the low intensities of the hookworm infections in this setting, poor conditions of latrines reducing their impact on hookworm transmission, and other factors such as malaria and inadequate diet causing anemia, stunting, and wasting.
| Integrating school health programs might yield an efficient solution to Ethiopia’s interrelated child health problems of malnutrition [1]; soil-transmitted helminth (STH) [2, 3] and schistosome infections [3–6]; and inadequate water, sanitation, and hygiene (WASH) coverage [3, 7]. The nutritional benefits of school feeding might be reinforced with preventive chemotherapy (PC) against STHs and schistosomes, which can cause malnutrition [5, 8, 9], while malnutrition may also increase children’s susceptibility to parasitic infection [10]. Hookworm infection in particular can be an important cause of anemia, due to the gastrointestinal blood loss and decreased appetite that it can cause [8, 11], and hookworm-induced anemia can slow the cognitive and physical development of children [12]. Furthermore, the impact of PC against schistosomes and STHs might be strengthened though the improvement of WASH, if such improvements curb the parasites’ transmission.
There is increasing interest in using the school as a platform for providing various nutrition and health interventions, such as PC [13], school feeding, and health education [14]. In addition to the overlapping health benefits, combining school feeding, PC, and WASH delivery into a single platform might lead to significant cost savings [15].
A collaboration between the Ethiopian Public Health Institute (EPHI) and the Partnership for Child Development at Imperial College London, has been investigating the optimal integrated delivery of these three school health interventions, in a three-year longitudinal study in 30 government primary schools in the Southern Nations, Nationalities, and Peoples’ Region (SNNPR), Ethiopia. Here, we present baseline associations between household sanitation, hookworm infection (the other parasites being much less prevalent), and anemia, stunting, and wasting. The wider program will be used to investigate the feasibility of this form of implementation, its costs, and health and education outcomes.
The study protocol was reviewed and approved by the Scientific and Ethical Review Committee (SERC) of the EPHI. On the Imperial College London side, this study was covered by ethical approval granted by the Imperial College London Research Ethics Committee for the monitoring and treatment of schistosomiasis and STHs (reference: ICREC_8_2_2). The aims and procedures of the study were explained to participants prior to enrolment in the study. Written informed consent for sample collection and student WASH surveys was obtained from the headteachers in place of the parents, and for the household WASH surveys from the heads of household. Each participant provided verbal consent, and was reminded of his or her right to withdraw from the study at any time, without consequences. Praziquantel (60 mg/kg) was administered by health officers or clinical nurses (HOCNs) to all children testing positive for schistosomiasis. In schools with non-zero prevalence of STHs, all children present were treated with 400 mg of albendazole; tablets were dispensed by the teachers, and records of the number of children treated were subsequently shared with the Regional Bureau of Health. A total of 22,258 children received PC with albendazole.
The 30 schools were selected by the Ethiopian Ministry of Education in partnership with the United Nations World Food Programme (WFP). They were chosen from schools already receiving school feeding from the WFP, on the basis of suitability for home-grown school feeding; this suitability stemmed from factors such as proximity to agricultural cooperatives and local agricultural practices. Subsequently, 15 of those 30 schools were randomly selected to receive a WASH upgrade from SNV (the Netherlands Development Organisation), to enable the broader project to generate evidence for the costs and benefits of combined school health and nutrition interventions.
At baseline, the schools had a combined enrolment of 30,705 students. They are distributed in four clusters, located in Konso, Alle, Kindo Koysha, Lanfero, Mareko, and Kokir Gedebano woredas, which are spread across SNNPR (Fig 1). The schools’ elevations vary from 785 m (in Konso) to 2,859 m (in Kokir Gedebano) above sea level. SNNPR has a predominantly (90.0%) rural population [16]. Mean annual rainfall in the region increases from south to north, with values between 300–1,000 mm in Konso in the south, and 1,500–3,000 mm in the north [17]. Temperature varies with elevation, but annual mean temperatures are around 20°C both in Konso, in the south, and in Jimma, which is just north of the region [18, 19].
The data were collected in June 2013 by eight teams, recruited from the zonal health offices of the included schools. Each team comprised two HOCNs and two laboratory technicians. The laboratory technicians carried out the parasitological, height, weight, and hemoglobin surveys, while the HOCNs carried out the WASH and school feeding surveys. Two days’ training prior to the survey refreshed the laboratory technicians’ understanding of the procedures to be used, familiarized the HOCNs with the survey forms to be used, and enabled standardization of the procedures across the data collectors. A third day of training included a school visit, where data was taken under close supervision and any misunderstandings were addressed.
Data were collected on paper and double-entered into CS-Pro version 5 (United States Census Bureau, Washington, DC, USA). The databases were exported to SPSS version 13 (IBM Corp., Armonk, NY) for cleaning. Analyses were then conducted in R version 3.3.0 (R Foundation for Statistical Computing, Vienna, Austria).
Participants’ Kato-Katz egg counts were used to classify them as hookworm-negative (0 eggs per gram of feces, EPG), or lightly, moderately, or heavily infected in accordance with WHO guidance [22] (1–1,999, 2,000–3,999, or 4,000 EPG or more, respectively).
Children were categorized as non-, mildly, moderately, or severely anemic as defined by the WHO [23], having adjusted blood hemoglobin concentrations to account for elevation, in accordance with the same guidelines. The body mass index (BMI) was calculated for each child as the weight in kilograms divided by the square of the height in meters. Height and BMI were then transformed to height-for-age Z-score (zHFA) and BMI Z-score (zBMI), using the Box-Cox power, median and coefficient of variation (L, M, and S) values from the WHO Growth Reference 2007 dataset [24, 25]. Since ages were recorded in years in this study, but the WHO standards are provided for ages in months, median standard values were used for each age group in years; for example, girls of age 10 years were compared with the standards for girls of age 10 years and six months. We adopted the frequently-used convention of defining wasting and stunting as zBMI and zHFA scores of less than -2 respectively, and severe wasting and stunting as values less than -3 [26–31].
Very few hookworm infections were of moderate or heavy intensity, so light, moderate, and heavy infections were pooled to give a binary uninfected/infected variable for the subsequent comparative analysis. The anemia, stunting, and wasting variables were also condensed to binary anemic/non-anemic, stunted/non-stunted, and wasted/non-wasted variables for the comparative analysis, in order to maximize statistical power.
Overall A. lumbricoides, S. haematobium, S. mansoni, and T. trichiura prevalences were all low (4.8%, 0.0%, 0.3%, and 0.6%, respectively) compared with hookworm (18%), so while all parasitic infections were summarized, only hookworm infection was compared with sanitation and other aspects of child health. Mixed-effects logistic regressions (implemented using version 1.1–12 of the lme4 package [32]) were used to compare household sanitation risk factors (the absence of a latrine, and evidence of open defecation at home) with hookworm infection (defined as a non-zero egg count by Kato-Katz), and hookworm infection with anemia, stunting, and wasting. These regressions also accounted for age, gender, and school cluster (all as fixed effects), and school (as a random intercept). Each model excluded participants missing any relevant data, or of an age outside the range of 5–18 years.
Logistic regression’s assumption of linearity of the logit with age (the only numerical variable) was checked using the Box-Tidwell method: for each model, statistical significance (P < 0.05) of an introduced age*log(age) interaction term was taken as being indicative of a non-linear dependence on age [33–36]. The logit of wasting did exhibit such nonlinearity with age, and therefore in all models, age was split into the classes of 5–10, 11–12, and 13–18 years (chosen to maximize equality in the class sizes). Multicollinearity between the independent variables incorporated as fixed effects was assessed through the inspection of variance inflation factors (VIFs). VIFs were calculated using the vif.mer function [37], and values above two would have been taken as requiring further investigation [36, 38], but none were found.
A total of 3,729 children provided blood, stool, and urine samples, of whom 1,955 (52%) were male. Student WASH surveys were conducted with 596 of these students, and the houses of 448 were visited for WASH assessment. The 3,729 children had a mean age of 11.8 years (SD = 2.1, data missing for 21 children).
Of the 30 schools, 16 (53%) reported having a designated handwashing time before serving food, but only one school (3.3%) had a handwashing area away from the latrines and with piped water. All schools but one had onsite latrines, and in total there were 56 latrines in the 30 schools. These were most commonly pit latrines with cement floors (47 latrines, 84%), but six (11%) were pit latrines without cement floors, and three (5.4%) were ventilated improved pit (VIP) latrines. Eight latrine floors (14%) were cracked, one (1.8%) had collapsed completely, and one more (1.8%) was missing data, but the rest were in good structural condition. Latrine floors were most commonly described as “unclean” (30 latrines, 54%), followed by “very unclean” (22 latrines, 39%), with only four latrines (7.1%) described as “clean”. On average, there was a usable latrine stall (that is, one whose floor had not completely collapsed) for every 104 boys and one for every 109 girls. Considering only latrines with doors in addition to floors, these figures rose to 177 boys and 174 girls. Evidence of open defecation was observed in 16 of the 30 schools (53%).
Of the 596 participants responding to student-level WASH questionnaires, the majority (388, 65%) reported no problem with school sanitation, and 434 (73%) reported no problem with home sanitation. The most common complaints were that school sanitation was too dirty (110, 18% of respondents), and that there was no toilet at home (33, 5.5% of respondents). Handwashing at various times was reported by the following numbers and proportions of students: after defecation by 299 students (50%), after urination by 116 students (19%), before eating by 526 students (88%), and when hands were visibly dirty by 467 students (78%).
Sanitation was absent in 122 (27%) of the 448 households. Of the 326 households that did have sanitation, this overwhelmingly (288, 88%) consisted of pit latrines without cement slabs. Latrines had no walls in 146 (45%) cases (data missing for two households, 0.6%), but the floors of 134 (41%) were described as “clean” (data missing for 14 households, 4.3%). Evidence of open defecation was observed at 131 (29%) households.
Schools have been used successfully as a platform for many different health and nutrition interventions around the world. To date, many of these interventions have been conducted in silos. In this project, the assessment of the parasitological status of students was complemented with the collection of a rich dataset pertaining to school feeding, WASH conditions of schools and households, and students’ WASH KAP. This data was collected by the same trained enumerators in a single school visit.
Of these schoolchildren, 23% were found to be infected with S. mansoni or an STH, and most of these infections were with hookworm (prevalence of 18%). This is consistent with other studies from the region. In particular, in two recent studies, Bayesian geostatistical models used climatic, socioeconomic, and parasitological survey data to predict a high prevalence of hookworm throughout much of SNNPR (and a national prevalence of 17.7%), along with much patchier distributions of A. lumbricoides and T. trichiura (and national prevalences of 8.5% and 6.1%, respectively), and very low levels of S. mansoni and S. haematobium infection (and national prevalences of 8.9% and 8.3%) [2, 6]. The high hookworm prevalence we found in Kindo Koysha aligns with the findings of a study by Taye et al. (2013), who, in a study of podoconiosis patients nearby, found a hookworm prevalence of 40.9% [40]. Around Butajira (a town between the Mareko and Kokir Gedebano schools), Davey et al. (2005) found prevalences of 16.6%, 14.7%, and 2.5% for A. lumbricoides, hookworm, and T. trichiura, respectively [41]. The prevalences of stunting and wasting that we found (28% and 14%, respectively) are comparable with the values of 45% and 9% found in a study of infants up to two years of age in Halaba (in the north of SNNPR) in 2013 [42]. Another study, in 2013 in Bule Hora, Oromia (just east of SNNPR) found prevalences of 47.6% and 13.4%, respectively, in children aged up to five years [43]. Regarding anemia, Deribew et al. (2010) found a prevalence of 32.4% in a study of under-fives in southern Ethiopia [44]; similar to the 23% that we found.
Hookworm transmission takes place when eggs in an infected person’s feces are deposited on the ground, before hatching to release larvae, which develop in the soil and infect a person during dermal contact with infested soil [9]. Consistent latrine use should contain these eggs and larvae, thereby reducing hookworm transmission. However, in this study, neither evidence of open defecation around the home compound, nor the absence of a latrine at home, were statistically significant predictors of hookworm infection. There are a number of possible explanations for this. Firstly, latrine presence may not lead to consistent latrine use, and although the evidence of open defecation at home was used as an indicator, this only reflects recent open defecation, while the hookworms’ lifespans have been estimated at 5–7 years [9]. Secondly, even if home sanitation has the effect of completely removing parasite eggs from soil around the household, people may still be exposed to infection away from home: indeed, it is likely that much hookworm transmission takes place between, rather than within, households [45].
These factors, along with the potential for socioeconomic confounding in observational studies (higher socioeconomic status, SES, being a cause of both better access to sanitation, and lower exposure to hookworms), may explain the mixed results of previous studies of the relationship between sanitation and hookworm infection. In a recent systematic review and meta-analysis, Strunz et al. (2014) did not find a statistically significant difference in odds of hookworm infection between those with and without access to sanitation [46]. Indeed, even a recent large cluster-randomized trial found no impact of sanitation on hookworm, a finding the authors attributed to insufficient coverage and use of the sanitation [47]. However, in another systematic review by Ziegelbauer et al. (2012), availability of sanitation was associated with a significantly lower odds of hookworm infection [48], while Ethiopia’s recent national mapping of schistosomes, STHs, and school WASH revealed a borderline statistically significant, negative association (Kendall’s τb = -0.039, 95% CI: -0.090–0.012, P = 0.067) between the adequacy of school sanitation and hookworm infection intensity [3]. In this survey, 88% of the latrines inspected had no cement slab. They may therefore actually be exacerbating hookworm transmission, by concentrating defecation spatially, and providing suitable conditions for larval development [49].
Participants with hookworm infections did not have statistically significantly higher odds of anemia, though the association might be described as borderline significant (P = 0.06). It is known that hookworms feed on their hosts’ blood, and this, as well as bleeding caused by hookworms attaching to the intestinal mucosa, can cause anemia [50, 51]. However, the impact of hookworm infection on blood hemoglobin concentration is strongly dependent on the infection intensity. A meta-analysis showed that children with heavy hookworm infections had statistically significantly lower blood hemoglobin concentrations than uninfected children, but that children with light infections did not [52]. Most (91.1%) of the hookworm infections we found were of light intensity, which may well explain the lack of a significant association between hookworm infection and anemia. Confounding by malaria and poor nutrition, both of which can also be associated with anemia [44, 53, 54], may also have weakened this association. Hookworm species were not differentiated in this study, but previous studies in Jimma (just north of SNNPR) and Mirab Abaya (in SNNPR), found the majority of hookworm infections to be with Necator americanus, rather than with Ancylostoma duodenale [55, 56]. N. americanus worms cause less blood loss than A. duodenale [12, 51]: a predominance of N. americanus worms may therefore have been another reason for the lack of an association between hookworm infection and anemia in this setting.
In all the models, statistically significant differences were seen between the clusters. As they were spread across SNNPR, there was substantial variation in elevation and climate. These climatic differences may explain some of the inter-cluster variation in all models: in the case of hookworm, a suitable environment is needed to allow development of larvae in the soil [9]. Anemia, stunting, and wasting may be caused by dietary differences, caused by climatic (as well as cultural and economic) differences between the clusters.
Hookworm-positive participants in this study did not have significantly higher odds of stunting, or wasting. This finding is supported by another study recently carried out in northern Ethiopia, which found that helminth infection (prevalence of 69.1%, 62.8% of which were hookworm infections or co-infections) was not a statistically significant determinant of either stunting or wasting [57]. A study in an area of extreme poverty in Peru, with a hookworm prevalence of 21.3% (mostly light-intensity infections), did reveal a significant association between hookworm infection and being both stunted and underweight (adjusted OR = 1.74, 95% CI: 1.05–2.86) [58]. However, no significant association between hookworm infection and stunting was found in a Brazilian community with a hookworm prevalence of 69.8% [59]. Similarly, analyses of data from the Lao People’s Democratic Republic [60], and Bolivia [61], did not reveal statistically significant associations between hookworm infection and anthropometric indicators.
Recent meta-analyses have investigated increases in height, weight, and blood hemoglobin concentration following treatment of STHs [62, 63]. These have found that mass treatment appears to have little to no effect on height or weight [62, 63], or hemoglobin [62], but that it is possible that targeted treatment of infected children does increase their weight gain, over the following six months [62]. These findings are broadly in line with ours.
In this survey, boys were at statistically significantly (P < 0.05) higher risk of anemia and wasting, and also at a higher risk of stunting (though statistically non-significant, P = 0.09). Svedberg (1990) demonstrated that boys in Sub-Saharan Africa frequently suffer worse nutrition than girls, and interpreted this as resulting from preferential treatment of girls, due to social factors such as high female participation in agricultural labor, polygamy, bridewealth (wealth transferred at marriage from the groom or his family to the bride’s parents), and early marriage of females [64]. Wamani et al. (2007) conducted a meta-analysis of 16 Demographic and Health Surveys in Sub-Saharan Africa, and found that boys suffered statistically significantly more stunting [65]. The authors also discussed how the cause may be biological, rather than social: boys may be more vulnerable to malnutrition because natural selection favors a sex ratio of 1:1, and the number of boys born slightly exceeds that of girls [65]. Regarding anemia, our finding that boys were at greater risk is supported by a cross-sectional survey of children aged 7–18 years in Tanzania [66].
Our analyses did not explicitly account for socioeconomic confounding, and this is perhaps their biggest drawback. Frequently-used indicators of SES, such as possession of a motorcycle or television, would have been of no use since these were completely absent in the households, but other indicators such as household floor construction, possessions at home, and ownership of livestock may have been useful. That said, most participants’ households were not visited. As recipients of WFP school feeding, these schools are all in areas considered to be food-insecure, and hence all the study participants might be considered to be of low SES.
Another limitation was the use of the Kato-Katz method on only one slide per participant, for the diagnosis of S. mansoni and STHs. This method is recognized to have a low sensitivity, particularly when slides are not examined immediately [67, 68]. It is therefore possible that some participants recorded as being hookworm-negative will in fact have had light infections.
Ideally, participants’ ages would have been recorded in months rather than years. This would have allowed for more robust comparisons with the WHO growth reference standards [25], which are provided for each month. Unfortunately, in this setting, children’s dates of birth were not available, but rather only their ages in years. Future studies that include anthropometry and interviews with parents, might benefit from the development of local calendars. Such local calendars should enable the assessment of children’s ages to within a few months. They could be developed using major events (such significant weather events including storms or droughts, or political events including elections) to ascertain the year, and then using questions about the season to estimate the month, of birth.
The student sanitation and hygiene indicators were self-reported, and shame and fear of reproach may therefore have influenced participants’ responses. For this reason, these indicators were not included in the models. Alternative approaches, such as sensors or global positioning systems (GPS) may enable the collection of more objective data, but will also be accompanied by substantial ethical and practical challenges. These include the invasions of privacy inherent in monitoring of location and WASH-related behaviors, as well as the possibility of sensors being removed.
No statistically significant associations were found between home sanitation and hookworm infection; or hookworm infection and anemia, stunting, or wasting, suggesting that in this setting, these different aspects of poor child health are not exacerbating each other. However, the lack of access to adequate sanitation, and the prevalences of hookworm, anemia, stunting, and wasting in and around these schools confirm the need for more interventions to improve child health in the region. Integrated approaches incorporating different types of interventions may prove to be the most efficient.
|
10.1371/journal.pgen.1006108 | Transposable Elements versus the Fungal Genome: Impact on Whole-Genome Architecture and Transcriptional Profiles | Transposable elements (TEs) are exceptional contributors to eukaryotic genome diversity. Their ubiquitous presence impacts the genomes of nearly all species and mediates genome evolution by causing mutations and chromosomal rearrangements and by modulating gene expression. We performed an exhaustive analysis of the TE content in 18 fungal genomes, including strains of the same species and species of the same genera. Our results depicted a scenario of exceptional variability, with species having 0.02 to 29.8% of their genome consisting of transposable elements. A detailed analysis performed on two strains of Pleurotus ostreatus uncovered a genome that is populated mainly by Class I elements, especially LTR-retrotransposons amplified in recent bursts from 0 to 2 million years (My) ago. The preferential accumulation of TEs in clusters led to the presence of genomic regions that lacked intra- and inter-specific conservation. In addition, we investigated the effect of TE insertions on the expression of their nearby upstream and downstream genes. Our results showed that an important number of genes under TE influence are significantly repressed, with stronger repression when genes are localized within transposon clusters. Our transcriptional analysis performed in four additional fungal models revealed that this TE-mediated silencing was present only in species with active cytosine methylation machinery. We hypothesize that this phenomenon is related to epigenetic defense mechanisms that are aimed to suppress TE expression and control their proliferation.
| Transposable elements (TEs) are enigmatic genetic units that have played important roles in the evolution of eukaryotic genomes. Since their discovery in the 1950s, they have gained increasing attention and are known today as active genome modelers in multiple species. Although these elements have been widely studied in plants, much less is known about their occurrence and impact on the fungal kingdom. Using a diverse set of basidiomycete and ascomycete fungi, we quantified and characterized a huge diversity of DNA and RNA transposable elements, and we identified species that had 0.02 to 29.8% of their genomes occupied by transposable elements. In addition, using our basidiomycete model Pleurotus ostreatus, we demonstrated how TE insertions produced detrimental effects on the expression of upstream and downstream genes, which were downregulated compared with the control groups. This silencing mechanism was present in the basidiomycetes tested but exhibited a patchy distribution in ascomycetes, and might be related to specific genome defense mechanisms that control transposon proliferation. This finding reveals the broader impact of transposable elements in fungi. In addition to their importance as long-term evolutionary forces, they play major roles in the more dynamic transcriptome regulation of certain species.
| Transposable elements (TEs) are mobile genetic units that colonize prokaryotic and eukaryotic genomes and generate intra- and inter-specific variability. Despite the ubiquity of TEs in the eukaryotic domain, the genome fraction occupied by these elements is highly diverse, accounting for approximately 3% in yeast genomes [1], up to 50% in mammalian genomes [2], and more than 80% in some plants, including wheat or maize [3,4]. The expansion of these elements is mediated by transposition events that can lead to their own duplication. TEs are classified into two classes based on transposition mechanisms. Class I elements transpose via RNA intermediates and include five orders (LTR, DIRS, PLE, LINE, and SINE) that are differentiated based on their structure and transposition system [5,6]. Class II encompasses elements that transpose directly from DNA to DNA. This class is divided into two subclasses. One includes the TIR and Crypton orders, and the other contains Helitrons and Mavericks [5]. The majority of transposable elements generate target site duplications at their insertion sites (TSD), which are formed as part of the insertion process. Exceptions include Helitrons [7] and the recently discovered Spy elements [8]. In addition, TE families are formed by both autonomous (coding for the proteins necessary for its transposition) and non-autonomous elements that rely on compatible transposases/retrotransposases for their mobilization.
Transposable elements can be considered selfish elements that parasitize their host genomes, and eukaryotes have developed defense mechanisms for preventing their expansion. Three mechanisms of TE silencing have been described in fungi: i) repeat-induced point mutations (RIP) [9], ii) transposon methylation [10,11], and iii) RNA-mediated gene silencing (quelling and meiotic silencing) [12,13]. Repeat-induced point mutations were originally described in Neurospora crassa and have been more recently studied in a broad range of filamentous fungi [14–16]. Transposon DNA methylation has been increasingly studied in the last few years, and recent genome-wide methylation analyses confirm the importance of this epigenetic mechanism in the control of TE proliferation in fungi [11,17,18]. Quelling and meiotic silencing occur through the detection of aberrant RNAs, which trigger RNAi pathway genes to silence. Meiotic silencing occurs when chromosomal regions are unpaired during meiosis, such as when a TE is present in one parent but not in the other. Previous studies have shown that meiotic silencing targets unpaired transposable elements [19].
Although TEs were originally considered “junk DNA”, we know today that the activity of these elements has strong consequences for genome architecture and that they are key drivers in rapid shifts in eukaryotic genome size [6,20]. Due to their repetitive nature, TEs promote chromosomal rearrangements through homologous recombination and alternative transposition [21]. TE activity can also shape genome function in multiple ways. Transposition events can lead to insertional mutations [22], which can modify or disrupt gene expression, as well as generate new proteins by exon shuffling and TE domestication [23,24]. In addition, TEs are powerful sources of regulatory sequences [25] that can be spread across the genome, rewiring pre-established networks or even creating new ones [26]. Transposable elements are associated with several classes of small RNAs that regulate the expression of multiple genes at the post-transcriptional level [27]. These reasons, among others, have transformed the originally underestimated importance of TEs into a new, exciting subject of study. This is especially relevant in fungi because international sequencing efforts are rapidly increasing the availability of genome sequences of divergent species with different lifestyles [28,29].
Fungal genomes are generally smaller than those of plants and animals, which greatly facilitates their assembly and annotation. However, the accurate annotation and quantification of transposable elements in a genome are not simple tasks, especially in draft assemblies with many scaffolds. Factors such the divergence between TE copies (due to mutations and rearrangements) or the occurrence of nested elements complicate the annotation process and necessitate the use of different algorithms to achieve reliable results [30,31]. With the rapid generation of fungal genomes, TE annotation has typically been performed using different strategies, thus limiting the ability to draw robust conclusions about the differences in TE family expansion in different species when copy differences can be ascribed to either methodological differences or biological variation. Recent comprehensive analyses of fungal TEs have described an exceptional variability in the repeat content [15,28,29], in which amplification events tend to be more related to the fungal lifestyle than to phylogenetic proximity [15,32]. LTR-retrotransposons are usually the most abundant mobile elements in fungal genomes, especially those that belong to the Gypsy and Copia superfamilies. In contrast, DNA elements generally constitute a smaller fraction of the fungal repeats, although in some species such as Fusarium oxysporum, they have undergone important amplifications in lineage-specific genomic regions [33].
In this study, we used a multi-approach pipeline for TE annotation in a collection of fungal genomes of varying phylogenetic distances and a detailed analysis of TEs in two strains of P. ostreatus. This species is a white rot basidiomycete fungus that grows on tree stumps in its natural environment. Its life cycle alternates between monokaryotic (haploid) and dikaryotic (dihaploid) mycelial phases. When two compatible monokaryotic hyphae fuse, a dikaryotic mycelium forms that is able to perform karyogamy, which occurs at the end of the life cycle, immediately before the onset of meiosis. Our results depict a P. ostreatus TE landscape dominated by Class I elements that tend to aggregate in non-homologous clusters. These clusters have profound impacts on the genome architecture at intra and inter-specific levels. In addition, we show that TE insertions modulate the global transcriptome of P. ostreatus and other fungi.
The two monokaryotic strains of P. ostreatus used in this study were sequenced by the Joint Genome Institute (JGI). PC15 was sequenced with the Sanger whole-genome shotgun approach [34], and PC9 was sequenced using Sanger whole genome shotgun and 454 paired end sequencing reads. PC15 genome assembly version 2.0 (34.3 Mb) was subjected to targeted genome improvement which led to a complete assembly of 12 scaffolds with a very low gap content (1 gap of 91 base pairs in the whole assembly) that matched the corresponding P. ostreatus chromosomes (eleven nuclear plus one mitochondrial chromosome) [35]. In contrast, PC9 assembly v1.0 (35.6 Mb) contains 572 scaffolds and a total of 476 gaps that cover 9.72% of the whole assembly.
Two monokaryotic strains of the basidiomycete P. ostreatus (PC9 and PC15) [34, 35] were used as a model to analyze differences in the occurrence and expansion of transposable element families. We identified and classified 80 TE families based on structural features and homology to previously described elements (Table 1). These families accounted for 6.2 and 2.5% of the total genome size in PC15 and PC9 genomes, respectively. In addition, we found 144 repeat-like consensus sequences that could not be reliably classified and occupied 3.6 and 2.3% of PC15 and PC9 assemblies, respectively. These elements are referred to hereafter as ‘unknown’ (S1 Table), and were not used in downstream analyses. Our integrated pipeline combined de novo predictions of LTRharvest [36] and RepeatModeler (http://www.repeatmasker.org), which were run on the two P. ostreatus genomes and merged to obtain a final TE library. This library was used then by RepeatMasker (http://www.repeatmasker.org) to detect and mask TE copies in each genome assembly. Our results showed that the merging strategy clearly outperformed the four independent approaches in terms of the number of detected families (Fig 1A). In fact, none of the TE families could be simultaneously detected by all four approaches, and very few were detected by three. In addition, up to 38 families (48% of the total) were detected by only one of the four methods. The distribution of family sizes showed that 9 of the 80 families accounted for the N50 repeat fraction in PC15 (50% of the total TE sequences), whereas 15 families accounted for the N50 repeat fraction in PC9 (Fig 1B).
The P. ostreatus repetitive element landscape was clearly dominated by Class I transposons, which accounted for 93% of the total TE content in PC15 and 89% in PC9. LTR-retrotransposons were the most abundant TE order, and were responsible for the main differences in TE content between PC15 and PC9. In fact, the four largest Gypsy families (Gypsy_1, Gypsy_2, Gypsy_3 and Gypsy_4) accounted for 2.2% of the PC15 genome size, but only 0.3% in the case of PC9. In addition, these families displayed 80 full-length copies in the former, whereas only fragments and two full-length copies were found in the latter (Table 1). A similar situation occurred with the most prominent Copia families (Copia_1 and Copia_2). Despite the important differences found between PC15 and PC9 in the number of full-length copies and the amount of LTR-retrotransposon masked sequences, the total number of detected TE fragments was closer (1,051 in PC15 vs 873 in PC9). The same was true with the amount of solo-LTRs (609 in PC15 vs 585 in PC9). Non-LTR retrotransposons (L1 elements) were found in similar abundance in PC9 and PC15, although at lower copy numbers than LTR-retrotransposons. The repertoire of Class II elements found in the genomes was dominated by the previously described Helitron families HELPO1 and HELPO2 [37]. In addition, we identified a family of Tc1-mariner transposons (TIR_1) showing putative autonomous elements as well as non-autonomous truncated copies. Autonomous elements of the latter family were present in both genomes, encoding a transposase carrying DDE3 endonuclease (pfam13358) and Tc3 transposase (cl09264) domains. Additionally, TIR_1 elements show terminal inverted repeats of 214 nt and generate a 2bp target site duplication (TA) upon insertion. Full TE annotations in PC15 and PC9 assemblies are deposited in the Supplementary Information (S1 and S2 Datasets, respectively).
Our screening of TE sequences in P. ostreatus genome assemblies uncovered that some of the most important LTR-retrotransposon families of PC15 were under-represented in PC9 (Table 1). We hypothesized that our estimation of TE content in PC9 could be underestimated in comparison to PC15 due to its lower assembling quality. In order to know whether this TE families were present in the genome but couldn’t be properly assembled, we analyzed the TE content of PC9 clean 454 sequencing reads (read length of 80 to 626 nt, median length of 364 nt). Datasets of 1.58x and 1.76x genome coverages were randomly sampled from two sequenced libraries, and repeat-masked using our curated TE library to provide an unbiased estimation of TE content. The analysis yielded an average TE content of 4.98%, being the amount of sequence masked by each TE family highly correlated between the two datasets (R2 = 0.98, S3 Dataset). In addition, the results showed that Gypsy_1, Gypsy_2 and Gypsy_3 LTR-retrotransposon families were the most abundant in PC9 genome, similarly to that found in the fully assembled PC15 strain.
The density of TEs in P. ostreatus was highly variable among the twelve chromosomes and regionally within each chromosome (Fig 2). TEs were not randomly distributed over the genome (Mann-Whitney-Wilcoxon p = 2.2e-16), and overlapped frequently with annotated genes (502 in PC15 and 339 in PC9, hereafter referred as “TE-associated genes”). The results of a hypergeometric test performed on the fully assembled PC15 strain revealed that 58% of the TEs were arranged in retrotransposon-rich clusters showing poor sequence conservation between the two genomes. A total of 2,108 genes out of 12,330 were present in these repeat-rich regions. Of these genes, 70 were annotated as lignocellulose-degrading enzymes such CAZymes, manganese and versatile peroxidases, although their presence in TE clusters was not over-represented in comparison to the whole genome (Fisher p value = 0.52). At an inter-specific level, the impact of TE insertions was even more striking, as the conservation of these transposon-enriched regions drops dramatically compared with other basidiomycetes (S1 Fig).
A whole genome alignment between PC15 and PC9 was performed to detect in silico polymorphic TE insertions. The alignment of every TE locus was extracted and parsed to detect the allelic state (genotype) based on the alignability of such regions. We used the same pipeline to analyze the allelic state of 11,630 protein-coding genes. While only 7.7% of the protein coding genes were heterozygous alleles, up to 50% of TE insertions were polymorphic. Bioinformatics predictions were validated by PCR in a subset of eight polymorphic insertions (Fig 3).
The insertion ages of all intact LTR-retrotransposons (carrying both Long Terminal Repeats, n = 189) were estimated based on the nucleotide divergence of LTRs using the approach described in [38] and the fungal substitution rate of 1.05 × 10−9 nucleotides per site per year [39,40]. Our results showed that 33% of the LTR-retrotransposon insertions occurred during a recent amplification burst (0 My), and up to 64% were amplified during the last 5 My (Fig 4). The oldest PC15 LTR-retrotransposon insertion clocked 41 My ago, while the oldest element in PC9 clocked 12 My ago. The phylogenetic reconstruction of the LTR-retrotransposon families revealed that some of the most prominent and recently amplified Gypsy families (Gypsy_1, Gypsy_2, Gypsy_5 and Gypsy_6) were phylogenetically close (S2 Fig).
We obtained the average expression of every TE family normalized per family size using RNA-seq (Fig 5). Among the main TE groups, LINE was the most abundantly expressed in both strains, followed by Helitrons (especially the HELPO1 family) in PC15 and Gypsy retrotransposons in PC9. At the family level, 60% were expressed in PC15 and 59% in PC9, while at the copy level only 14% and 17% showed transcription, respectively. In addition, 16 out of the 80 families were transcriptionally silent in both strains. Notably, the three strain-specific families in P. ostreatus (Copia_17, DIRS_4 and Gypsy_53, present only in PC9) were transcriptionally active.
To investigate the impact of TEs on the functional genome of P. ostreatus, we explored the effect of TEs on the expression of the surrounding genes. The closest TE insertion to each gene was identified in the three following scenarios (TE-associated genes were excluded from the analysis): i) a TE was present in a 1kb window upstream of the gene start codon, ii) a TE was present in a 1 kb window downstream of the gene end, and iii) a TE was present in both upstream and downstream regions in a window of 1 kb (gene “captured” between two TEs). This window size was selected based on the small intergenic distance of P. ostreatus (1.14 Kb). When we analyzed the gene expression distribution in every scenario, significant differences were uncovered between controls and genes under TE influence (Fig 6A and 6B). In particular, a strong repression was found for genes captured between two TEs (scenario III), while a discontinuous repression was found when the TE was present upstream or downstream of the gene body (scenarios I and II). In the latter case, distribution shapes indicate that approximately half of the genes were repressed and the other half remained unaltered.
To investigate whether this silencing effect could be influenced by the TE distribution along the chromosomes, we split the analysis of the PC15 strain in two additional scenarios: i) the gene under TE influence was located inside a significant TE cluster (Fig 6C) and ii) the gene under TE influence was located outside a significant TE cluster (isolated TE) (Fig 6D). The results showed that the impact of TEs on gene expression was more intense when insertions occurred inside TE clusters. Additionally, significant differences were found between the distribution of gene expression of genes inside clusters that were not under the influence of TEs (control plot, Fig 6C) and that of the genes in the same condition but outside TE clusters (control plot, Fig 6D, p = 1.22e-8).
To corroborate the hypothesis of TE-mediated gene repression we studied the transcription of orthologous genes displaying polymorphic insertions (always in a window size of 1 Kb), where a TE was present in PC15 and absent in PC9 and vice versa. Tables 2 and 3 show 21 genes that were inactive under TE influence and active in the orthologous, TE-free allele. Gypsy LTR-retrotransposons were the main TEs involved in the repression with only two exceptions, which involved the Copia_5 (LTR-retrotransposon) and HELPO1 (Helitron) families. The inactivated genes displayed a broad range of functions. Additional orthologous pairs showing strong repression in the allele under TE influence (5 fold) are shown in S2 Table.
Our pipeline for the identification, classification and annotation of transposable elements was performed in eighteen Ascomycetes and Basidiomycetes genomes (Fig 7). The results demonstrated great variability in TE content at the phylum, genus and species levels (Fig 7, S3 Table). Elements belonging to 20 different TE superfamilies (11 of Class I and 9 of Class II) were identified and classified into the main groups shown in Fig 7. The genome percentage occupied by these TE families showed a positive correlation with genome size (R2 = 0.38). Within the genera analyzed, Serpula showed a surprisingly high TE content in proportion to its genome size, especially due to LTR-retrotransposon expansions in the Gypsy and Copia superfamilies. In fact, when excluding the two Serpula genomes from the analysis, the correlation between TE content and genome size in the remaining species was much higher (R2 = 0.71). The Ascomycete species analyzed had a ratio of Class I / Class II elements ranging from 0.78 to 4.23 and a low content of repetitive sequences, with the exception of the plant pathogen F. oxysporum. Interestingly, this species showed a 15-fold enrichment of transposable elements compared with F. graminearum as a result of important expansions of Class II elements (Tc1-mariner and hAT families). The variability in the TE content in the analyzed Basidiomycetes ranged from species practically free of TE repeats, such as in the Pseudozyma genera (0.02% of the genome), to species with almost one third of their genome masked by the TE library, such as Serpula lacrymans or Puccinia graminis. TE expansions seemed to be constrained in basidiomycete yeasts such Pseudozyma or Mixia compared to the rest of the basidiomycetes analyzed. LTR-retrotransposons in the Gypsy and Copia superfamilies families were the main elements responsible for differences in TE content, with the Class I / Class II ratio much higher in basidiomycetes than in ascomycetes (9.3 in average). In fact, these two superfamilies were detected in all species analyzed in this study. When we studied the differential TE amplifications at the genus/species level, we found six pairs that displayed similar content (Botrytis, Cryptococcus, Phanerochaete, Serpula, Pleurotus and Pseudozyma) and two pairs (Fusarium and Puccinia) that showed important differences between counterparts.
The effect of TE insertions in nearby genes was analyzed in four additional fungal models: Laccaria bicolor, Fusarium graminearum, Botrytis cinerea B05.10 and Saccharomyces cerevisiae S288C. These species were chosen based on the public availability of genomic (full genome sequence) and transcriptomic (RNA-seq) data. In addition, L. bicolor and S. cerevisiae were chosen based on their opposite methylation patterns (evidence of methylation vs absence of methylation, respectively [11]). The analysis uncovered two clear profiles. First, L. bicolor and F. graminearum showed a pattern of TE-mediated repression similar to P. ostreatus, in which an important number of genes carrying TE insertions within a 1 kb upstream/downstream window were repressed (Fig 8). Second, B. cinerea and S. cerevisiae genes under TE influence did not show any alteration in expression, with distributions identical to the control (p > 0.05, Fig 8)
During the process of TE classification using BLASTX against Repbase peptide database we noticed high similarity between the P. ostreatus TIR_1 family and the previously described Mariner2_PPa [41] (71% nucleotide identity over 71% of the sequence), a Tc1-mariner element identified in the moss Physcomitrella patens. According to the nucleotide divergence estimated by K2P distance and the fungal nucleotide substitution rate, TIR_1 and Mariner2_PPA diverged 517 My ago, despite mosses and fungi diverged about 1,600 My ago [42]. To investigate if horizontal transfers could have played a role in the distribution of fungal and other eukaryotic Tc1-mariners, we reconstructed the phylogeny of their encoded transposases (Fig 9). Our dataset included fungal, animal, plant and bacterial Tc1-mariner transposases, which were obtained based on best BLAST hits against NCBI and JGI reference proteins databases. The topology of the gene tree shows clear incompatibilities with the phylogenetic relationships of the species analyzed, which might be explained by horizontal transfers of Tc1-mariners. Specifically, basidiomycete and animal transposases were placed in a single clade with very high support, separated from ascomycete transposases. Other phylogenetic incongruences were the presence of the moss Physcomitrella patens and the mucoral Rhizopus oryzae in the basidiomycete clade, as well as the endosymbiont bacteria Wolbachia present in the animal clade.
Fungal TE content is highly diverse, even within species that are phylogenetically close [28]. However, studies analyzing the intra-specific variability in TE content have been infrequent. According to our results, transposable elements accounted for a small to moderate amount of the genome size in the two P. ostreatus strains analyzed (6.2% in PC15 and 2.5–4.9% in PC9). Although the number of TEs detected varies according to the pipeline used, the TE content in P. ostreatus fell within the range reported for most fungal genomes (from 0 to 25%) [15,28,43,44,45], with the exception of some plant pathogens and ectomycorrhizal species that have undergone massive TE amplifications [32,44]. Despite all TE groups are generally more abundant in PC15 than in PC9, major differences between the strains were observed in LTR-retrotransposons. Most of the LTR-retrotransposon families under-represented in PC9 were actually present in the genome, but could not be assembled into the main scaffolds due to its length and repetitive nature. Assembling transposable elements is technically challenging because identical TE copies require sequencing reads exceeding the TE length to be resolved [46]. This is especially relevant in P. ostreatus, as we show that most of its LTR-retrotransposons underwent a recent amplification burst, thus sharing high nucleotide similarity. The presence of TE sequences in the unassembled reads is common in plants and animals [47,48]. In fungi, a recent study performed on several Amanita species identified many TEs that could not be found in the assembled regions, especially Gypsy elements [32]. In addition to the difficulty in assembling TE repeats, their structural complexity, which is caused by internal rearrangements, mutations, nested elements and DNA fragment acquisition events, complicated their identification using generic annotation tools. Our multi-way approach used for TE detection greatly improved the discovery of repeats, as revealed by the number of detected families in our combined TE library (Fig 1A). Using this approach was of particular importance for TE detection in PC9, because families that could not be detected by de novo searches in the assembly due to its high gap content could be found in PC15 and thus were present in the TE library.
P. ostreatus repeat content is enriched in Class I transposons, especially in the Gypsy and Copia superfamilies. LTR-retrotransposons are divided into five superfamilies, but these two are the most abundant in the fungal kingdom [28,49]. The replicative transposition mechanism of autonomous LTR-retrotransposons makes them efficient genome colonizers because the copy number increases with every transposition event. Autonomous LTR-retrotransposons contain gag and pol genes flanked by long terminal repeats, and they differ from retroviruses in that they do not have infection capacity [50]. The difference between the Gypsy and Copia superfamilies lies in the order of the internal protease, integrase, reverse transcriptase and RNAse H domains present in the pol gene. We also found retrotransposons of the DIRS superfamily, which contains a gag, pol and tyrosine recombinase ORFs flanked by terminal repeats. This group of TEs is less abundant than other retrotransposons, and it exhibited patchy distribution in the fungal phylogeny [51].
One necessary condition for an active TE family is the presence in the genome of autonomous elements encoding the structural features and protein domains necessary for their own transposition. In this sense, the Gypsy architecture seems to be the most successful, as shown by the number of families and number of full-length copies per family. A second condition for TE transposition is that autonomous elements must be transcribed. We showed that although most genomic regions containing TEs are silenced, about 60% of the TE families showed at least one transcriptionally active copy. Interestingly, Class I transposons show high transcriptional levels, which are essential because they are propagated through RNA intermediates that can be translated into proteins necessary for replication or can act as replication templates. In parallel to the successful amplification of LTR-retrotransposons in P. ostreatus, the presence of solo-LTRs suggests the occurrence of homologous recombination between LTRs leading to retrotransposons elimination. Class II DNA transposons are less abundant than Class I RNA elements and are represented by the Helitron and Tc1-mariner superfamilies. In a previous work, we reported the presence and structure of the two Helitron families in P. ostreatus [37]. Helitrons were discovered by bioinformatics approaches in Arabidopsis thaliana and Caenorhabditis elegans more than a decade ago [7]. Nevertheless, the experimental demonstration of their transposition was not described until very recently [52]. Their rolling-circle transposition mechanism and their ability to capture and amplify gene fragments make them interesting subjects of study. Helitrons are present in all eukaryotic kingdoms [53], although they show patchy distribution in some phylogenetic clades, such as mammals. In plants, they play an important role in genome evolution, introducing functional diversity by creating new genes and isoforms [54]. In this study, we showed that Helitrons are the most abundant DNA transposons in the P. ostreatus genome and are the second superfamily in transcriptional activity. Our results add a piece of evidence to the fact that this superfamily is actively populating the P. ostreatus genome. Interestingly, within the 19 described superfamilies of cut and paste DNA transposons, only Tc1-mariner is present in P. ostreatus. According to our results, this superfamily would be the most efficient fungal cut and paste transposon, as it is the most represented in the species analyzed. Nevertheless, most of the copies present in P. ostreatus are truncated, and the putative autonomous elements encoding transposases are not expressed in the condition tested. Our phylogenetic reconstruction of TIR_1-like Tc1-mariner transposases shows important discordances with organismal phylogenies, suggesting that horizontal transfer has shaped the distribution of these Class II transposons within the eukaryotic kingdom. Specifically, the presence of animal, plant, bacterial, mucoral and basidiomycete transposases in a monophyletic group separated from ascomycetes supports the hypothesis that multiple horizontal transfers occurred after the divergence of basidiomycetes and ascomycetes, event that took place about 1200 My ago [42]. It is known that transposable elements are horizontally transferred in eukaryotes at a higher frequency than regular genes [55], and this ability allows them to persist in the course of evolution escaping from vertical extinction [56]. Our data suggests that horizontal gene transfer has played an important role in the dynamics of eukaryotic Tc1-mariners. Nevertheless, the diversity of TE copies, their repetitive nature and the limitations of the taxonomic sampling make difficult to reconstruct the full evolutionary history of TIR_1-like Tc1-mariner transposases.
Most fungal species have streamlined, compact genomes. Owing to international efforts and advances in genome sequencing over the last decade, there is genomic information for nearly 500 fungal species covering most of the fungal phylogenetic diversity, with more being produced (http://1000.fungalgenomes.org). The assembled genome sizes in fungi range from about 2 to 190 Mb, while flow cytometry estimations have uncovered genome sizes of up to 893 Mb in the Pucciniomycotina subphylum [57] (Gymnosporangium confusum). The available data demonstrate the impressive variability in fungal genome size, and our results suggest that an important part of this variability could be explained by differential expansions of TEs that seem to be related to the fungal lifestyle. Our results confirm that obligate biotrophs such P. graminis and P. striiformis are highly enriched in TEs [45]. By contrast, the (not obligate) biotroph M. osmundae is practically free of TEs, similarly to other basidiomycete yeasts such the P. hubeiensis and P. antarctica. Previous studies have shown that TE-driven expansions have played important roles in the genomes of filamentous plant pathogens [58]. An example of the impact of TEs in host adaptation and pathogen aggressiveness is the Leptosphaeria genus [59]. According to [58], faster adaptation occurs because genes encoding proteins for host interactions are frequently polymorphic and reside within repeat-rich regions of the genome. Due to the presence of P. ostreatus lignin degrading enzymes within TE clusters, is tempting to hypothesize that TEs could play an important role in the evolution of wood decayers.
Transposable elements are undoubtedly an important source of genetic variation in fungi. As previously found in other fungal species [43], P. ostreatus TEs are preferentially arranged in non-homologous genomic regions that display low conservation at both the intraspecific and interspecific levels. These genomic blocks are hotspots for LTR-retrotransposon accumulation, which could target these regions due to specific chromatin structures adopted by pre-existing elements [60].
The compatible monokaryotic strains PC9 and PC15 can mate to form a dikaryon, the nuclei of which coexist in the same cell [35]. Thus, the unpaired long blocks of repetitive DNA are unlikely to undergo crossover and are likely inherited as supergenes after meiosis. We show that the transcription of these TE-rich regions tended to be strongly repressed (Figs 2 and 6) and we hypothesize that genes with essential functions might eventually be captured and silenced during the formation of these TE clusters, leading to a looseness of fit by the monokaryotic genotypes carrying these genomic regions. Selection against these TE blocks would lead to the loss of these alleles in the course of evolution. On the other hand, the higher plasticity of these repeat regions might create novel opportunities for diversification and adaptation. In addition to the permanent genomic modifications that TEs can promote, we showed that both isolated and clustered TE insertions modulate the expression of surrounding genes. In addition to the disruption-mediated changes originated by TE insertions into promoter regions, there are additional mechanisms by which TEs can alter the expression of surrounding genes. TEs often carry cis-regulatory elements that can be spread over the genome [26]. Similarly, LTR-retrotransposons and solo-LTRs contain promoters that can activate the expression of dormant genes [60]. Additionally, transcripts from full-length TEs can read through into a neighbor gene, producing spurious transcripts that can be subjected to transcriptional and post-transcriptional control [61]. Finally, TEs can be targeted for heterochromatin formation, thus potentially silencing the transcription of the adjacent gene [26]. Several studies have shown that Arabidopsis genes close to TEs had lower expression than the average genome-wide expression [62, 63]. Similarly, a recent study showed that the insertion of SINE retrotransposons close to human and mouse gene promoters led to transcriptional silencing mediated by the acquisition of DNA methylation [64]. The few studies available on the subject in fungi indicate that methylation targets transposon sequences selectively, leading to TE transcriptional silencing [11,17,18]. Although methylation within fungal genes tends to be low, studies in the plant pathogen Magnaporthe oryzae showed that genes that were methylated in upstream or downstream regions resulted in lower transcription than un-methylated genes [17]. We hypothesize that the transcriptional repression of genes surrounded by TE insertions could be related to the epigenetic status of the given TE. In fact, the discontinuous repression found in P. ostreatus genes under TE influence (gene repressed vs non-repressed) fits with the putative methylated vs non-methylated status of the involved TEs. Although we lack experimental evidence of methylation in PC15 or PC9, the presence in both strains of transcriptionally active homologs of the Dim-2 DMTase (S3 Fig) responsible for cytosine methylation in fungi [65] suggests that the methylation machinery is active in P. ostreatus. In addition to P. ostreatus, we used the same transcriptional analysis pipeline in two species with well-known methylation profiles [11]: S. cerevisiae (methylation-free) and L. bicolor (TE regions highly methylated). The expression distribution of S. cerevisiae genes under TE influence was identical to the control (p < 0.05), while the distribution in L. bicolor showed a severe bias towards low expressed genes. Additional analyses performed in other species uncovered that the ascomycetes F. graminearum and B. cinerea showed different expression patterns for genes under TE influence. Whereas B. cinerea genes remained unaltered, the expression in F. graminearum genes was lower than the control. Bisulfite sequencing of Gibberella zeae (anamorph: F. graminearum) showed that this species has low cytosine methylation levels, although it displays related mechanisms of TE silencing, such as RIP and meiotic silencing [66]. Regarding B. cinerea, the unique reference found on the subject showed that no or very little methylation occurred in this species, according to HpaII/MspI restriction patterns [67]. In summary, we show that transposable element dynamics differentially impact fungal genome-wide transcription patterns, likely as a result of the epigenetic machinery evolved to control TE proliferation.
Eighteen Ascomycetes and Basidiomycetes species were selected in this study as sample sets of closely related species for genomes comparisons. Publicly available genomic assemblies were downloaded from the Joint Genome Institute’s fungal genome portal MycoCosm [68] (http://jgi.doe.gov/fungi), the Broad Institute (https://www.broadinstitute.org/) and FungiDB [69]. The genome sequences of the P. ostreatus monokaryotic strains PC15 v2.0 [34] and PC9 v1.0, which were obtained by de-dikaryotization of the dikaryotic strain N001 [35], were used as models for building the pipelines described in this paper.
De novo identification of repetitive sequences in the genome assemblies was performed by running the RECON [70] and RepeatScout [71] programs (integrated into the RepeatModeler pipeline). LTRharvest [36] was used to improve the detection of full length LTR-retrotransposons. LTRharvest results were filtered to avoid false positives as follows: elements were de-duplicated and used as queries for BLASTN searches (cutoff E-value = 10−15) against the genome assembly and for BLASTX (cutoff E-value = 10−5) against the Repbase peptide database [54]. Only sequences longer than 400 bp with more than five copies or yielding a significant hit to a described LTR-retrotransposon were kept for further analysis. The outputs of the above programs were merged and clustered at 80% similarity using USEARCH [72] to create species-specific (i.e., P. ostreatus PC15 and PC9) or genus-specific (i.e., F. oxisporum and F. graminearum) TE libraries. Each consensus sequences library was classified using BLASTX against the Repbase peptide database, and the final libraries were used as input for RepeatMasker (http://www.repeatmasker.org). Consensus sequences without similarity to any Repbase entry were labeled as ‘unknown’. The RepeatMasker output was parsed using the One_code_to_find_them_all script [73] to reconstruct TE fragments into full-length copies and estimate the fraction of the genome occupied by each TE family.
To identify solo-LTRs, the left terminal repeat of every autonomous copy was extracted, and a BLASTN against each assembly was performed. The flanking sequences of every hit (5,000 bp, cutoff E-value = 10−15) were extracted and screened for retrotransposon internal sequences. Solo-LTRs were defined as those hits lacking internal retrotransposon sequences at the flanking sites.
To determine whether TEs were non-randomly distributed, the distribution of inter-TE distances was compared (Mann-Whitney-Wilcoxon text) with that of the inter-element distances of a randomly generated subset of 1,196 elements. In addition, TEs and gene model annotations were merged and used as reference for a hypergeometric test to test for the presence of regions enriched in TEs. The analysis was performed using REEF [74] with a Q-value of 0.05 (FDR 5%), a window width of 100 kb with a shift of 10 kb and a minimum number of 10 features in clusters.
The P. ostreatus PC15 and PC9 genome assemblies were aligned using the Mercator and MAVID pipeline [75], using the fully assembled PC15 genome as a reference. Gene model positions and TE hits of the PC15 strain were used to extract individual alignments and to check the homozygous vs. heterozygous nature of the insertions. A locus was considered homozygous if the alignment spanned at least 80% of the whole locus length, and heterozygous when the PC9 allele was absent.
Long Terminal Repeats of every intact, full-length element were extracted and aligned. Kimura 2-Parameter distance was obtained using a Python script and transformed to My using the approach described in [39] and the fungal substitution rate of 1.05 × 10−9 nucleotides per site per year [40].
Mycelia were harvested, frozen and ground in a sterile mortar in the presence of liquid nitrogen. DNA was extracted using a Fungal DNA Mini Kit (Omega Bio-Tek, Norcross, GA, USA). Sample concentrations were measured using a Qubit 2.0 Fluorometer (Life Technologies, Madrid, Spain), and purity was measured using a NanoDrop 2000 (Thermo-Scientific, Wilmington, DE, USA). PCR reactions were performed according to Sambrook et al [76] using primers designed after the TE flanking sequences (S1 Text, Supplementary Information). Total RNA was extracted from 200 mg of deep frozen tissue using Fungal RNA E.Z.N.A Kit (Omega Bio-Tek, Norcross, GA, USA), and its integrity was estimated by denaturing electrophoresis on 1% (w/v) agarose gels. Nucleic acid concentrations were measured using a Nanodrop 2000 (Thermo Scientific, Wilmington, DE, USA), and the purity of the total RNA was estimated by the 260/280 nm absorbance ratio. Messenger RNA was purified using a MicroPoly(A) Purist kit (Ambion, USA). Transcriptome libraries were generated and sequenced by Sistemas Genomicos S.L. (Valencia, Spain) on a SOLiD platform, following the manufacturers’ recommendations (Life Technologies, CA, USA).
P. ostreatus RNA-seq datasets corresponding to PC15 and PC9 strains (8.4 and 9.7 million reads in PC15 and PC9, respectively) cultured in SMY medium and harvested during the exponential growth phase, were used to analyze the transcription of genes and TEs. The quality of the SOLiD RNA-seq reads was verified using FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/), and they were mapped to their corresponding PC15 v2.0 or PC9 v1.0 assemblies using TopHat [77], restricting the multihits option to 1. HTseq-count [78] was used to determine the number of reads mapping to every feature. SAMtools [79], BEDTools [80] and custom Python scripts were used to manipulate the data, to calculate RPKMs and to obtain genome coverages. Public RNA-seq data from other species were downloaded from the NCBI SRA database and were analyzed using the same pipeline (accessions SRR1257938 Saccharomyces cerevisiae S288C [81], SRR1284049 Botrytis cinerea B05.10 [82], SRR1592424 F. graminearum [83] and SRR1165053 Laccaria bicolor [84]).
For analyzing the expression of TE families, reads were mapped to the extracted transposon sequences using Bowtie [85] and allowing multi-mapping. RSEM software was used to calculate TE expression because its algorithm is especially designed to handle multi-mapped reads [86]. Afterwards, the FPKMs of each family were normalized to the number of elements.
Gene and TE annotations were intersected to obtain TE-associated genes (genes overlapping with any TE) and non-TE genes (genes not overlapping with any TE). Afterwards, the closest TE upstream and downstream to each non-TE gene was obtained at a maximum distance of 1 kb. The resulting genes were organized in three groups: i) genes with an upstream TE, ii) genes with a downstream TE and iii) genes with both upstream and downstream TEs. Control groups were obtained by subtracting target genes (three previous scenarios) to all the non-TE genes.
The predicted proteomes of all species were downloaded from the Mycocosm database (http://genome.jgi.doe.gov/programs/fungi/index.jsf). After all-by-all BLASTP, proteins were clustered with MCL [87] using an inflation value of 2. Clusters containing single copy genes of each genome were retrieved (allowing two missing taxa per cluster) and proteins were aligned with MAFFT [88]. The alignments were concatenated after discarding poorly aligned positions with Gblocks [89]. Maximum-likelihood phylogeny was constructed using RaxML [90] under PROTGAMMAWAGF substitution model and 100 rapid bootstraps.
Using the P. ostreatus JGI browser we identified the internal transposase gene of a full length element of TIR_1 family. This protein was used as query for BLASTP searches (cutoff = E-5) against NCBI RefSeq protein database (independent searches were carried out against animal, plant and bacterial databases). The best five animal, plant and bacterial hits were retrieved when possible (i.e. only one hit was obtained using plant database). The same search was performed in the JGI database to retrieve the best five basidiomycete hits, and the best five non-basidiomycete hits. Proteins were aligned with MUSCLE [91], and the alignments were trimmed using trimAl [92] with the default parameters. An approximate maximum likelihood tree was constructed using FastTree [93] and edited with Figtree (http://tree.bio.ed.ac.uk/software/figtree/). Transposases from P. patens, Wolbachia and Rhizopus oryzae were further analyzed to exclude the possibility of being a result of database contamination: Using TBLASTN against NCBI Whole-genome shotgun contigs or JGI genomic scaffolds, we identified their genomic position and verified that they were assembled in long scaffolds and surrounded by other host genes.
Raw sequencing data was deposited in GEO database under the accession number GSE81586.
|
10.1371/journal.pgen.1007254 | Large scale variation in the rate of germ-line de novo mutation, base composition, divergence and diversity in humans | It has long been suspected that the rate of mutation varies across the human genome at a large scale based on the divergence between humans and other species. However, it is now possible to directly investigate this question using the large number of de novo mutations (DNMs) that have been discovered in humans through the sequencing of trios. We investigate a number of questions pertaining to the distribution of mutations using more than 130,000 DNMs from three large datasets. We demonstrate that the amount and pattern of variation differs between datasets at the 1MB and 100KB scales probably as a consequence of differences in sequencing technology and processing. In particular, datasets show different patterns of correlation to genomic variables such as replication time. Never-the-less there are many commonalities between datasets, which likely represent true patterns. We show that there is variation in the mutation rate at the 100KB, 1MB and 10MB scale that cannot be explained by variation at smaller scales, however the level of this variation is modest at large scales–at the 1MB scale we infer that ~90% of regions have a mutation rate within 50% of the mean. Different types of mutation show similar levels of variation and appear to vary in concert which suggests the pattern of mutation is relatively constant across the genome. We demonstrate that variation in the mutation rate does not generate large-scale variation in GC-content, and hence that mutation bias does not maintain the isochore structure of the human genome. We find that genomic features explain less than 40% of the explainable variance in the rate of DNM. As expected the rate of divergence between species is correlated to the rate of DNM. However, the correlations are weaker than expected if all the variation in divergence was due to variation in the mutation rate. We provide evidence that this is due the effect of biased gene conversion on the probability that a mutation will become fixed. In contrast to divergence, we find that most of the variation in diversity can be explained by variation in the mutation rate. Finally, we show that the correlation between divergence and DNM density declines as increasingly divergent species are considered.
| Using a dataset of more than 130,000 de novo mutations we show that there is large-scale variation in the mutation rate at the 100KB and 1MB scales. We show that different types of mutation vary in concert and in a manner that is not expected to generate variation in base composition; hence mutation bias is not responsible for the large-scale variation in base composition that is observed across human chromosomes. As expected, large-scale variation in the rate of divergence between species and the variation within species across the genome, are correlated to the rate of mutation, but the correlation between divergence and the mutation rate is not as strong as it could be. We show that biased gene conversion is responsible for weakening the correlation. In contrast, we find that most of the variation across the genome in diversity can be explained by variation in the mutation rate. Finally, we show that the correlation between the rate of mutation in humans and the divergence between humans and other species, weakens as the species become more divergent.
| Until recently, the distribution of germ-line mutations across the genome was studied using patterns of nucleotide substitution between species in putatively neutral sequences (see [1] for review of this literature), since under neutrality the rate of substitution should be equal to the mutation rate. However, the sequencing of hundreds of individuals and their parents has led to the discovery of thousands of germ-line de novo mutations (DNMs) in humans [2–6]; it is therefore possible to analyse the pattern of DNMs directly rather than inferring their patterns from substitutions. Initial analyses have shown that the rate of germ-line DNM increases with paternal age [4], a result that was never-the-less inferred by Haldane some 70 years ago [7], maternal age [6], varies across the genome [5] and is correlated to a number of factors, including the time of replication [3], the rate of recombination [3], GC content [5] and DNA hypersensitivity [5].
Previous analyses have demonstrated that there is large scale (e.g. 1MB) variation in the rate of DNM in both the germ-line [3, 5] and the somatic tissue [8–12]. Here we focus exclusively on germ-line mutations. We use a collection of over 130,000 germ-line DNMs to address a range of questions pertaining to the large-scale distribution of DNMs. First, we quantify how much variation there is at different scales and investigate whether the variation in the mutation rate at a large-scale can be explained in terms of variation at smaller scales. We also investigate to what extent the variation is correlated between different types of mutation, and to what extent it is correlated to a range of genomic variables.
We use the data to investigate a long-standing question–what forces are responsible for the large-scale variation in GC content across the human genome, the so called “isochore” structure [13]. It has been suggested that the variation could be due to mutation bias [14–18], natural selection [13, 19, 20], biased gene conversion [21–24], or a combination of all three forces [25]. There is now convincing evidence that biased gene conversion plays a role in the generating at least some of the variation in GC-content [26–28]. However, this does not preclude a role for mutation bias or selection. With a dataset of DNMs we are able to directly test whether mutation bias causes variation in GC-content.
The rate of divergence between species is known to vary across the genome at a large scale [1]. As expected this appears to be in part due to variation in the rate of mutation [3]. However, the rate of mutation at the MB scale is not as strongly correlated to the rate of nucleotide substitution between species as it could be if all the variation in divergence between 1MB windows was due to variation in the mutation rate [3]. Instead, the rate of divergence appears to correlate independently to the rate of recombination. This might be due to one, or a combination, of several factors. First, recombination might affect the probability that a mutation becomes fixed by the process of biased gene conversion (BGC) (reviewed by [26]). Second, recombination can affect the probability that a mutation will be fixed by natural selection; in regions of high recombination deleterious mutations are less likely to be fixed, whereas advantageous mutations are more likely. Third, low levels of recombination can increase the effects of genetic hitch-hiking and background selection, both of which can reduce the diversity in the human-chimp ancestor, and the time to coalescence and the divergence between species. There is evidence of this effect in the divergence of humans and chimpanzees, because the divergence between these two species is lower nearer exons and other functional elements [29, 30]. And fourth, the correlation of divergence to both recombination and DNM density might simply be due to limitations in multiple regression; spurious associations can arise if multiple regression is performed on two correlated variables that are subject to sampling error. For example, it might be that divergence only depends on the mutation rate, but that the mutation rate is partially dependent on the rate of recombination. In a multiple regression, divergence might come out as being correlated to both DNM density and the recombination rate, because we do not know the mutation rate without error, since we only have limited number of DNMs. Here, we introduce a test that can resolve between these explanations.
As with divergence, we might expect variation in the level of diversity across a genome to correlate to the mutation rate. The role of the mutation rate variation in determining the level of genetic diversity across the genome has long been a subject of debate. It was noted many years ago that diversity varies across the human genome at a large scale and that this variation is correlated to the rate of recombination [31–33]. Because the rate of substitution between species is also correlated to the rate of recombination, Hellmann et al. [31, 32] inferred that the correlation between diversity and recombination was at least in part due to a mutagenic effect of recombination, an inference that has been confirmed by recent studies of recombination [3, 34, 35]. However, no investigation has been made as to whether variation in the rate of mutation explains all the variation in diversity, or whether biased gene conversion, direct and linked selection have a major influence on diversity at a large scale.
To investigate large scale patterns of de novo mutation in humans we compiled data from three studies which between them had discovered more than 130,000 autosomal DNMs: 105,385 from Jonsson et al. [36], 26,939 mutations from Wong et al. [6], and 11016 mutations from Francioli et al. [3] The datasets are henceforth referred to by the name of the first author. We divided the mutations up into 9 categories reflecting the fact that CpG dinucleotides have higher mutation rates than non-CpG sites, and the fact that we cannot differentiate which strand the mutation had occurred on: CpG C>T (a C to T or G to A mutation at a CpG site), CpG C>A, CpG C>G and for non-CpG sites C>T, T>C, C>A, T>G, C<>G and T<>A mutations.
The proportion of mutations in each category in each of the datasets is shown in Fig 1. We find that the pattern of mutation differs significantly between the studies (Chi-square test of independence on the number of mutations in each of the 9 categories, p < 0.0001). This appears to be largely due to the relative frequency of C>T transitions in both the CpG and non-CpG context; a discrepancy which has been noted before[37, 38]. In the data from Wong et al. [6] the frequency of C>T transitions at CpG sites is ~13% whereas it is ~16–17% in the other two datasets. For non-CpG sites the frequency of C>T transitions is ~24% in all studies except that of Wong et al. in which it is 26%. It is not clear whether these patterns reflect differences in the mutation rate between different cohorts of individuals, possibly because of age [3, 4, 6] or geographical origin [39] or whether the differences are due to methodological problems associated with detecting DNMs.
To investigate whether there is large scale variation in the mutation rate we divided the genome into non-overlapping windows of 10KB, 100KB, 1MB and 10MB and fit a gamma distribution to the number of mutations per region, taking into account the sampling error associated with the low number of mutations per region. We focussed our analysis at the 1MB scale since this has been extensively studied before. However, we show that the variation at 1MB forms part of a continuum of variation. We also repeated almost all our analyses at the 100KB scale with qualitatively similar results (these results are reported in supplementary tables).
We find that the amount of variation differs significantly between the three studies (likelihood ratio tests: p < 0.001), although, the differences are quantitatively small at the 1MB (Fig 2) and 100KB (S1 Fig) scales. The variation between datasets might be due to differences in age or ethnicity between the individuals in each study, or methodological problems–for example, there might be differences between studies in the ability to identify DNMs. We can test whether callability is an issue in the Wong dataset because Wong et al. [6] estimated the number of trios at which a DNM was callable at each site. If we reanalyse the Wong data using the sum of the callable trios per MB, rather than the number of sites in the human genome assembly, we obtain very similar estimates of the distribution: the coefficient of variation (CV) for the distribution is 0.27 when we use the number of sites and 0.24 when we use the sum of callable trios.
As expected the number of DNMs per site is significantly correlated between the datasets (1MB Francioli v Wong r = 0.15, p<0.001; Francioli v Jonsson r = 0.19 p<0.001; Wong v Jonsson r = 0.29, p<0.001). The correlation is weak, but this is likely to be in part due to sampling error. If we simulate data assuming a common distribution, estimating the shape parameter as the mean CV of the distributions fit to the individual datasets, the mean simulated correlations are: Francioli v Wong r = 0.20; Francioli v Jonsson r = 0.29; Wong v Jonsson r = 0.41. This suggests that a substantial proportion of the variation is common to the three datasets, however in each case less than 5% of the simulated correlations are less than the observed correlation suggesting that some portion of the variation in the three datasets is uncorrelated.
The CV of the gamma distribution fitted to the density of DNMs is 0.18, 0.27 and 0.15 for the Francioli, Wong and Jonsson datasets respectively (Fig 2). The level of variation is significant (i.e. the lower 95% confidence interval of the CV is greater than zero), however the level of variation is modest (Fig 2). A gamma distribution with a coefficient of variation of 0.18 is one in which 90% of regions have a mutation rate within 30% of the mean (i.e. if the mean is one, between 0.7 and 1.3). The gamma distribution fits the distribution of rates qualitatively quite well (S2 Fig; S3 Fig for 100KB), even though a goodness-of-fit test rejects the model at both the 100KB and 1MB scales in all three datasets (p<0.001 in all cases). At the 1MB the observed distribution is more peaked than the fitted gamma distributed; there are too many regions with very low, very high and intermediate numbers of DNMs.
If we include estimates of the distribution for 10KB, 100KB and 10MB we find, as expected, that the variance in the mutation rate declines as the scale gets larger (Figs 3 and 4). This is more marked for the Francioli dataset than for the Wong and Jonsson datasets (Figs 3 and 4). If we plot the CV of the fitted gamma distribution against the window size we find that the log of the CV of the gamma distribution is approximately linearly related to the log of the window size for the Francioli and Wong datasets (Fig 4); the relationship appears curvi-linear for the Jonsson dataset. The fact that the CV declines gradually across scales suggests that the variation at the 1MB scale is part of a continuum of variation at different scales. The linearity of the relationship in two of the datasets suggests that a simple phenomenon may underlie the variation at different scales.
If all the variation at the larger scales is explainable by variation at a smaller scale, then the CV at scale x should be equal to the CV at some finer scale, y, divided by the square-root of x/y; on a log-log scale this should yield a slope of -0.5. The slope for each dataset is shallower than this (Francioli b = -0.25; Wong b = -0.10; Jonsson b = -0.16). This therefore suggests that there is variation at a larger scale that cannot be explained by variation at a smaller scale. To test whether this is the case, we ran a series of one-way ANOVAs; testing variation at the 100KB scale using 10KB windows, 1MB using 100KB windows and 10MB using 1MB windows. The results were significant for all datasets (p<0.001 in all cases).
If we estimate the distribution for individual mutational types we find that in many cases the lower CI on the CV is zero; this might be because we do not have enough data to reliably estimate the distribution for each individual mutational type. We therefore combined mutations into a variety of non-mutually exclusive categories. In each case we estimated the distribution for the relevant category of sites–e.g. in considering the distribution of CpG rates we consider the number of CpG DNMs at CpG sites, not at all sites. We find that the estimated distributions are similar for different mutational types except that there is rather more variation at CpG sites in the Francioli dataset (Fig 3; 100KB results S1 Table). Although the distributions are fairly similar for different mutational types, likelihood ratio tests demonstrate that there are significant differences between mutational categories (S2 Table for 1MB and 100KB results); this is particularly apparent for the Jonsson dataset, probably as a consequence of the size of this dataset. Never-the-less the differences between different mutational categories are relatively small.
Given that there is variation in the mutation rate at the 1MB scale and that this variation is quite similar in magnitude for different mutational types, it would seem likely that the rate of mutation for the different mutational types are correlated. We find that this is indeed the case. We observe significant correlations between all categories of mutations in the three datasets (Table 1; S3 Table for 100KB). The correlations are weak but this is to be expected given the large level of sampling error. To compare the correlation to what we might expect if the two categories of mutation shared a common distribution and were perfectly correlated, we simulated data under a common distribution, estimating the CV of the common distribution as the mean of the distributions fitted to the two mutational categories. We find that generally the observed correlations are similar, and not significantly different, to the expected correlations. In some cases, we observe that the simulated correlation is actually consistently weaker than the observed correlation; this may reflect the inadequacy of the gamma distribution in describing the distribution of rates.
The fact that the rates of Strong to Weak base pairs (S>W) and W>S mutation covary (Table 1) suggests that mutational biases are unlikely to generate much variation in GC-content across the genome. To investigate this further, we used two approaches to test whether there was variation in the pattern of mutation that could generate variation in GC content. First, we used the DNM data for each window to predict the equilibrium GC content to which the sequence would evolve, fitting a model by maximum likelihood (ML) in which this equilibrium GC-content could vary across the genome. The ML estimate for the mean equilibrium GC-content is similar in all datasets at ~0.32. The ML estimate and its 95% CIs for the standard deviation for the equilibrium GC-content are 0.02 (0, 0.060), 0.001 (0, 0.036) and 0.011 (0, 0.024) for the Francioli, Wong and Jonsson respectively; in each case confidence intervals encompass 0, suggesting that a model with no variation in equilibrium GC-content fits the data well. Furthermore, the upper confidence interval is small, suggesting that at most variation in the pattern of mutation generates little variation in GC-content.
However, the ML method does not rule out the possibility that there is some variation in the pattern of mutation. Furthermore, the method does not take into account the difference in the mutation rate between CpG and non-CpG sites. We therefore used a second approach in which we grouped windows together based on their current GC-content. We then estimated the mutation rates for the 9 categories of mutation using the DNM data and used these estimated mutation rates in a simulation of sequence evolution, in which we evolved the sequence to its equilibrium GC content. We find no correlation between the equilibrium GC content to which the sequence evolves and the current GC content (Fig 5; S4 Fig for 100KB).
It has been suggested that the mutation rate at a site is predictable based on genomic features, such as replication time, by Michaelson et al. [5], or the 7-mer sequence in which a site is found, by Aggarwala et al. [40]. To investigate whether these models can explain the variation at large scales we used the models to predict the average mutation rate for each 100KB or 1MB region and correlated these predictions against the observed number of DNMs per site.
We find that the density of DNMs is significantly correlated to the rates predicted under the 7-mer model of Aggarwala et al. [40]. This correlation is significantly positive for the Wong and Jonsson datasets, as we might expect, but significantly negative for the Francioli dataset (Table 2; S4 Table for 100KB results). To compare these correlations to what we might expect if the Aggarwala model explained all the variation at large scales, we simulated the appropriate number of DNMs across the genome according to this model. The observed correlation is significantly smaller than the expected correlation for all datasets, however, the observed and expected correlations are quite similar for the Wong dataset suggesting that much of the variation in DNM density in this dataset is explainable by the model of Aggarwala et al. [40]. However, the model explains almost none of the variation in the Jonsson dataset.
In contrast, the density of DNMs is significantly positively correlated to the predictions of the Michaelson model in the Francioli and Jonsson datasets, but not for the Wong dataset. However, in all cases the correlation is substantially and significantly smaller than it could be if the model explained all the variation (Table 2; S4 Table for 100KB results) suggesting that this model fails to capture much of the variation at the 1MB and 100KB scales.
To try and understand why there is large scale variation in the mutation rate, we compiled a number of genomic variables which have previously been shown to correlate to the rate of germline or somatic DNM, or divergence between species: male and female recombination rate, GC content, replication time, nucleosome occupancy, transcription level, DNA hypersensitivity and several histone methylation and acetylation marks [3, 5, 9, 41, 42].
Surprisingly, the three datasets yield different patterns of correlation. The overall density of DNMs is significantly positively correlated to male and female recombination rates across all datasets, but otherwise there is no consistency (Table 3; 100KB results S5 Table); for example, DNM density is negatively correlated to replication time (later replicating regions have higher mutation rates) in the Francioli and Jonsson datasets, but positively correlated in the Wong dataset, and despite containing 10-times as much data, the correlation is weaker in the Jonsson than the Francioli dataset. Overall, the correlations are more similar in their direction in the Francioli and Jonsson datasets.
Many of the genomic variables are correlated to each other. If we use principle components to reduce the dimensionality, the first principle component (PC) explains 58% of the variation in the genomic variables, the second 13%, the third and fourth 6.9 and 5.7% of the variation. We find that the density of DNMs is significantly negatively correlated to the first PC in the Francioli data (r = -0.14, p<0.001), significantly positively in the Wong data (r = 0.14, p<0.001) and uncorrelated in the Jonsson data (r = -0.013, p = 0.54). All are significantly positively correlated to the second PC (Francioli, r = 0.14, p<0.001; Wong, r = 0.27, p<0.001; Jonsson, r = 0.15, p < 0.001), uncorrelated to the third component and Wong and Jonsson are significantly correlated to the fourth component but in opposite directions (Wong, r = -0.059, p = 0.005; Jonsson, r = 0.1, p<0.001).
It is possible that the differences between Wong and the other datasets are due to biases in the ability to call DNMs. However, analysing the Wong data using the number of callable trios at each site does not qualitatively alter the pattern of correlation in the Wong dataset (Table 3) or the correlations to the principle components of the genomic features (PC1, r = 0.11 p<0.001; PC2, r = 0.25, p<0.001; PC3, r = -0.019, p = 0.37; PC4, r = -0.048, p = 0.019).
To investigate whether these patterns are consistent across mutational types, we calculated the correlation between the density of each mutational type (e.g. CpG C>T mutations at CpG sites) and the first two PCs of the genomic features. For the Francioli and Jonsson datasets the patterns are perfectly consistent; all mutational types, if they show a significant correlation, are significantly negatively correlated to the first PC, and significantly positively correlated to the second (S6 Table). For the Wong data, the patterns are more heterogeneous; all mutational types are positively correlated to the second PC, but some mutational types are significantly positively correlated to the first PC and others significantly negatively correlated.
In order to try and disentangle which factors might be most important in determining the rate of mutation we used stepwise regression. We find, as expected, that the models selected for the three datasets are different (Table 4); only male recombination rate is common to and correlated in the same direction in all three models. The differences are not due to variation in the ability to call DNMs in the Wong dataset since repeating the analyses using the sum of callable trios rather than sites, does not alter the patterns (Table 4). At the 100KB scale, replication time joins male recombination factor as a common factor in all three datasets (S7 Table).
The differences between the three datasets could be due to paternal age since Francioli et al. [3] showed that the correlation between DNM density and replication time was only evident amongst individuals born to young fathers (<28 years), and paternal age differs between the three studies: the average paternal age was 27.7 years in the Francioli dataset (Laurent Francioli pers comm), 33.4 years in the Wong data [6] and 32.0 in the Jonsson data (calculated from their supplementary data). To investigate whether this could explain the differences between the datasets we divided the DNMs into those discovered in individuals with young (<28 years) and old fathers (≥28 years), and regressed the normalised DNM density (dividing by the mean DNM density for each dataset in each age cohort) against replication time and PC1. We find no evidence that the relationship between DNM density and replication time (or PC1) is stronger in individuals born to young fathers in the Wong and Jonsson datasets (Table 5).
The amount of variation explained by the multiple regression models is small– 0.044, 0.10 and 0.042 for Francioli, Wong and Jonsson respectively—but this might be expected given the small number of DNMs per MB and hence the large sampling error. To investigate how much of the explainable variance the model explains we sampled rates from the gamma distribution fitted to the distribution of DNMs across the genome and generated DNMs using these rates and then correlated these simulated rates to the true rates (i.e. those sampled from the gamma distribution). The average coefficient of determination for the simulated data is 0.11, 0.39 and 0.42 for the Francioli, Wong and Jonsson datasets respectively suggesting that the regression model explains ~37%, ~26% and ~10% of the explainable variance for the three datasets. In all cases, none of the simulated datasets have a coefficient of determination that is as low as the observed.
The rate of divergence between species is expected to depend, at least in part, on the rate of mutation. To investigate whether variation in the rate of substitution is correlated to variation in the rate of mutation we calculated the divergence between humans and chimpanzees, initially by simply counting the numbers of differences between the two species. There are at least three different sets of human-chimpanzee alignments: pairwise alignments between human and chimpanzee (PW)[43] found on the University of California Santa Cruz (UCSC) Genome Browser, the human-chimp alignment from the multiple alignment of 46 mammals (MZ)[44] from the same location, and the human-chimp alignment from the Ensembl Enredo, Pecan and Ortheus primate multiple alignment (EPO) [45].
We find that the correlation depends upon the human-chimpanzee alignments used and the amount of each 1MB window covered by aligned bases (Fig 6). The correlation is significantly negative if we include all windows for the UCSC PW and MZ alignments at the 1MB scale, but becomes more positive as we restrict the analysis to windows with more aligned bases. In contrast, the correlations are always positive when using the EPO alignments, and the strength of this correlation does not change once we get above 200,000 aligned bases per 1MB. Further analysis suggests there are some problems with the PW and MZ alignments because divergence per MB window is negatively correlated to mean alignment length (r = -0.31, p < 0.0001) for the PW alignments and positively correlated (r = 0.57, p < 0.0001) for the MZ alignments (S5 Fig). The EPO alignment method shows no such bias and we consider these alignments to be the best of those available. Therefore, we use the EPO alignments for the rest of this analysis.
To gain a more precise estimation of the number of substitutions we used the method of Duret and Arndt [21], which is a non-stationary model of nucleotide substitution that allows the rate of transition at CpG dinucleotides to differ to than that at other sites. As expected the divergence along the human lineage (since humans split from chimpanzees) is significantly correlated to the rate of DNMs (Francioli, r = 0.20 p<0.001; Wong, r = 0.16, p<0.001; Jonsson, r = 0.31, p<0.001). However, the correlation between the rate of DNMs and divergence is not expected to be perfect even if variation in the mutation rate is the only factor affecting the rate of substitution between species; this is because we have relatively few DNMs and hence our estimate of the density of DNMs is subject to a large amount of sampling error. To investigate how strong the correlation could be, we follow the procedure suggested by Francioli et al. [3]; we assume that variation in the mutation rate is the only factor affecting the variation in the substitution rate across the genome between species and that we know the substitution rate without error (this is an approximation, but the sampling error associated with the substitution rate is small relative to the sampling error associated with DNM density because we have so many substitutions). We generated the observed number DNMs according to the rates of substitution, and then considered the correlation between these simulated DNM densities and the observed substitution rates. We repeated this procedure 1000 times to generate a distribution of expected correlations. Performing this simulation, we find that we would expect the correlation between divergence and DNM density to be 0.30, 0.44 and 0.68 for the Francioli, Wong and Jonsson datasets respectively, considerably greater than the observed values of 0.20, 0.16 and 0.31 respectively. In none of the simulations was the simulated correlation as low as the observed correlation.
There are several potential explanations for why the correlation is weaker than it could be; the pattern of mutation might have changed [39, 46–48], or there might be other factors that affect divergence. Francioli et al. [3] showed that including recombination in a regression model between divergence and DNM density significantly improved the fit of the model; a result we confirm here; the coefficient of determination when the sex-average recombination rate is included in a regression of divergence versus DNM density increases from 0.039 to 0.14, 0.026 to 0.12 and 0.095 to 0.18 for the Francioli, Wong and Jonsson datasets respectively; similar patterns are observed for male and female recombination rates separately.
As detailed in the introduction there are at least four explanations for why recombination might be correlated to the rate of divergence independent of its effect on the rate of DNM: (i) biased gene conversion, (ii) recombination affecting the efficiency of selection, (iii) recombination affecting the depth of the genealogy in the human-chimpanzee ancestor and (iv) problems with regressing against correlated variables that are subject to sampling error. We can potentially differentiate between these four explanations by comparing the slope of the regression between the rate of substitution and the recombination rate (RR), and the rate of DNM and the RR. If recombination affects the substitution rate, independent of its effects on DNM mutations, because of GC-biased gene conversion (gBGC), then we expect the slope between divergence and RR to be greater than the slope between DNM density and RR for Weak>Strong (W>S), smaller for S>W, and unaffected for S<>S and W<>W changes. The reason is as follows; gBGC increases the probability that a W>S mutation will get fixed but decreases the probability that a S>W mutation will get fixed. This means that regions of the genome with high rates of recombination will tend to have higher substitution rates of W>S mutations than regions with low rates of recombination hence increasing the slope of the relationship between divergence and recombination rate. The opposite is true for S>W mutations, and S<>S and W<>W mutations should be unaffected by gBGC. If selection is the reason that divergence is correlated to recombination independently of its effects on the mutation rate, then we expect all the slopes associated with substitutions to be less than those associated with DNMs. The reason is as follows; if a proportion of mutations are slightly deleterious then those will have a greater chance of being fixed in regions of low recombination than high recombination. If the effect of recombination on the substitution rate is due to variation in the coalescence time in the human-chimp ancestor, then we expect all the slopes associated with substitution to be greater than those associated with DNMs; this is because the average time to coalescence is expected to be shorter in regions of low recombination than in regions of high recombination. Finally, if the effect is due to problems with multiple regression then we might expect all the slopes to become shallower. Since the DNM density and divergences are on different scales we divided each by their mean to normalise them and hence make the slopes comparable.
The results of our test are consistent with the gBGC hypothesis; the slope of divergence versus RR is greater than the slope for DNM density versus RR for W>S mutations and less for S>W mutations (Fig 7); we present the analyses using sex-averaged RR, but the results are similar for either male or female recombination rates, and for 100KB windows (S6 and S7 Figs and S8 and S9 Tables). These differences are significant in the expected direction for all comparisons except W>S from the Wong data (Table 6)(significance was assessed by bootstrapping the data by MB regions 100 times and then recalculating the slopes). There are no significant differences between the slope for W<>W and S<>S mutations and the slope for substitutions, consistent with gBGC, except for the Jonsson dataset in which the DNM slope is significantly less than the slope for substitutions. This latter result suggests that there might also be an effect of linked selection, but this result should be treated with caution given that the other two datasets show the opposite pattern.
Just as we expect there to be correlation between divergence and DNM rate, so we might expect there to be correlation between DNA sequence diversity within the human species and the rate of DNM. To investigate this, we compiled the number of SNPs in 1MB and 100KB windows from the 1000 genome project [49, 50]. There is a positive correlation between SNP density and DNM density in all datasets (Francioli r = 0.18 p<0.001; Wong r = 0.31, p<0.001; Jonsson r = 0.43, p<0.001).
Using a similar strategy to that used in the analysis of divergence we calculated the correlation we would expect if all the variation in diversity was due to variation in the mutation rate by assuming that the level of diversity is known without error, and hence is a perfect measure of the mutation rate (we have on average 31,000 SNPs per MB, so there is little sampling error associated with the SNPs). We then simulated the observed number of DNMs according to these inferred mutation rates. The expected correlations are 0.24, 0.35 and 0.58 in the Francioli, Wong and Jonsson datasets, which are slightly higher than the observed correlation, significantly so for Francioli and Jonsson (p<0.01 in both cases). The observed correlations are 74%, 89% and 74% of the expected correlations for Francioli, Wong and Jonsson respectively. A similar pattern is observed for individual mutational types at both the 1MB and 100KB scale, with some being greater and others smaller than expected (S10 Table). These results suggest that much of the variation in diversity at the 1MB scale is due to variation in the mutation rate.
Although much of the variation in diversity appears to be due to variation in the mutation rate we tested for the effect of gBGC. We find the slopes are consistent with gBGC for the Francioli dataset, but the other datasets show inconsistent patterns; in the Wong data, the slope of DNM versus RR is significantly greater than the slope of SNP density versus RR across all mutational categories and the opposite pattern is found in Jonsson (p<0.01 in all cases) (Fig 7).
The divergence between species, usually humans and macaques, is often used to control for mutation rate variation in various analyses. But how does the correlation between divergence and the DNM rate in humans change as the species being compared get further apart? Terekhanova et al. [48] showed that the rate of S<>S and W<>W substitutions (chosen to eliminate the influence of gBGC) along the human lineage at the 1MB scale is correlated to that along other primate lineages, but that the correlation declines as the evolutionary distance increases. This suggests that the mutation rate evolves at the 1MB relatively rapidly. However, they did not consider DNMs in detail. To investigate further, we compiled data from a variety of primate species–human/chimpanzee/orangutan (HCO) considering the divergence along the human and chimp lineages, human/orangutan/macaque (HOM) considering the divergence along the human and orangutan lineages, and human/macaque/marmoset (HMM) considering the divergence along the human and macaque lineages. This yields two series of divergences of increasing evolutionary divergence: the human lineage from HCO, HOM and HMM, and chimp from HCO, orangutan from HOM and macaque from HMM. We estimated the divergence using the non-stationary method of Duret and Arndt [21] that treats CpG sites separately. We do not restrict ourselves only to DNMs in the aligned regions but used all DNMs in each window. In this way, the average number of DNMs per window is independent of the evolutionary divergence. As expected, we find that the correlation between the density of DNM and the rate of substitution declines as the evolutionary divergence increases, except the correlation between the density of DNMs in the Francioli dataset and the divergence along the human lineage since the divergence from orangutan which is slightly lower than the correlation with divergence since humans split from macaques (Fig 8). It is also notable that the decrease in the correlation is quite modest in many cases.
We have considered the large-scale (1MB or 100KB) distribution of DNMs along the human genome using an analysis of 3 datasets obtained by the sequencing of trios (an individual and their parents). Unfortunately, there are significant differences between these datasets; most conspicuously they show different patterns of correlation to genomic variables. For example, the density of DNMs at the 1MB scale is significantly negatively correlated to the density of H3K4me1 epigenetic marks in the dataset of Francioli et al. [3], significantly positively correlated in Wong et al. [6] and uncorrelated in Jonsson et al. [36] despite this being by far the largest dataset. However, these correlations to genomic variables are weak, and explain only a small fraction of the explainable variance, and there are many commonalities between datasets, which likely represent true patterns. There appears to be rather little variation in the mutation rate at a large scale in all datasets. However, there is variation at a large scale that cannot be explained by variation at smaller scales, and large-scale variation forms part of a continuum of variation across different scales. Furthermore, the level of variation for different mutational types is similar and different mutational types covary together. There is no evidence that variation in the pattern of mutation generates variation in GC content that would underlie the maintenance of isochores. In all datasets, the correlations to genomic variables are weak and explain little of the explainable variance. We confirm that the correlation between the mutation rate, as measured by DNM density, and divergence, is not as strong as it could be across datasets, and demonstrate that this is in part due to BGC. In contrast, we find that variation in diversity at large scales is largely a consequence of variation in the mutation rate. Finally, we demonstrate that the correlation between the rate of DNM and the rate of substitution, declines as increasingly divergent species are considered.
It is possible that the differences between datasets are due to parental age, since Francioli et al. [3] found that the correlation between DNM density and replication time was only evident in individuals born to young fathers, and paternal age differs between our datasets. However, like Besenbacher et al. [51], we find no evidence that paternal age affects the relationship between the mutation rate and replication time or genomic variables, as summarised by the first principle component of the genomic variables, in either the Wong or Jonsson datasets.
It is also possible that the differences between the datasets are due to ethnicity, since it has been shown that the rate and pattern of mutation, at the single nucleotide scale, varies over short timescales, such that it can vary between human populations [39, 46, 47]; for example, the rate of TCC to TTC is elevated in Europeans [39, 46]. It has also been demonstrated that the mutation pattern evolves at larger scales. Terekhanova et al. [48] considered the correlation between the rate of S<>S and W<>W substitution along the human and other primate lineages at the 1MB scale. They showed that the strength of the correlation declines as more distant species are considered suggesting that the mutation rate evolves at this scale. However, the rate of decline was fairly slow, and human populations would not be predicted to show very different patterns from this analysis. Furthermore, it seems that the populations considered by the three studies were dominated by individuals from the same population, Europeans: Dutch in the study of Francioli et al. [3], Icelanders in Jonsson et al. [36] and mostly North American Europeans in Wong et al. [6] (see [52] for ethic details).
Without any other obvious explanation, it therefore seems likely that the differences between datasets are due to sequencing technology, or the pipelines used to call the DNMs. The Francioli [3] and Jonsson [36] datasets were largely sequenced using Illumina Hiseq at 13x and 35x coverage respectively. The Wong [6] dataset was sequenced using the DNA nanoball technology at 60x coverage. The datasets were subject to a variety of different methods to call DNMs. One potential problem is a GC-bias that has been documented for Illumina sequencing [53], in which high and low GC-content reads are under-represented [54]. To investigate whether this might be the cause of the differences between datasets we regressed the number of DNMs per MB against GC content, and the square of the GC content, to allow for non-linearity. We find that both linear and quadratic terms are significant for the Francioli (p<0.05 for both terms) and Wong (p<0.001 for both terms) datasets, but neither coefficient is significant in the Jonsson dataset. In the Francioli dataset high GC-content regions have fewer DNMs, whereas in Wong it is the low GC-content regions that have a deficit (Fig 9). If we take the residuals from the regression and correlate these against genomic variables we find consistent patterns across datasets (Table 7): the GC-content corrected DNM density is significantly positively correlated to male and female recombination rates, and significantly negatively correlated to replication time and H3K4me3 across all datasets. There are some other significant correlations to histone marks in each of the datasets, with the sign of the correlation being consistent across datasets. If we calculate the principle components for the genomic variables, excluding GC-content, we find that the first four components explain 55, 14, 7.4 and 6.2% of the variance respectively. We find consistent patterns of correlation across datasets in terms of the sign of the correlation (Table 7)—the GC-corrected density of DNMs is negatively correlated to the first PC, but only significant for Jonsson, significantly positively correlated to the second PC in all datasets, uncorrelated to the third and only significantly correlated to the fourth in Jonsson (Table 7).
Despite the fact that the GC-corrected densities of DNMs show similar correlations to genomic variables, we do not find similar models selected by forward selection in a multiple regression (Table 7). Only one feature is common to all datasets–replication time. The differences between the datasets may reflect the strong correlations between genomic variables, which makes it difficult for any procedure to select the correct model.
The fact that correcting for GC content yields similar patterns of correlation to genomic variables, suggests that there is a GC-bias in detecting DNMs. However, regressing against GC-content does not necessarily yield the correct pattern, because there may be a genuine relationship between the mutation rate and GC-content. For example, if we regress the number of W<>W and S<>S substitutions, chosen to remove the influence of BGC, between human and chimpanzee against GC-content we find a U-shaped relationship, unlike that seen for any of the DNM datasets (Fig 9); this might reflect the true pattern. Are there any clues as to which dataset reflects the true pattern of correlation? If we consider the correlation between S<>S and W<>W substitutions, and genomic variables (1MB Table 3; 100KB S5 Table) we find the correlations most closely parallel those in the Francioli dataset; when there is a significant correlation both the substitution date and the Francioli DNMs are significant in the same direction. In contrast, the sign of the correlation is usually the same in the substitution and Jonsson datasets, but the correlations are often non-significant in the Jonsson data. Wong shows very different patterns with some significant correlations in opposite directions.
If the differences between datasets are due to sequencing and processing technology then this has important implications for understanding the reasons the mutation rate varies across the genome because no two datasets show identical patterns of correlation between DNM density and genomic variables. We would suggest that unless a pattern can be shown to be consistent across datasets generated by different sequencing and processing technologies then it must be treated with some caution.
There are two additional points to make about correlations to genomic variables. First, it is evident that many genomic variables are highly correlated to each other so disentangling them will be difficult. Applying multiple regression may not be informative because few of the genomic variables are known without error, so the variables which come out as correlated may not be the causative factors, but those known with the least error. Second, genomic variables explain rather little of the variance in the rate of DNM. This may be because the genomic variables are measured with considerable error, or it may be that we are not assaying the factors which are important; but what these might be, is far from clear.
The evolution of the large-scale variation in GC-content across the human genome has been the subject of much debate [25]; mutation bias [14–18], selection [13, 19, 20] and biased gene conversion [21–24] have all been proposed as explanations. There is good evidence that biased gene conversion has some effect on the base composition of the human genome [24, 26–28]. However, this does not preclude a role for mutation bias. We have tested the mutation bias hypothesis using the DNM data and two different tests. We find no evidence that the pattern of mutation varies across the genome in a way that would generate variation in GC-content. Instead we provide additional evidence that biased gene conversion influences the chance that mutations become fixed in the genome.
We find that previous models of mutation rate variation do not explain the variation in DNM density seen in our datasets at large scales. This is perhaps not surprising. The model of Michaelson et al. [5] was derived by regressing a small number (~600) of DNMs against a suite of genomic variables at multiple scales. So, whilst the model took into account genomic variables at large scales it was principally aimed at estimating the rate of mutation at a single site. The model of Aggarwala et al. [40] estimated the mutation rate at individual sites based on the 7-mer context. It therefore contained no explicit information about large-scale variation.
As expected the rate of divergence between species is correlated to the rate of DNM, however, the strength and even the sign of the correlation depends on the alignments being used. The correlations between divergence and DNM density are actually negative if no filtering is applied to the UCSC alignments, and there is a negative correlation between divergence and alignment length for the pairwise alignments from the UCSC genome browser, and a positive correlation for the multi-species alignment. It is clear that there are problems with these alignments and that they should be used with caution.
As Francioli et al. [3] showed, the correlation between divergence and DNM density is worse than it would be if variation in the mutation rate was the only factor affecting divergence. This is perhaps not surprising because the substitution rate depends both on the rate of mutation and the probability of fixation, both of which may vary across the genome. Francioli et al. [3] further demonstrated that although the rate of DNM is correlated to the rate of recombination, divergence is correlated to the rate of recombination independently of this effect. There are at least four explanations for the effect of recombination on divergence: (i) biased gene conversion, (ii) direct selection, (iii) linked selection and (iv) problems with multiple regression. We have provided evidence for an effect of biased gene conversion, but no clear evidence of three other factors–i.e. the slope of the regression between DNM density and RR is not significantly different to the slope of the regression between divergence and RR for S<>S and W<>W mutations, except in the Jonsson data. However, whilst the slope of DNMs versus RR is lower than the slope of divergence versus RR in Jonsson, we see the opposite pattern in the other two datasets.
The fact that there is no obvious effect of indirect selection is surprising given the results of McVicker et al. [29]. They showed that the divergence between humans and chimpanzees was significantly lower near exons and other regions of the genome subject to evolutionary constraint. A similar reduction was not observed in the divergence of human and macaque and human and dog, suggesting that the pattern was not due to selection outside exons, or regions identified as being subject to selection (though see Phung et al.[30] who detected a correlation between divergence and proximity to functional DNA in the divergence between humans and rodents). They therefore inferred that the reduction was due to the effect of linked selection reducing diversity in the human-chimpanzee ancestor. There are several possible reasons why we see no evidence of this effect in our analysis. First, our test may not be powerful enough. Second, the effects may be counteracted by direct selection which is expected to affect the slope of the regression between divergence and RR in the opposite direction to indirect selection. Third, the scale, magnitude and variation in the effects of indirect selection may be not large enough to affect the relationship between divergence and the rate of mutation; if there is little variation in the magnitude of the indirect effects of selection across the genome at the 1MB (or 100KB) level then indirect selection will have no effect on the correlation between the rate of mutation and divergence. McVicker et al. [29] and Phung et al.[30] presented evidence of indirect selection affecting the divergence between humans and chimpanzees, but over short scales of <100KB. It is possible that at fine scales indirect selection may be more important.
In contrast to the pattern with divergence, we find that much of the variation in diversity, at least at the 1MB and 100KB scales, can be explained by variation in the mutation rate. This suggests that much of the correlation between diversity and RR [31–33, 55–57] is due to variation in the mutation rate not to linked selection. However, the correlation between DNM density and diversity is not as strong as it could be and this could be due to linked selection. Considering that much of the variation in diversity is due to variation in the mutation rate, it is perhaps not surprising that the analysis of DNM density versus RR and SNP density versus RR slopes are inconclusive. The results from Francioli are consistent with BGC affecting the relationship between SNP and DNM density, but the data from Wong and Jonsson are not.
Divergence between species has often been used to control for mutation rate variation in humans (for example [29, 55, 58, 59]). This is clearly not satisfactory given that the correlation between divergence and rate of DNM is only about half as strong as it could and this correlation gets worse as more divergent species are considered (see also Terekhanova et al. [48]). Unfortunately, correcting for mutation rate variation is likely to be difficult because attempting to predict mutation rates from genomic features is unreliable, given that regression models explain less than half the explainable variation. Furthermore, the largest amounts of variation are at the smallest scales (Fig 4) where we have the lowest density of DNMs.
We find, as others have before [3, 5], that the rate of germ-line DNM is correlated to a number of genomic features. However, we find that these features explain less than 50% of the explainable variance leaving the majority of the variance unexplained. Our inability to predict the mutation rate might be because the genomic features have not been assayed in the relevant tissue, the germ-line, or that there are important features that have yet to be assayed. Interestingly, Terekhanova et al. [48] showed that this unexplained component of the substitution rate evolves more rapidly than the explained component. They demonstrated that the substitution rates at the 1MB level in a range of primate species were almost as well correlated to genomic features in humans, as the substitution rate along the human lineage. This implies that the variance in the substitution rate not explained by genomic features, evolves rapidly, given that the correlation between the substitution rate in humans and other lineages declines as they get more distant. There is clearly much we do not currently understand about the why there is large scale variation in the mutation rate and how it evolves through time. Understanding these patterns is challenging given that different datasets show different patterns. Never-the-less there are some patterns which are common to all datasets.
Details of DNM mutations were downloaded from the supplementary tables of the respective papers or from the relevant web-sites: 105,385 mutations from Jonsson et al. [36], 26,939 mutations from Wong et al. [6] and 11016 mutations from Francioli et al. [3]. The data from Jonsson et al. was mapped to hg38 so the liftover tool was used to map these to hg19. Only autosomal DNMs were used.
Three sets of alignments were used in this analysis, all based on human genome build hg19/GRCh37: (i) the University of California Santa Cruz (UCSC) pairwise (PW) alignments [43] for human-chimpanzee (hg19-panTro4 downloaded from http://hgdownload.cse.ucsc.edu/goldenpath/hg19/vsPanTro4/) (ii) the UCSC MultiZ (MZ) 46-way alignments [44] downloaded from http://hgdownload.cse.ucsc.edu/goldenpath/hg19/multiz46way/ and (iii) Ensembl Enredo, Pecan, Ortheus (EPO) 6 primate multiple alignment, release 74, [45] downloaded from ftp://ftp.ensembl.org/pub/release-74/emf/ensembl-compara/epo_6_primate/. We found that the EPO alignments were the most reliable–see main text–and they were used for the majority of the analyses.
All SNPs from the 1000 genomes project phase 3 [50] were downloaded from http://hgdownload.cse.ucsc.edu/gbdb/hg19/1000Genomes/phase3/. After removing all multi-allelic SNPs and, structural variants and indels we were left with 77,818,368 autosomal SNPs. After filtering out windows which had less than 50% of nucleotides aligning between human-chimpanzee-orangutan and no recombination rate scores we were left with 71,917,321 SNPs.
We considered how well the variation at the 100KB and 1MB scale was predicted by two models of mutation rates: the rates estimated by Aggarwala et al. [40] based on the 7-mer context surrounding a site, and the rates estimated for each site by Michaelson et al. based on a variety of genomic features. The rates for Aggarwala et al. [40] were taken from their S7 Table, and the context of each site was used to predict the average mutation rate for each 100KB or 1MB window using their model. The mutability indices from the Michaelson et al. study [5] were provided by the authors. The analysis of the model of Michaelson et al. [5] is more complex since they give the probability of detecting a DNM in their data at each site in the genome, referred to as the mutability index (MI), but these do not translate directly into mutation rates. Using their DNM data we tabulated the number of sites in the genome with a given MI along with the number of DNMs from their study that had been observed at those sites. Because DNMs are not observed at some MIs we grouped MIs into groups of ten starting from the first MI with at least one DNM. We then regressed the log of the number of DNMs over the number of sites against the mean MI (see S8 Fig). The regression line was estimated to be log(mutation rate) = -6.73 + 0.0103 x MI. Using this equation, we predicted the mutation rate at each site in the genome. Michaelson et al. [5] give MIs mapped to hg18; we lifted these over the hg19 using the liftover tool.
Male, female and sex-averaged standardised recombination rate data [60] were downloaded from http://www.decode.com/additional/male.rmap, which provides recombination rates in 10KB steps. For each 100KB and 1MB windows the recombination rate was calculated as the mean of these scores with a score assigned to the window in which the position of its first base resided. GC content was calculated directly from the human genome (hg19/GCRh37) for 100kb and 1Mb windows. All other feature data was taken from the ENCODE project [61] and downloaded from the UCSC genome browser. Where possible we used data from the embryonic stem cell line H1-hESC. The mean value was taken for each genome feature across the window. For replication time data, we downloaded the ENCODE Repli-seq wavelet smoothed signal data [62, 63], provided in 1KB steps, for the GM12878, HeLa, HUVEC, K562, MCF-7 and HepG2 cell lines. Replication times were assigned to windows based upon their start coordinates. We computed the mean replication time for all autosomes for 100KB and 1MB windows across all 6 cell lines. We measured transcription rate using RNA-seq data. Nucleosome occupancy was taken from the GM12878 cell line, histone modifications and RNA-seq data from the stem cell line H1-hESC. We only included windows in our analysis in which >50% of the window had data from all features.
SPSS version 22 and Mathematica version 10 were used for all statistical analyses.
To estimate the mutation rate distribution we use the method of [8]. In brief, we assume that the mutation rate in each window is αu¯ where u¯ is the average mutation rate per site and α is the rate above or below this mean. α is assumed to be gamma distributed. The number of mutations per window is assumed to be Poisson distributed with a mean αu¯l where l is the length of the window. This means that the number of mutations per window is a negative binomial. In considering a particular category of mutations, such as CpG transitions, we considered the number of CpG transition DNMs at CpG sites. We fit the distribution using maximum likelihood using the NMaximize function in Mathematica. Initial analyses suggested that the maximum likelihood value of the mutation rate parameter was very close to the mean estimate of the mutation rate; as a consequence, to speed up the maximization we fixed the mutation rate to its estimated mean and found the ML estimate of the shape parameter of the gamma distribution.
We investigated the correlation between different types of mutation across windows by fitting a single distribution to both types of mutation, estimating the shape parameter of the shared distribution as the mean of the CV of the ML estimates of distributions fitted to the two categories independently. We then used this distribution to simulate data; we drew a random variate for each window from the distribution assigning this as the rate for that window. We then generated two Poisson variates with the appropriate means such that the total number of DNMs for each type of mutation was expected to be equal the total number of DNMs of those types.
To test whether the mutation pattern varied across the genome in a manner that would generate variation in the mutation rate we fit the following model. Let us assume that the mutation rate from strong (S) to weak (W) base pairs, where strong are G:C and weak are A:T, be μ(1 − fe), where μ is the mutation rate and fe is the equilibrium GC-content to which the sequence would evolve if there was no selection or biased gene conversion. Let the mutation rate in the opposite direction be μfe and the current GC-content be f. Then we expect the proportion of mutations that are S->W to be
x(fe,f)=fμ(1−fe)fμ(1−fe)+(1−f)μfe=f(1−fe)f(1−fe)+(1−f)fe
(1)
Let us assume that fe is normally distributed. Then the likelihood of observing i S>W mutations out of a total of n S>W and w>S mutations is
L=∫01N(fe;fe¯,σ)B(n,i,x(fe,f))dfe/∫01N(fe;fe¯,σ)dfe
(2)
The total log-likelihood is therefore the sum of the log of Eq 2 for each MB or 100KB window across all the windows in the genome. The maximum likelihood values were obtained by manually searching for the ML values in Mathematica.
In a number of analyses, we simulate DNMs under assumed model; for example, using the 7-mer model of Aggarwala et al. [40]. In these simulations, we calculate the expected number of DNMs given the window’s mutation rate, the number of relevant sites and the total number of DNMs, and then generated a random Poisson variate from this expectation. In each simulation, we generated 1000 simulated datasets.
|
10.1371/journal.pcbi.1005968 | Reactome graph database: Efficient access to complex pathway data | Reactome is a free, open-source, open-data, curated and peer-reviewed knowledgebase of biomolecular pathways. One of its main priorities is to provide easy and efficient access to its high quality curated data. At present, biological pathway databases typically store their contents in relational databases. This limits access efficiency because there are performance issues associated with queries traversing highly interconnected data. The same data in a graph database can be queried more efficiently. Here we present the rationale behind the adoption of a graph database (Neo4j) as well as the new ContentService (REST API) that provides access to these data. The Neo4j graph database and its query language, Cypher, provide efficient access to the complex Reactome data model, facilitating easy traversal and knowledge discovery. The adoption of this technology greatly improved query efficiency, reducing the average query time by 93%. The web service built on top of the graph database provides programmatic access to Reactome data by object oriented queries, but also supports more complex queries that take advantage of the new underlying graph-based data storage. By adopting graph database technology we are providing a high performance pathway data resource to the community. The Reactome graph database use case shows the power of NoSQL database engines for complex biological data types.
| To better support genome analysis, modeling, systems biology and education, we now offer our knowledgebase of biomolecular pathways as a graph database. We have developed a tool to migrate the Reactome content from the relational database used in curation to a graph database during each quarterly release process. The new graph database has two main advantages; higher performance and simpler ways to perform complex queries. Reactome has already adapted its software infrastructure to benefit from this growing in popularity storage technology, significantly improving query efficiency, by reducing the average query time by 93%. We strongly believe that the successful adoption of a graph database by Reactome demonstrates the positive impact this new technology could potentially have in the field and could provide a practical example for other community projects with similar complex data models to move their storage to a graph database while retaining their data models.
| Reactome (https://reactome.org) is a free, open-source, open-data, curated and peer-reviewed knowledgebase of biomolecular pathways. Reactome annotates processes in a consistent pathway model to create an online resource for researchers as a core reusable pathway dataset for systems biology approaches. Reactome provides infrastructure and intuitive bioinformatics tools for search, visualisation, interpretation and analysis of pathways [1].
Reactome contains a detailed representation of cellular processes, as an ordered network of molecular reactions, interconnecting terms to form a graph of biological knowledge. Like most biomolecular pathway knowledgebases, Reactome has relied on a relational database to store its content. Although widely used among pathway knowledgebases for data management, relational databases are not always the best fit to deal with today’s performance requirements and increasing data complexity [2, 3]. Relational databases cope well with modeling and storing complex pathway information, but the final product is very likely to contain many intermediate tables to represent many-to-many relationships. As a result, database queries across a network of highly interconnected pathway data are often difficult to formulate and require a high number of join operations, ultimately resulting in degradation of performance and excessive response times.
The Reactome data model naturally forms a large interconnected network that can be seen as a directed graph, which consists of a set of nodes and a collection of directed edges connecting ordered pairs of nodes [4]. Storing Reactome pathway data in its natural form has multiple benefits. Most significantly, it does not require any transformation of data into a flat or denormalised table format. As a result, data can be persisted as originally designed, reducing the complexity of the database and thus allowing a more straightforward access to the Reactome knowledgebase [3].
Here we describe the motivation behind our adoption of a graph database and show how Reactome benefits from this change in the underlying storage technology to overcome the previously mentioned limitations imposed by relational databases. The main target audiences for this manuscript are bioinformatics developers, who might be inspired to apply a graph database in a similar domain, and bioinformaticians involved in pathway analysis, who might benefit from using our graph database directly. While users of the Reactome web interface take advantage of the described gains in performance, features, and stability, the Reactome web interface is described in detail in [1].
Reactome uses a frame-based knowledge representation [5]. The data model (https://reactome.org/content/schema) consists of classes (frames) that describe different concepts like reaction or entity. Classes have attributes (slots) that hold properties of the represented class instances, like names or identifiers. The value types contained in the slots can be primitive (string, numbers, or boolean) or references to other class instances. Therefore, knowledge in Reactome is captured as instances of these classes with their associated attributes.
While implementing its relational database, Reactome opted for a physical design that favoured flexibility over performance. Simply put, the relational database incorporated an increased level of abstraction in its physical design resulting in easier adoption of new concepts but at the same time heavily impacting the complexity and execution time of its queries. However, since the graph database natively stores Reactome content in a graph following its model, this trade-off between flexibility and performance is no longer needed.
The Event and PhysicalEntity (PE) classes hold prominent positions in the Reactome model. Events are the building blocks used in Reactome to represent biological processes and are further subclassed into Pathways and ReactionLikeEvents (RLE). RLEs are single-step molecular transformations. RLE includes Reaction among other types like FailedReaction, Polymerisation, Depolymerisation, and BlackBoxEvent. Examples discussed here all involve transformations of the “Reaction” type but all types are handled in the same way with the same results. Pathways are ordered groups of RLEs that together carry out a biological process. PEs are the participants in these events. PE types include SimpleEntity for chemicals, EntityWithAccessionedSequence for proteins, Complex for multi-molecular structures and EntitySet for PEs grouped together on the basis of their shared function.
Persistence of a model, like the one described above, can be achieved with flat files, a relational database, or a non-relational database (e.g. a graph database). The selected underlying storage mechanism determines how data are physically stored and accessed. Consequently, each of these options comes with both advantages and disadvantages in terms of performance and scalability. Until recently, Reactome relied on a relational database (MySQL) for both storing its content during curation and accessing it in its production phase. Among the factors that contributed to this decision were that (1) Protégé (http://protege.stanford.edu) was used as the curator tool during Reactome's nascent years with a Perl script processing the Protégé files to store content into a MySQL database, which was modeled according to the Protégé schema, (2) at the time a relational database met Reactome’s needs for data integrity and consistency, and (3) relational databases were well established for biological data whereas graph based solutions were hardly used in the field [6, 7].
It was not until recently that graph databases became a popular technology in different areas of computational biology. Henkel et al. proposed the concept of graph databases for storage and retrieval of computational models of biological systems [7]. Summer et al. developed a Cytoscape application that takes advantage of the Neo4j database to perform server-side analysis of large and complex biological networks [8]. In [9] the authors explored the potential of using a graph database to facilitate data management and analysis to provide biological context to disease-related genes and proteins. Toure et al. developed a Java-based framework that transforms biological pathways represented in SBGN format into the Neo4j graph database, enabling more powerful management and querying of complex biological networks [10]. Balaur et al. demonstrated that advanced exploration of highly connected and comprehensive genome-scale metabolic reconstructions can benefit from an integrated graph representation of the model and associated data [11]. Swainston et al. described biochem4j that enables complex queries by linking a number of widely used chemical, biochemical and biology resources within a graph database [12].
Reactome has gradually introduced a Neo4j graph database (https://neo4j.com/) to store and query its content in the production phase since July 2016 (version 57). Neo4j is an open source, transactional and ACID (Atomicity, Consistency, Isolation, and Durability) compliant graph database [13]. Native graph databases, such as Neo4j, naturally store, manage, analyze, and use data within the context of connections to improve performance and flexibility when handling highly interconnected data compared to that in SQL. Neo4j’s greatest advantage and probably its most defining feature is Cypher: a declarative, pattern matching query language, specifically designed for dealing with graph data structures [14, 15].
The Reactome knowledgebase has many use cases, like the one in Fig 1, where the use of a graph model together with a query language like Cypher can greatly improve response times and simplify the code necessary to access the data. For instance, recursively retrieving all reactions of a pathway, retrieving the participants of a reaction or a pathway, deconstructing a complex or a set into its participating molecules, or enumerating the chain of consecutive reactions that lead to the formation of a signalling complex are typical use cases that benefit greatly from traversing the graph version of the Reactome knowledgebase.
Fig 1 provides a simplified example where reactions only contain lists of reactants and products, instances of the PE class. In the relational use case, two junction tables, Reaction-input and Reaction-output, are required to model these many-to-many relationships (Fig 1A). Each junction table contains foreign keys of the Reactions and the associated PEs. The SQL query to retrieve input and output entities of a given reaction requires two join operations per junction table (Fig 1B). In the first stage of its execution, each join operation forms the cartesian product between the tables and, during the filtering process, all rows of the result set that are not of interest are discarded.
The same structure of a reaction with inputs and outputs can be modelled in a simpler way with Neo4j as exemplified by the reaction presented in Fig 1C. The reaction (green node), contains named outgoing relationships to corresponding input and output entities (purple nodes). Taking advantage of Cypher, the same query, can be written in a shorter but more intuitive manner thanks to its ASCII-Art syntax [3] to represent patterns (Fig 1D). The query describes a pattern that includes a Reaction, again identified by its identifier, with its outgoing input and output relationships. Finally, all nodes matching the specified pattern are returned.
Since their introduction in the 1970’s, relational database engines have been optimised to provide efficient execution of SQL queries. This is particularly the case with global queries that aggregate large amounts of data without the need to perform any traversal operations. However, Reactome data contain many relationships, like those illustrated in Fig 1, and thus many join tables, so queries generally require traversal operations, a computational intensive task that tends to result in poor performance compared to graph databases [16]. To address this issue and improve query performance, some resources have created redundant denormalised copies of their relational database [17, 18, 19]. Nowadays, graph databases, such as Neo4j, offer a more appropriate alternative for cases of highly interconnected data.
The graph database batch importer (https://github.com/reactome/graph-importer) was developed to migrate the content from the relational database used in curation, to a graph database during each quarterly release process. Although the underlying data storage was changed, the original data model used by MySQL was kept the same. The conversion was done following a depth-first approach starting from the top level pathways and traversing all the content, ensuring that each object is processed only once during the conversion. Every object constitutes a node in the graph and the edges that connect the nodes correspond to the names of the slots as defined in the domain model (Fig 2). As a result, a Neo4j graph database is generated and contains all the Reactome data. It can be directly used for third parties in order to use Cypher to retrieve the target data.
A number of integrity tests have been put in place to ensure that both the graph and relational database have the same content after conversion. These tests are part of the graph-core and they are executed after migrating the relational database to the graph database to ensure that the data has been properly stored. The tests include checks to verify: that the number of top level pathways present in the graph database corresponds to the number of those present in the relational database; that a given pathway in the graph database has the same ancestors as its counterpart in the relational database; that the content of a given complex is the same in both databases.
Fig 3 presents a schematic illustration of the new Reactome graph database ecosystem. A library called graph-core (https://github.com/reactome/graph-core) was developed on top of the graph database to serve as a data access layer. The aim of the library is to provide easy access and data persistence as well as to reduce the boilerplate code in third party projects that require accessing and traversing Reactome content. The graph-core uses Spring Data Neo4j (SDN) [20] to access the graph content and AspectJ to enable lazy loading [21]. Lazy loading commonly refers to a design pattern, that postpones the retrieval of object attributes until the point at which they are needed. In our case, AspectJ weaver is used to intercept the getter methods and run specific code to silently retrieve more data when needed.
The ContentService (https://reactome.org/ContentService) is a REST based web service [22], built on top of the graph-core, to provide programmatic access to the Graph Database for third party developers (https://github.com/reactome/content-service). Implemented on top of Spring MVC (https://spring.io/), the ContentService utilises the graph-core library and is fully documented with Open API (https://www.openapis.org/).
Among its main advantages, this new solution is faster and less computationally intensive than the previous one based on the relational database. Performing queries against the graph database constitutes a more scalable approach, resulting in higher throughput and, ultimately, to a more robust ContentService able to cope with an always increasing number of requests. Additionally, the resulting product is easier to maintain as most new methods can be added by simply writing the respective Cypher queries, avoiding writing complex algorithms in a given programming language (Fig 1B).
The use cases above are available as methods in the ContentService API (https://reactome.org/ContentService/). Fig 4 emphasises how queries of Reactome data have been simplified by the adoption of the graph database. The query in Fig 4A shows how to retrieve the participating molecules for a pathway. The reverse query, identifying pathways where a molecule participates, is shown in Fig 4B, which follows a similar pattern to Fig 4A, but fixes the end-bound and leaves the upper-side open for traversing results. Based on feedback provided by people contacting our help desk ([email protected]) and attending our training sessions, the new way of querying Reactome is easy and intuitive to learn, and researchers, who are interested in performing queries against Reactome data, can learn to write them in Cypher in a relatively short amount of time.
To assess the improvement we designed a set of stress tests to measure the impact of adopting the graph database in Reactome. All stress tests were executed on a standard laptop featuring an Intel Core i7 at 2.6 GHz, 16 GB of DDR3 memory at 1,600 MHz, and 256 GB of flash storage. The tests do not aim to compare the two storage technologies (MySQL and Neo4j) but instead their usage by Reactome. The stress tests were run against the web services build on top of each storage technology and included two scenarios: (1) simulation of one user sequentially querying 5,000 reactions for Homo sapiens and (2) simulating an increasing set of users simultaneously performing the previous task. In each case the resulting data for every reaction had to be marshalled as an instance of the correspondent model class. The test comprised four executions; two against the previous web service running on top of the relational database and the other two accessing the new web service running on top of the graph database through the newly created graph-core library (https://github.com/reactome/graph-core). The reactions were accessed in a sequential fashion to ensure that caching did not provide any sort of advantage for any of the approaches, because a queried object would never be retrieved again in the same test. It should be mentioned that prior to any stress test’s execution, both Neo4j and MySQL databases were configured to allocate 50% of the available physical memory (8GB).
As illustrated in Fig 5, querying the data stored in the relational database resulted in significantly longer response times. In particular, in the case of the relational implementation of the Reactome knowledgebase the average query time was 173.11 ms (±25.81) while in the case of the graph implementation, the average response time dropped to 12.56 ms (±2.94),a 93% reduction in the average query time. The new implementation supported higher throughput, in terms of transactions per second (TPS), reaching 79.5 TPS compared to 5.8 TPS. As a result of this boost in performance, all 5,000 queries to the graph database were performed in 63 seconds while the relational implementation required more than 14 minutes for the same task.
A second stress test simulated a more realistic scenario where multiple users perform concurrent database queries (Fig 6). Once again, querying the Reactome knowledgebase in its relational implementation resulted in significantly longer response times. For instance, in case of 10 concurrent threads performing queries to the relational implementation of the Reactome knowledgebase the average response time was 1,516 ms while in the case of the graph implementation, the average response time dropped to 49.05 ms. In addition, the new implementation achieved higher throughput reaching 203.6 TPS compared to 6.6 TPS. Consequently, the graph implementation of Reactome provides higher scalability enabling Reactome to handle larger volumes of user requests.
Fig 7 presents a comparison between the throughputs achieved by both systems against the number of users performing concurrent queries. The graph implementation achieved a higher number of transactions per second that reached a plateau after the point where the number of active threads becomes equal to the available processor cores; in this case 4. On the other hand, the measured throughput in case of the relational implementation is stable and does not seem to take advantage of any concurrency.
Many users choose to download the Reactome graph database and access the data through Cypher queries directly in their computers. Our usage statistics show that a growing number of users have downloaded the Reactome graph database and, based on the questions gathered by our help desk service, we believe that they have used it to perform local queries against the complete Reactome knowledgebase. In particular, during the first year that Reactome provided the graph database, there were 2,385 downloads by 912 unique users. 118 of those users downloaded the graph database after each data release. It is worth mentioning that during writing of this manuscript, the size of the Reactome relational database in its current data release (v62) is around 2.0GB while the size of the graph database is approximately 1.8GB. Fig 8 provides a summary of the graph database.
With a tool so powerful at managing highly connected data sets and complex queries at our disposal, Reactome is providing faster and more stable services to researchers around the world. In the near future, Reactome plans to upgrade its services and leverage the full potential of Cypher in order to provide answers to questions that require diving deeper into our data. In particular, the integration of a graph database lowers the complexity of problems that require traversing of our knowledgebase, such as identifying causal interactions or revealing all possible paths between two molecules.
Future development in Reactome is not likely to be affected by the fact that Neo4j is by nature schema-less, mainly because the rigid schema of our relational database with all the applied constraints is used to ensure data consistency during the curation phase. Currently, data are migrated to Neo4j during each quarterly release process and are used to speed up queries in production.
In conclusion, through the adoption of the Neo4j graph database, and by harnessing the power of its query language, Reactome provides efficient access to its pathway knowledgebase. As a result of this shift in the underlying data storage technology, the average query time has been reduced up to 93%. In addition, the graph-core library and the ContentService leverage these benefits of this shift and can be used by third party applications to efficiently access Reactome.
Reactome’s successful use case constitutes a strong argument in favour of the positive impact this new technology can have in the field. By following Reactome’s use case, other community projects with similar complex models could benefit from moving their storage to a graph database while keeping their data model. While we have demonstrated the major impact of moving the Reactome public database to a graph database in terms of usability, stability, and response time, we think this is only a milestone in the growing ecosystem of network-oriented biomolecular data resources that will enable entirely new functionalities through moving to modern database technology that better reflects the graph-like structure of their source data. While we will work directly with internal and external resources to move along that path, we would also like to invite the community to use the open data Reactome graph database to develop their own novel uses of Reactome data.
The Reactome graph database is freely available at: https://reactome.org/dev/graph-database. The API for the ContentService is available at https://reactome.org/ContentService with documentation and tutorials available at: https://reactome.org/dev/content-service. The source code, in Java, is freely available at: https://github.com/reactome (See the graph-core, graph-importer and content-service repositories).
Future development will focus on updating the version of SDN and integrating interaction data from IntAct (http://www.ebi.ac.uk/intact/) directly to the Reactome graph database.
|
10.1371/journal.pcbi.1003538 | Correction of Distortion in Flattened Representations of the Cortical Surface Allows Prediction of V1-V3 Functional Organization from Anatomy | Several domains of neuroscience offer map-like models that link location on the cortical surface to properties of sensory representation. Within cortical visual areas V1, V2, and V3, algebraic transformations can relate position in the visual field to the retinotopic representation on the flattened cortical sheet. A limit to the practical application of this structure-function model is that the cortex, while topologically a two-dimensional surface, is curved. Flattening of the curved surface to a plane unavoidably introduces local geometric distortions that are not accounted for in idealized models. Here, we show that this limitation is overcome by correcting the geometric distortion induced by cortical flattening. We use a mass-spring-damper simulation to create a registration between functional MRI retinotopic mapping data of visual areas V1, V2, and V3 and an algebraic model of retinotopy. This registration is then applied to the flattened cortical surface anatomy to create an anatomical template that is linked to the algebraic retinotopic model. This registered cortical template can be used to accurately predict the location and retinotopic organization of these early visual areas from cortical anatomy alone. Moreover, we show that prediction accuracy remains when extrapolating beyond the range of data used to inform the model, indicating that the registration reflects the retinotopic organization of visual cortex. We provide code for the mass-spring-damper technique, which has general utility for the registration of cortical structure and function beyond the visual cortex.
| A two-dimensional projection of the visual world, termed a retinotopic map, is spread across the striate and extra-striate areas of the human brain. The organization of retinotopic maps has been described with algebraic functions that map position in the visual field to points on the cortical surface. These functions represent the cortical surface as a flat sheet. In fact, the surface of the brain is intrinsically curved. Flattening the cortical surface thus introduces geometric distortions of the cortical sheet that limit the fitting of algebraic functions to actual brain imaging data. We present a technique to fix the problem of geometric distortions. We collected retinotopic mapping data using functional MRI from a group of people. We treated the cortical surface as a mass-spring-damper system and corrected the topology of the cortical surface to register the functional imaging data to an algebraic model of retinotopic organization. From this registration we construct a template that is able to predict the retinotopic organization of cortical visual areas V1, V2, and V3 using only the brain anatomy of a subject. The accuracy of this prediction is comparable to that of functional measurement itself.
| The human occipital cortex contains multiple representations of the visual field, starting with primary visual cortex (V1; also called striate cortex). V1 lies primarily within the calcarine sulcus and represents the contralateral visual hemifield. The cortical surface dorsal and ventral to V1 contains the neighboring extrastriate regions V2 and V3, each of which represents a complete visual hemifield that is split into the upper visual quarterfield, ventral to V1, and the lower visual quarterfield, dorsal to V1. These three distinct retinotopic maps are organized on the cortical surface by distance from the fovea (eccentricity) and angle from the vertical meridian (polar angle) [1]. Polar angle sweeps dorsally down and ventrally up from the horizontal meridian in V1 (lying along the calcarine sulcus) around the foveal confluence then reverses direction at the V1/V2 and V2/V3 borders (Fig. 1A). Eccentricity radiates uniformly outward from the foveal confluence in all three visual field maps (Fig. 1B).
This visual area organization is readily demonstrated in people by performing retinotopic mapping using functional magnetic resonance imaging (fMRI). When examined in a population of subjects, the qualitative topographical organization of V1–V3 has been found to be consistent [2]. An important advance in the study of retinotopic organization has been the development of software tools for cortical surface registration [3], [4]. The cortical surface is a topological sheet (specifically a sphere), which is thrown into folds (gyri and sulci). The continuous gray matter layer can be identified on an anatomical brain image, represented as a tessellation of vertex points of triangles, and then digitally inflated and flattened to a 2D surface. The pattern of gyral and sulcal curvature is retained and expressed on the flattened cortical sheet. The pattern of cortical surface curvature is then used to drive between-subject registration of brain anatomy on the cortical surface [4]–[6]. When registered in this way, the cortical location of area V1 is found to be consistent across people [7]. More recent work has shown that the size and location of V1, V2, and V3 are also similar across subjects when cortical surface topology is brought into alignment [7], [8].
The functional organization of retinotopic cortex may be captured in an algebraic model. Early algebraic models of V1–V3 used a log-polar transform to relate visual field position to location on the flattened cortical surface [9]. These 2D models were later improved by Schira et al. to better capture banding near the foveal confluence and cortical magnification in V2 and V3 [10] (Fig. 1C). Although such models are conceptually useful, the cortical surface as measured in imaging experiments does not respect the details of their idealized geometry. To compare the 2D models to functional measurements, a topological transformation must be applied to the measurements to produce a representation of the data on a flattened surface. Such a transformation, however, necessarily introduces non-trivial geometric distortions that cause the flattened cortical representation to deviate from the idealized, 2D plane in which the algebraic model is defined.
In the limited case of area V1, which resides in a single sulcal fold, we have shown that an algebraic model can be fit to retinotopic mapping data on the flattened cortical surface [11]. In this approach, the fMRI data is brought into alignment across subjects by digital inflation and registration of the cortical surface to a standard anatomic atlas [3]–[5], [12]. Within the 2D, flattened cortical atlas space, we were able to aggregate retinotopic mapping data across subjects and then fit the aggregate data with an algebraic model of retinotopic organization [11]. This linking of algebraic model to the 2D cortical surface atlas then allowed us to accurately predict the functional, retinotopic cortical organization of individual subjects by registering their idiosyncratic brain anatomy to the cortical atlas. The algebraic model provided both a regularization of the data in the presence of noise and generalization of the prediction beyond the boundaries of data itself. Despite the success of this approach within area V1, local geometric distortions of the cortical surface were introduced by the 2D flattening, which in turn distorted the functional prediction of retinotopy (e.g., violation of the equal areal magnification property of retinotopic maps [13], [14]). If we wish to extend this approach to the extrastriate visual areas, we will need to contend with the much greater degree of geometric distortion found in the flattened representation of a larger cortical area that reaches over multiple gyral ridges.
Here we provide a means to link fMRI data from visual areas V1–V3 to an algebraic 2D model of retinotopy in the presence of geometric cortical distortion. One might first consider solving this challenge by modifying the algebraic model to better match the data. Mathematically, however, it is both difficult and poorly descriptive of the fundamental structure of retinotopic organization to tailor a 2D model to the local distortions in geometry present in flattened cortical data. Instead, we propose to register the functional data of the flattened cortical surface to the algebraic model. Such a technique distorts the flattened cortical representation to align the functional data to the algebraic model and is thus flexible enough to correct the geometric distortions introduced by flattening. In this approach, the challenge becomes devising a registration technique that is flexible enough to correct the undesired distortions and adequately align to the algebraic model yet sufficiently constrained so that the resulting, registered anatomy retains its structure enough to support generalization of the algebraic model beyond the extent of the data used in the registration.
Mass-spring-damper (MSD) systems are commonly used in the simulation of the deformation of materials and objects [15]–[17]. These systems approximate surfaces or volumes as a series of point masses connected in a mesh by ideal springs (i.e., a spring whose applied force is proportional only to the displacement of the spring). Because of the simplicity of the forces enacted by an ideal spring, simulation of such systems by means of numerical integration is relatively straightforward.
Similar to our prior work in V1, we obtain across-subject retinotopic mapping data that is then aggregated within a standard cortical surface atlas [4]–[6]. We then represent the cortical atlas surface and aggregate retinotopic mapping data as an MSD system which places two sets of springs in opposition. First, all cortical vertices are treated as point masses connected by ideal springs to their neighbors. This spring set resists warping the anatomy of the cortex. A second set of springs connects each cortical point that has a retinotopic mapping value to a fixed position in an overlying algebraic model of retinotopy. This spring set works to modify the cortex to bring the functional data into best alignment with the algebraic model. The simulation identifies a low-energy state of the system which balances these competing forces.
The result of the MSD simulation maps individual vertices within the cortical surface atlas to a specific visual area and visual field position. We show that this mapping may be used to accurately predict the retinotopic organization of extrastriate cortex in a novel subject who's brain anatomy is brought into register with the cortical surface atlas. Further, because the algebraic model is continuous, we find that the mapping may be used to accurately predict retinotopic data collected from beyond the eccentricity range of data used in the aggregate to derive the mapping.
This study was approved by the University of Pennsylvania Institutional Review Board, and all subjects provided written consent.
A total of 25 subjects (15 female, mean age 24, range 20–42) participated in fMRI scanning experiments. All subjects had normal or corrected-to-normal vision. Experimental data from all subjects have been reported previously [11]. Each subject contributed to only one of two datasets.
The first dataset, D10°, contained 19 subjects all of whom were scanned for 27 minutes using a sweeping bar stimulus that extended to 10° of eccentricity within a central 20° aperture. The bar stimulus consisted of a single sweeping 2.5°-thick bar that flickered at 5 Hz [18]. Bars moved 1.25° every 3 s in 4 directions (horizontal, vertical, oblique +45°, oblique −45°) while subjects maintained central fixation.
The second dataset, D20°, contained 6 subjects. Subjects fixated on either the left or right edge of the screen for 64 minutes while 16 iterations of standard “ring and wedge” stimuli swept in the periphery [19].
BOLD fMRI data (TR = 3 s, 3 mm isotropic voxels) and anatomical images (T1-weighted, 1 mm isotropic voxels) were collected at 3 Tesla. The FMRIB Software Library (FSL) toolkit (http://www.fmrib.ox.ac.uk/fsl/) was used to process anatomical images which were then reconstructed and inflated using FreeSurfer (v5.1) (https://surfer.nmr.mgh.harvard.edu/) [3]–[6]. Hemispheres from individual subjects were aligned via surface registration to FreeSurfer's common left-right symmetric pseudo-hemisphere (fsaverage_sym) [3], [12].
For subjects in the D10° dataset, a population average hemodynamic response (HRF) [20] was used to model the BOLD signal. For subjects in dataset D20°, a subject-specific HRF was derived from a separate blocked visual stimulation scan. Global signal, cardiac and respiratory fluctuations (when available) [21], effects of the scan, and spikes (i.e., instances in which the signal deviates from the mean by ≥2 standard deviations) were modeled as nuisance covariates. Polar angle and eccentricity were either modeled (with receptive field size) using the population receptive field (pRF) method [18] (datasets D10°) or by identification of the peak of a Gaussian fit to the weights of a set of finite impulse response covariates (dataset D20°) [22].
Aggregate retinotopic maps of each dataset were produced separately for polar angle and eccentricity by finding the weighted mean polar angle and eccentricity of all subjects at each aligned vertex position. Mean polar angles and eccentricities were weighted by the F-statistic of the confidence of each subject's polar angle and eccentricity assignments. A confidence for each vertex in the aggregate was calculated as the sum of squares of the F-statistics of all significant vertices divided by the sum of the same F-statistics. For a set of subjects Q, each of whom have a vertex at position p on the cortical surface with a polar angle and eccentricity assignment whose significance is above threshold, the confidence of aggregate vertex p is (Σq•Q F(q, p)2)/(Σq•Q F(q, p)) where F(s, x) is the confidence of the polar angle and eccentricity assignment in subject s at vertex position x. The assignment of any vertex whose confidence was below a minimum threshold chosen for the dataset (see Supplemental Mathematica Notebook, §3.2), was discarded. Because averaging produces bias in the direction of the mean near the borders of a finite stimulus range (e.g., values near 0° and 180° of polar angle tend to attenuate toward 90° in the aggregate), the aggregate polar angle values were corrected and eccentricity was truncated by 1.25°. Polar angle correction was performed by forcing the distribution of polar angles in the corrected aggregate to match the distribution of the union of all significant polar angle values of all subjects. More specifically, the uncorrected aggregate polar angle θ of each vertex in the aggregate was changed to a corrected polar angle θ′ such that C(A, θ) = C(M, θ′) where C(D, t) is the cumulative density function of the distribution D, evaluated at t, and A and M are the distributions of the uncorrected aggregate polar angles and union of all significant polar angle values for all subjects, respectively. Eccentricity values below 1.25° and within 1.25° of the outer stimulus border were excluded due to measurement bias near the edge of the stimulus range [23].
All vertices within π/3 radians on the inflated spherical hemisphere of the point p0, defined as the most anterior point on the anatomically defined V1 border [7], were rotated such that p0 lay at the intersection of the equator of the spherical fsaverage_sym brain hemisphere and prime meridian, then flattened via projection onto the plane tangent to the sphere at p0. A shear transformation, present also in our previous treatment of V1 [11], was applied to the flattened data to render the V1 region more elliptical. These flattened and sheared data formed a “flattened occipital region” on the cortical surface.
Data from D10° were registered to a modified version of the banded double-sech model proposed by Schira et al. [10] using a simulated mass-spring-damping system. Each vertex in the flattened occipital region was assigned an initial position identical to its position in the flattened occipital region and a mass of 1 g. Vertex coordinates were measured in radians (rad) according to their angular latitude (y-coordinate) and longitude (x-coordinate) relative to p0 (described above) on the fsaverage_sym spherical hemisphere. All pairs of vertices whose initial positions were within 0.015 rad of each other were connected by a spring whose ideal length was equal to the initial distance between the vertices and whose stiffness was 1.0 g/s2. These “anatomical springs” ensured that warping introduced during the simulation would respect anatomical constraints. Additionally, for each vertex with an above confidence threshold assignment of eccentricity and polar angle in the dataset aggregate, a “model spring” with one fixed and one free end was connected between the vertex (free end) and the position predicted by the algebraic model for the aggregate observed polar angle and eccentricity of the vertex (fixed end). Because there are multiple such points (i.e., in V1, V2, and V3) for each polar angle and eccentricity, the fixed end of the spring was constantly updated throughout the simulation to be positioned at the nearest such point. Model springs were assigned an ideal length of 0 rad and a stiffness of 10 g/s2. To prevent vertices distant from the algebraic model but with polar angle and eccentricity assignments nonetheless above our F-value threshold from having an overly large influence on the simulation due to their high spring length, the potential function of the vertex attached to a model spring was represented as an inverted Gaussian whose center was the ideal position for the vertex in the algebraic model of retinotopy instead of a parabola with the same center. The choice of a Gaussian potential function for use in aligning retinotopic data on the cortical surface is similar to the energy function proposed by Fischl et al. [3] for aligning hemispheres by curvature. Note that because the force acting on the vertex is the gradient of the potential of that vertex, this choice of potential function effectively means that the force acting on a vertex either very close to or very far from its ideal position is near zero. For a spring of length d and stiffness k, the magnitude of the force acting on the ends of an anatomical spring with ideal length d0 is k |d - d0|; for the endpoint of a model spring, the magnitude of the force is 4k |d - d0| exp(−64(d - d0)2), which approximately models the force of a parabolic spring at small distances. An additional force was applied to all pairs of vertices not bonded by springs such that any such pair of vertices within a given distance d, less than some cutoff c, of each other were repelled by (4 c/(d+c)−2) rad g/s2; in our simulations, c was chosen to be half the average anatomical spring length. This “van der Waals”-like force prevents vertices from passing through each other. The motion of all vertices was dampened by 0.1% after each step (i.e., each vertex's velocity was multiplied by 0.999 after each simulation step). Further details concerning the parameterization of the simulation and the stability of these parameters can be found in the Supplemental Materials.
The algebraic model of retinotopic organization was modified from that of Schira et al. [10] by the addition of parameters for translation, rotation, and horizontal and vertical stretch, all of which were necessary to produce an initial fit to the aggregate functional data. The original double-sech model includes parameters a, b, k, and λ. We retain a, b, and λ, but replace k, the scale parameter, with horizontal and vertical scales. Although this breaks certain features of the original Schira model such as the consistency of areal magnification, we note that this point is essentially moot as we are dealing with distorted data already and are further warping it during registration to the model. Accordingly, we focus on the parameters a, b, and λ, which define the shape of the model and for which we use values 1.5, 60, and 2.5 respectively. This parameterization was found by manipulating parameters “by hand” to align them with the aggregate retinotopy; code for experimenting with this fit is provide in our Supplemental Mathematica Notebook (§1.6.5). An additional “V4-like” dorsal and ventral region was added to the model to stabilize vertices in both V3A and hV4 whose retinotopy would otherwise cause them to be attached via model springs to V3. The full parameterization of the algebraic model of retinotopy and source code for calculating and inverting it are provided in the Supplemental Mathematica Notebook, §1.6.
Simulation was performed by numerical integration of the system using a time-step size of 5 ms. At each step t, acceleration values were calculated for each vertex using Newton's second law of motion. Positions were updated such that xt+1 = xt+vt ∂t+at ∂t2/2 and velocities were updated such that vt+1 = vt+at ∂t, where xt, vt, and at are the position, velocity, and acceleration vectors of a given vertex at step t, and ∂t is the step size. Vertices were given small random initial velocities such that the net velocity at time 0 was 0 but such that the total KE of the system was 10 rad2⋅g/s2. Energies were examined every 10 steps and KEs were rescaled whenever the total energy (PE+KE) exceeded the initial energy (PE0) by at least 2 rad2 g/s2 due to numerical drift. Simulations were run with a step size of 2 ms for 5,000 steps (10 s). After simulation, the resulting configuration was minimized by a simple gradient descent search using a gradient step distance of 0.005 for 500 steps or until convergence. Source code for the simulation is provided via a gitHub repository (http://github.com/NoahBenson/SpringRegister/).
By simulating the system until a low PE is achieved, we allow the constraints imposed by both the cortical anatomy and the functional model to relax into a solution that respects both kinds of information. Because the simulation incorporates KE, a nonlocal energy minimum may be found; it is therefore beneficial to use simulated annealing. Four simulations of 10 s (5,000 steps) each were performed such that the final arrangement of vertices in each simulation was used as the starting arrangement for the next simulation; spring ideal lengths were not recalculated, however, and the velocities were re-randomized such that the KE of the system was 10 rad2 g/s2 at the beginning of each simulation. The final arrangement of the four simulations with the lowest PE was chosen as the arrangement of the corrected topology. Vertices were assigned model polar angle and eccentricity values from their positions in the corrected topology by inverting the algebraic model of retinotopy. In other words, if the algebraic model of retinotopy predicts that a point (θ, ρ) in the visual field should lie at position (x, y) on the cortical surface, then a vertex with position (x, y) in the corrected topology would be assigned a polar angle of θ and an eccentricity of ρ.
Retinotopic mapping data was obtained from 19 subjects to an eccentricity of 10° of visual angle (dataset D10°) (Fig. 1A, 1B). The brain anatomy from each subject was registered to an atlas of cortical surface topology (fsaverage_sym), and the across-subject, confidence-weighted mean aggregate of polar angle and eccentricity obtained. The reversals of polar angle that mark the boundaries of visual areas, and the regular progression of eccentricity from the occipital pole, is readily seen in the aggregate data. The goal of our work is to register the measured retinotopy in the volume with an algebraic model on the flattened cortical surface (Fig. 1C). The registration is performed within a flattened patch of the cortical surface atlas (Fig. 1D).
Registration via MSD simulation brought the aggregate retinotopic mapping values into alignment with the algebraic model by warping the cortex. The magnitude and direction of warping induced by this registration (i.e., the distance and angle between each vertex position in the flattened fsaverage_sym atlas space and its position in the corrected topology) is shown in Figs. 2A and 2B. Notably, the greatest displacement of vertices is found around the occipital pole. We presume that the warping of the cortex in our registration is correcting the geometric distortions created during flattening of this region of high curvature. The sulcal folding pattern of the original cortical surface atlas and the corrected topology following MSD simulation are shown in Figs. 2C and 2D respectively, along with the regional assignment (V1, V2, or V3) predicted by applying the algebraic model of retinotopic organization to vertices in the corrected topology.
When aggregated within the cortical surface atlas, polar angle organization is largely consistent across subjects. A flattened aggregate map of the confidence-weighted mean polar angle of the 19 subjects in our 10° eccentricity dataset D10° is shown in Fig. 3A. Although regional boundaries in the aggregate map are apparent, the iso-angular curves in this organization do not resemble the smooth curves found in the algebraic model of retinotopic organization (Fig. 1C), suggesting an opportunity for registration via simulation to improve the predictive accuracy of the algebraic model. The polar angle organization following the MSD simulation is shown in Fig. 3B. As would be expected, minimization of energy in the MSD simulation has warped the cortex to bring the aggregate polar angle data into better alignment with the algebraic model of retinotopic organization. The algebraic model of retinotopy can then be projected back to the original cortical surface atlas (Fig. 3C). This smooth, continuous map of polar angle organization should resemble the measured polar angle functional data of any subject following the registration of their brain anatomy to the cortical surface atlas. We therefore refer to this representation as an anatomical template of retinotopy.
To examine how well our template predicts a subject not previously seen, we calculated leave-one-out errors. To do so, the aggregate polar angle data was obtained from 18 subjects. The cortical surface atlas was then warped by MSD simulation to match the aggregate to the algebraic model of retinotopy. Finally, the algebraic model was projected back to the surface atlas and used to predict the polar angle organization of the left out subject.
Leave-one-out errors in the polar angle prediction were non-uniform across striate and extrastriate cortex (Figs. 3D and 3E). The highest errors are visible near the foveal confluence where all iso-angular lines converge, as well as in the dorsal region of V3 where V3 borders V3A. Although the median absolute leave-one-out error was uniformly low for a given predicted polar angle when aggregated across all three regions (Fig. 3E), errors in V3 were higher than those in V1 or V2, particularly close to the outer borders (Fig. 3D; Fig. S1). Overall, the median absolute and signed leave-one-out errors across all subjects and all vertices between observed and predicted polar angle were 10.93° and −0.48° respectively (Tab. 1). Additional reports and plots of the error in these predictions can be found in our Supplemental Mathematica Notebook (§5 and §6.3-6).
The quality of the polar angle predictions provided by the MSD approach may be compared to the prediction accuracy obtained using only the aggregate retinotopy data (similar to the approach of [8]). We calculated a mean-weighted average polar angle map for each subset of 18 of the 19 subjects in D10° and used each of these maps to predict the polar angle and eccentricity of the excluded subject. The median absolute leave-one-out polar angle error between all significant vertices of all subjects and the appropriate leave-one-out aggregate vertices was 23.27° (Tab. 1), twice as large as was obtained using the MSD approach. This indicates that the MSD approach serves as an informed regularization of noise that is present even in the average retinotopic mapping data from 18 subjects.
Finally, we examined how well the algebraic model of retinotopic organization, prior to spring registration to the aggregate data, predicts retinotopy in individuals. Again, the median absolute polar angle error of 34.12° was much greater than that obtained following MSD warping of the aggregate data to the algebraic model. This indicates that our approach corrects consequential distortions introduced by cortical flattening.
Fig. 4A presents the aggregate, confidence-weighted mean eccentricity of the 19 subjects in D10°. As with polar angle, the organization of eccentricity is consistent across subjects combined in the cortical surface atlas space, but sharp bends in the iso-eccentric contours, e.g. at the borders of V1 near 8–10° of eccentricity, do not match the properties of the algebraic model. The aggregate eccentricity, when warped to the algebraic model using MSD simulation, now follows the smooth lines of the idealized model (Fig. 4B). As with polar angle, the algebraic model may be projected back to the cortical surface atlas (Fig. 4C), to create an anatomical template of retinotopy which may be used to predict the retinotopic organization of novel subjects.
Median absolute leave-one-out errors for eccentricity were low across V1–V3 (Fig. 4D), with only slightly higher errors at greater eccentricities (Fig. 4E). Eccentricity error, unlike polar angle error, was uniform in V1, V2, and V3 (Fig. S2). The absolute and signed median leave-one-out errors for all subjects and vertices in D10° were 0.41° and 0.05° respectively (Tab. 1).
As was found for polar angle, simply using the aggregate polar angle data without MSD registration to the algebraic model resulted in substantially worse prediction accuracy for left-out subjects (median absolute error of 1.53°; Tab. 1). This was true as well for the attempt to predict eccentricity using the algebraic model but without MSD driven warping of the cortex (median absolute error of 2.44°).
The accuracy of polar angle and eccentricity prediction suggests that the algebraic model following MSD warping fits the retinotopic arrangement in regions V1–V3 well. This accuracy of prediction, however, does not necessarily indicate that the algebraic model is a good general representation of retinotopic organization. This is because MSD warping could in principle force the retinotopic data to match any locally smooth model which would then serve to regularize the data in the face of noise and thus improve prediction. While the “anatomical” springs used in the MSD simulation make an extreme warping to a very poor algebraic model implausible, an explicit test of the generalizability of the approach is desirable. If the algebraic model of retinotopy accurately describes the functional arrangement of the visual cortex, our approach should extrapolate to the prediction of eccentricity and polar angle in regions of visual cortex beyond the retinotopic mapping data.
To test the generality of the algebraic model and our template, we compared the anatomical template of retinotopy derived from D10° to the aggregate retinotopic mapping data from D20°, which consists of a separate set of subjects whose retinotopic maps were found using different techniques that doubled the mapped eccentricity range to 20° (see Methods).
For polar angle, the median absolute and signed errors between the measured and predicted value were 14.58° and 0.99° respectively. Note that these errors are comparable to those from the D10° leave-one-out analyses despite the fact that the D20° data extends beyond the 10° of eccentricity used to fit the model.
We next examined eccentricity prediction. Fig. 4F presents the median aggregate eccentricity map from D20° in the corrected cortical surface space found using D10° (the D20° aggregate in the original cortical atlas space is presented in the Supplemental Mathematica Notebook, §6.4.1). The overall median absolute error between vertices in D20° and the eccentricity template was 0.77°. Notably, this error is lower in the region from 1.25° to 8.75° (median absolute error: 0.59°) and higher in the region from 8.75°–18.75° (median absolute error: 2.33°). This suggests that our ability to fit extended data with our template is good but imperfect.
Our prediction error incorporates both the imperfections of our template as well as error in the measurement of retinotopy in the individual subject to be predicted. We have previously reported that the error in measured polar angle and eccentricity between two identical 20 minute retinotopic scans is ∼0.75° of eccentricity and ∼7.76° of polar angle in area V1 [11]. Similar statistics, using the new definition of region V1 we have derived here as well as the definition of V2 and V3, are given in Table 1 and plotted in Fig. S3. Measurement error grows from V1 to V3 as does prediction error. Notably, measurement error is actually greater than the prediction error of our anatomical template of retinotopy in all visual areas except for polar angle in V3.
We have described a technique to register functional data on the cortical surface to a 2D algebraic model of cortical organization. This approach allows us to predict the location and organization of visual areas V1–V3 in individual subjects based only upon an anatomical image of their brain. The overall prediction error for V1–V3 (10.93° of polar angle, 0.51° of eccentricity) is actually somewhat lower than the error we observed in our previous V1 template alone (11.43° of polar angle, 0.91° of eccentricity) [11]. We attribute this improvement to the correction of geometric distortions in the cortex introduced by flattening. Overall, we found that the accuracy of anatomically-based prediction of retinotopy in the individual subject is comparable to that provided by functional measurement itself.
In addition to good prediction accuracy, the anatomical template of V1–V3 retinotopy had generally small and uniform prediction bias. An exception to this property was found at the dorsal V3/V3A border, where our template consistently over-predicts the observed polar angle values (Fig. 3E). The error near the dorsal V3 border is substantially higher than that of any other region we studied (Fig. S1). The much smaller error found near the ventral V3/hV4 boundary suggests that the error is not due to a general inability of the approach to fit the outer boundaries of a model. Instead, we find that the dorsal V3 error can be understood as the effect of the V3A region extending into the V3 region during registration. Examination of the aggregate polar angle map (Fig. 3A, B) indicates that the predicted dorsal V3 boundary passes through a region of cortex that should be assigned by the template to V3A. This misalignment results from an attenuated polar angle reversal near the dorsal V3 border as compared to other reversals, which can be observed in Fig. 3A.
One possible explanation of this attenuated reversal is that it is the result of a poor across-subject anatomical registration due to variability in sulcal topology between subjects that could not be aligned. Such a problem would result in poorly aligned vertices and variable values contributing to the aggregate at this location. Examination of the sulcal curvature of individual subjects, however, does not support the idea of greater sulcal variability in this region (Fig. S4A).
An alternate explanation of the error near the dorsal V3 boundary is that individual differences in the mapping between structure and function create an area of relatively poor fit. Indeed, if we assume that the anatomical registrations provided by FreeSurfer are unbiased, there do appear to be significant differences in the location of the V3/V3A boundary between subjects (Fig. S4B). However, Fig. S3 shows a concentration of error in this same dorsal region, as well as near the foveal confluence, for the split-halves (test-retest) measurement error. Retinotopy in this region may simply be more difficult to measure.
It is entirely possible that the dorsal V3 border, and more generally the quality of the entire template, could be improved with modifications of our approach. We presented here a particular algebraic model of retinotopy [10] linked to the cortical surface with a particular deformation technique (MSD simulation). Neither of these choices are integral to the approach we describe. We selected the MSD approach as it provided an explicit means to balance maximizing registration of retinotopic values to the algebraic model against minimizing anatomical warping. Other approaches are certainly possible. In the Supplemental Mathematica Notebook we provide an example of an alternative registration method (see §1.8, Delaunay Mesh Registration, for implementation; §5.1.2 and §5.2.2 for error reports; and §6.2.3, §6.5.4-6, and §6.6.4-6 for figures).
More broadly, we consider the key insight of our work to be that geometric distortion of the flattened cortical surface limits the application of idealized models of cortical organization to empirical measurements of cortical function. These distortions, whether introduced by the developmental process of cortical folding or the digital process of cortical flattening, may be corrected by warping the cortical surface to bring function and model into alignment. Here, we demonstrated the practical value of this approach by creating an anatomical template of retinotopic organization. We expect that other early sensory areas such as the sensorimotor and auditory cortex, as well as higher level visual areas such as motion and face sensitive cortex, could be modeled using similar methods. This paper and its supplemental materials are intended as a guide for these kinds of studies.
|
10.1371/journal.pmed.1002756 | Potential effectiveness of prophylactic HPV immunization for men who have sex with men in the Netherlands: A multi-model approach | Men who have sex with men (MSM) are at high risk for anal cancer, primarily related to human papillomavirus genotype 16 (HPV16) infections. At 8.5 per 100,000 per year, the incidence rate of anal cancer among MSM is similar to that of cervical cancer among adult women in the Netherlands. However, MSM are not included in most HPV vaccination programs. We explored the potential effectiveness of prophylactic immunization in reducing anogenital HPV16 transmission among MSM in the Netherlands.
We developed a range of mathematical models for penile–anal HPV16 transmission, varying in sexual contact structure and natural history of infection, to provide robust and plausible predictions about the effectiveness of targeted vaccination. Models were informed by an observational cohort study among MSM in Amsterdam, 2010–2013. Parameters on sexual behavior and HPV16 infections were obtained by fitting the models to data from 461 HIV-negative study participants, considered representative of the local MSM population. We assumed 85% efficacy of vaccination against future HPV16 infections as reported for HIV-negative MSM, and age-specific uptake rates similar to those for hepatitis B vaccination among MSM in the Netherlands. Targeted vaccination was contrasted with vaccination of 12-year-old boys at 40% uptake in base-case scenarios, and we also considered the effectiveness of a combined strategy. Offering vaccine to MSM without age restrictions resulted in a model-averaged 27.3% reduction (90% prediction interval [PI] 11.9%–37.5%) in prevalence of anal HPV16 infections, assuming similar uptake among MSM as achieved for hepatitis B vaccination. The predicted reduction improved to 46.1% (90% PI 21.8%–62.4%) if uptake rates among MSM were doubled. The reductions in HPV16 infection prevalence were mostly achieved within 30 years of a targeted immunization campaign, during which they exceeded those induced by vaccinating 40% of preadolescent boys, if started simultaneously. The reduction in anal HPV16 prevalence amounted to 74.8% (90% PI 59.8%–93.0%) under a combined vaccination strategy. HPV16 prevalence reductions mostly exceeded vaccine coverage projections among MSM, illustrating the efficiency of prophylactic immunization even when the HPV vaccine is given after sexual debut. Mode of protection was identified as the key limitation to potential effectiveness of targeted vaccination, as the projected reductions were strongly reduced if we assumed no protection against future infections in recipients with prevalent infection or infection-derived immunity at the time of immunization. Unverified limitations of our study include the sparsity of data to inform the models, the omission of oral sex in transmission to the penile or anal site, and the restriction that our modeling results apply primarily to HIV-negative MSM.
Our findings suggest that targeted vaccination may generate considerable reductions in anogenital HPV16 infections among MSM, and has the potential to accelerate anal cancer prevention, especially when combined with sex-neutral vaccination in preadolescence.
| Anal and genital human papillomavirus (HPV) infections are sexually transmitted and may cause cancer in the anogenital area.
HPV vaccines protect against cancer by lowering the risk of getting infected with HPV, and are especially effective when given before becoming sexually active.
Men who have sex with men (MSM) are at high risk for anal cancer, but are not included in most HPV vaccination programs.
Decisions about their inclusion need to be informed by transmission models, but this is a challenge due to uncertainties regarding vaccine efficacy in those already exposed to HPV, and regarding HPV infection dynamics among MSM.
To give robust and plausible predictions about the effectiveness of targeted vaccination, we developed various models for HPV transmission among MSM that were parameterized using data from a Dutch cohort study.
We assessed the effectiveness of various vaccination strategies targeting MSM or 12-year-old boys or a combination thereof, and assuming vaccine uptake in targeted campaigns comparable to that of hepatitis B vaccine among MSM in the Netherlands.
In the models, targeted vaccination reduced the occurrence of anogenital HPV infections by around 30% after 40 years, with a range from 10% to 50%, depending on the recruitment of MSM into targeted campaigns and on the assumed mode of vaccine protection.
This figure increased to 75% after 60 years when targeted vaccination was combined with sex-neutral vaccination in preadolescence, assuming 40% uptake among 12-year-old boys and 85% efficacy against future HPV16 infections in MSM.
Our results are helpful for prioritizing male HPV vaccination, especially when deciding on the implementation of a selective campaign among MSM.
Offering HPV vaccine to sexually experienced MSM need not impede the efficiency of targeted vaccination, if vaccination protects against future HPV16 infections.
Targeted vaccination deserves consideration, at least temporarily, to protect adult MSM at high risk for anal cancer.
| Sexually transmitted oncogenic types of human papillomavirus (HPV) are known as the causative agents of cervical cancer [1–3]. They may also cause cancers in males, notably penile cancer, anal cancer, and a subset of head and neck cancers [2]. Relative to heterosexual males, men who have sex with men (MSM) are at increased risk for HPV-related cancers, especially for anal cancer [3]. With an estimated incidence of 8.5 per 100,000 per year, the incidence rate of anal cancer among MSM is similar to that of cervical cancer among adult women in the Netherlands [4].
In many countries, including the Netherlands, HPV-related disease prevention efforts are still entirely directed at females, through vaccination of preadolescent girls and screening for cervical cancer [2,4,5]. Over time, heterosexual males may receive indirect benefit from female vaccination through herd immunity, but MSM will not [6,7]. Vaccination of preadolescent boys along with girls can ultimately lead to control of HPV-related diseases in men and women alike, but might not constitute the most efficient use of resources [8–10]. Moreover, preadolescent vaccination will not protect currently active MSM, who might benefit from a targeted immunization campaign [11–14]. However, the effectiveness of selective vaccination targeting MSM past sexual debut could be hampered by prior exposure to HPV vaccine types [15,16].
Extrapolating the population-level effectiveness of selective vaccination from vaccine trials is difficult, for several reasons. First, it is not clear which estimates of vaccine efficacy to use; intention-to-treat estimates are difficult to apply outside the specific study settings (e.g., to age groups other than those included in a trial), whereas per-protocol estimates require a correct interpretation outside the protocol conditions. Additionally, the population-level effectiveness of a selective vaccination program targeting high-risk individuals strongly depends on the infrastructure by which vaccines can be delivered to this group, and the implications in terms of vaccine coverage by age. Finally, herd effects are expected to play a major role in determining the ultimate impact of targeted prevention efforts, and their assessment typically relies on mathematical modeling.
The purpose of this paper is to explore, by means of mathematical modeling, the potential effectiveness of a targeted immunization campaign among MSM in the Netherlands. We focus on reductions in anogenital HPV genotype 16 (HPV16) infections, as HPV16 causes the majority of anogenital cancers in males—i.e., around 85% of anal HPV-related cancers [17] and over 60% of penile HPV-related cancers [18]—and is included in all registered HPV vaccines [2]. To assess the temporal benefit of targeted vaccination, post-vaccination dynamics in HPV16 prevalence among MSM are contrasted to those induced by sex-neutral vaccination (i.e., vaccination of boys in addition to girls) in preadolescence.
We developed a range of mathematical models for penile–anal HPV16 transmission to assess the potential effectiveness of selective vaccination of MSM. Throughout we assumed that MSM acquire penile infections via insertive anal intercourse, whereas anal infections are acquired through receptive anal intercourse. The models differed in terms of sexual contact structure of the MSM population and assumptions regarding the natural history of HPV16 infection. Sexual contact parameters were estimated, whenever possible, from self-administered questionnaire data regarding sociodemographic characteristics and recent sexual behavior among 778 MSM (median age: 40 years, 5th–95th percentile 28–61 years) participating in the H2M study [19]. This is an observational cohort study on HIV and HPV infections in MSM recruited in Amsterdam in 2010–2011 and followed every 3–6 months for at least 2 years.
We constructed a deterministic dynamic population model of MSM by estimating age-specific rates of entering and exiting the population of individuals forming same-sex male partnerships (S1–S4 Figs). To account for penile-to-anal and anal-to-penile transmission, we incorporated sexual behavior by distinguishing insertive anal intercourse from receptive anal intercourse. The probabilities of engaging in either insertive, receptive, or both insertive and receptive anal sex within a partnership were obtained by fitting a mixture model to self-reported activities with anal sex partners in the last 6 months (S5 Fig). We modeled partner acquisition on the basis of self-reported numbers of anal sex partners in the last 6 months by HIV-negative H2M study participants, considered representative of the MSM population as validated by comparison to HIV incidence rates in the local community and an internet survey on sexual behavior among MSM throughout the Netherlands (S1 Text). To account for heterogeneity in partner acquisition rates in the model population, we considered 18 distinct settings of sexual contact structure for penile–anal HPV16 transmission, namely 3 distributions according to level of sexual activity (conditional on the age-specific mean and variance in age-specific contact rates) times 3 degrees of assortative mixing with respect to sexual activity times 2 degrees of assortative mixing with respect to preference for insertive/receptive anal sex (Table 1).
We stratified the model population into separate compartments by penile and anal HPV16 infection status. Individuals could be susceptible or infected at either or both anatomic sites separately, yielding a minimum of 4 compartments: {SS,SI,IS,II}, with SS denoting the proportion of the population susceptible for both penile and anal infection, SI denoting the proportion susceptible for penile infection while infected at the anal site, and so forth. We defined separate (age- and time-dependent) infection hazards for acquiring HPV16 infection at the penile site only, at the anal site only, and at both sites from the same partner. We distinguished between penile-to-anal transmissibility, β01, defined as the per-partnership probability of HPV16 transmission from the penis to the anus when engaging in insertive anal sex, and anal-to-penile transmissibility, β10, defined as the per-partnership probability of HPV16 transmission from the anus to the penis when engaging in receptive anal sex (S2 Text).
In order to provide robust and plausible predictions about the effectiveness of targeted vaccination in light of structural model uncertainties, we constructed several models for the natural history of penile and anal HPV16 infections. The simplest model included only the minimum of 4 compartments {SS,SI,IS,II}, with independent clearance of penile and anal infections at a constant rate γ10 and γ01, respectively. In a modified version, we considered separate compartments with persistent infections developing at a rate ζ10 and ζ01 and clearing at a rate ξ10 < γ10 and ξ01 < γ01 for penile and anal infections, respectively.
Further modifications were obtained by considering natural immunity or latency. For natural immunity, we considered both the possibility of systemic and local immunity, and in the latter case we also considered the options that immunity would only be induced at the penile or anal site. In addition, we considered separate scenarios for immunity following clearance in all instances, in 1:3 instances, or in 1:10 instances. In all scenarios, natural immunity could be lost at a constant rate κ, assumed similar for the penile and anal site in case of local immunity. Latency was incorporated in a similar fashion; either all, 1:3, or 1:10 incident infections would turn into latent infections, with the remainder becoming either susceptible again or systemically immune. Reactivation of latent infections was modeled at a rate ϱ, assumed to be similar for both anatomic sites.
Parameters related to HPV16 infection and transmission were obtained by fitting the models to HPV16 prevalence and clearance among the 461 H2M study participants who provided penile and anal samples and were HIV-negative for the entire follow-up [20]. An overview of the models used in prediction is given in Table 2, with mathematical descriptions in S2 Text. For each model, parameter estimates were obtained by an approximate maximum-likelihood procedure (S3 Text), consisting of separate optimization of progression and clearance parameters from longitudinal data, and conditional optimization of other parameters from site-specific HPV16 infection prevalence at study baseline.
In evaluating the potential effectiveness of targeted vaccination, we considered offering vaccine to MSM in the following age groups: ≤26 years (based on evidence from vaccine trials) [15,16], ≤40 years (recommended for selective vaccination of MSM in the UK) [13], and all ages (without an upper age for eligibility). In base-case analysis, we assumed age-specific uptake rates similar to those for hepatitis B (HepB) vaccine among MSM throughout the Netherlands [21]. In 2002, the Netherlands initiated a selective vaccination program targeting groups at high risk for HepB infection, including MSM. Because HepB vaccination was added to the childhood vaccination program in 2011, we restricted estimates of age-specific annual vaccination rate to estimated HepB vaccine uptake rates among 15- to 70-year-old MSM over the period up to 2010 (S4 Text). We also considered a scenario where HPV vaccine acceptance among MSM was double that of HepB vaccine, by using 2-fold increased age-specific uptake rates (S9 Fig). Effectiveness of targeted vaccination was contrasted with sex-neutral preadolescent vaccination by assuming 40% of MSM were vaccinated against HPV16 upon entrance into the sexually active population. The value 40% was based on the estimated uptake among boys in countries with sex-neutral HPV immunization programs [22]. In sensitivity analyses, we also examined a combined strategy of preadolescent and targeted vaccination under base-case assumptions—i.e., 40% uptake among boys and uptake among MSM similar to that for the HepB vaccine—and a scenario of 80% uptake among 12-year-old boys (S4 Text).
Prophylactic efficacy was taken from a quadrivalent HPV vaccine trial conducted among 16- to 26-year-old males [15]. We based our analysis on the 85.6% (97.5% CI 73.4–92.9) incidence rate reduction of infection detected for ≥6 months with vaccine-type HPV in the per-protocol population, consisting of participants who were seronegative on day 1 and PCR-negative from day 1 through month 7 for the relevant vaccine types. Prophylactic efficacy was incorporated in the transmission models by assuming that 85% of vaccinees became fully protected against future HPV16 infections and 15% were unaffected by vaccination. In the latter category, vaccine recipients remained fully susceptible if so at the time of immunization, or reverted back to susceptibility upon loss of natural immunity. This interpretation of vaccine efficacy by “take” rather than “degree” has also been used in assessing the population-level impact of sex-neutral vaccination by heterosexual HPV transmission models [10,23]. In sensitivity analysis, we considered the conservative scenario of restricted efficacy, where HPV16 infection hazards were reduced by 85%, i.e., “leaky” protection by degree, and only if vaccinees were fully susceptible at the time of immunization (Table 1). Following previous models, we assumed 98% efficacy in the scenario of preadolescent boys’ vaccination [9,10]. Note that vaccine efficacy against reactivation of latent infections was not assumed in any scenario.
For each vaccination scenario, we formed a model-averaged prediction of the reduction in anogenital HPV16 prevalence among MSM that may be achieved via prophylactic immunization. This started by calculating Akaike weights for each model under consideration [24], based on the relative quality of all candidate models with respect to H2M study data (S4 Text). As some natural history models can be viewed as a subset of more generic models (e.g., models with natural immunity converge to those without in case of short-lasting immunity), we employed upper bounds on parameter estimates for loss of immunity κ and reactivation rate ϱ in order to avoid duplicates in the set of candidate models. Models with negligible weight, i.e., without empirical support, were omitted from further consideration.
Eventually, we included 360 models (20 natural history models combined with 18 settings of sexual contact structure) per vaccination scenario in assessing the effectiveness of vaccination. We calculated the site-specific HPV16 prevalence prior to vaccination and its (relative) reduction, specifically after 25 years and at the post-vaccination equilibrium, and summarized results using the Akaike-weighted predictions with 90% prediction intervals (PIs), defined as the 5th–95th percentile range of the 360 models.
Patterns of site-specific HPV16 infection prevalence and clearance among HIV-negative H2M study participants were compatible with a range of mathematical models for penile–anal HPV16 transmission (Fig 1). Site-specific transmission probabilities varied widely across the models (Table 2), but penile-to-anal transmissibility almost invariably exceeded anal-to-penile transmissibility. Models that assumed a higher degree of natural immunity generally required increased transmissibility to reproduce the observed prevalence of penile and anal HPV16 infections. In addition, penile-to-anal transmissibility was higher when presuming a stronger degree of assortative mixing with respect to sexual activity (S7 Fig).
The relative quality of each model with respect to H2M study data was more dependent on the assumed natural history of HPV16 infection than on the sexual contact structure of the model (S8 Fig). In models that allowed for reactivation of latent infections, HPV16 prevalence mostly increased with increasing age, whereas HPV16 prevalence peaked around 40 years in models without latency (Fig 2). Likewise, models without latency predicted most sexually active HPV16-positive MSM to be in their 30s, whereas models with latency predicted this group to be somewhat older (S10 Fig). The model-averaged prevalence in the total MSM population prior to vaccination was 3.9% (90% PI 3.8%–4.1%) for penile HPV16 infection and 12.6% (90% PI 12.1%–13.1%) for anal HPV16 infection.
Offering vaccine to MSM aged ≤26 years achieved 9.4% vaccine coverage among MSM at the post-vaccination equilibrium in the base-case analysis (Table 3). This figure improved to 19.2% by extending vaccine eligibility to 40 years, and to 21.2% if the upper age for vaccine eligibility was discarded. Overall, the vaccine coverage among MSM achieved by targeted vaccination surpassed that of preadolescent boys’ vaccination in the first 12, 22, and 24 years of vaccination when offered to ≤26-year-old, ≤40-year-old, and all MSM, respectively, assuming similar HPV vaccine acceptance to that of HepB vaccine among MSM (Fig 3A). The combined strategy was projected to achieve 52.3% vaccine coverage among MSM. Adopting 2-fold increased uptake rates led to 17.6% vaccine coverage when vaccination was offered to MSM aged ≤26 years, 33.8% when offered until 40 years of age, and 36.5% if there was no age restriction. With doubled uptake, the vaccine coverage among MSM achieved by targeted vaccination surpassed that of preadolescent boys’ vaccination in the first 11, 20, and 22 years after initiating the vaccination strategy, respectively (Fig 3B).
With base-case uptake, offering vaccine to MSM aged ≤26 years resulted in a model-averaged reduction of 14.3% (90% PI 9.4%–18.8%) in equilibrium penile HPV16 infection, and a 13.4% reduction (90% PI 7.5%–17.8%) in equilibrium anal HPV16 infection, compared to the pre-vaccine prevalence of anogenital HPV16 infections among MSM. The predicted reductions improved to 27.2% (90% PI 15.3%–37.2%) and 25.5% (90% PI 11.6%–34.8%) in penile and anal HPV16 infections, respectively, if vaccine eligibility was extended to 40 years, and to 29.2% (90% PI 15.9%–40.2%) and 27.3% (90% PI 11.9%–37.5%), respectively, without an upper age for eligibility. HPV16 prevalence reductions in the post-vaccination equilibrium exceeded vaccine coverage projections (S11 Fig), and most of these reductions were realized within the first 30 years of a targeted immunization campaign, during which they exceeded those induced by vaccinating 40% of 12-year-old boys (Fig 4A). Moreover, the reductions in anogenital HPV16 prevalence from preadolescent boys’ vaccination were largely confined to younger MSM as compared to the reductions achieved by targeted vaccination (S12 Fig). However, the declines induced by preadolescent boys’ vaccination were sustained for much longer, ultimately leading to a 64.1% (90% PI 53.2%–79.7%) reduction in penile HPV16 infection among MSM, and a 61.6% (90% PI 48.4%–75.7%) reduction in anal HPV16 infection (S13 Fig). The combined strategy resulted in a 77.1% (90% PI 64.9%–95.2%) reduction in penile HPV16 infection and a 74.8% (90% PI 59.8%–93.0%) reduction in anal HPV16 infection among MSM (Table 3). Post-vaccination equilibria were achieved after 40 years in the case of targeted vaccination, but only after 60 years in scenarios involving preadolescent boys’ vaccination (S13–S15 Figs).
With doubled uptake, offering vaccine to MSM aged ≤26 years resulted in a model-averaged reduction of 26.2% (90% PI 17.4%–34.4%) in penile HPV16 infection and a 24.6% reduction (90% PI 14.1%–32.3%) in anal HPV16 infection. The corresponding reductions amounted to 46.2% (90% PI 27.4%–63.0%) and 43.6% (90% PI 21.4%–58.7%) in penile and anal HPV16 infections, respectively, if vaccine eligibility was extended to 40 years, and to 48.8% (90% PI 28.1%–66.9%) and 46.1% (90% PI 21.8%–62.4%), respectively, if no upper age for eligibility was considered. The reductions in anogenital HPV16 prevalence were sustained for over 40 years (S14 Fig), again exceeding vaccine coverage projections in the post-vaccination equilibrium (S11 Fig). Assuming 80% vaccine uptake among preadolescent boys resulted in the near elimination of anogenital HPV16 infections among MSM, with ≥90% reductions in anal HPV16 infection in 95% of model projections (Table 3). However, for the first 30 years of the vaccination strategies, the reductions induced by 80% preadolescent boys’ vaccination were smaller than those induced by vaccinating 40% of preadolescent boys in combination with offering selective vaccination to MSM without age restrictions, with base-case uptake (Fig 4B).
Targeted vaccination was only marginally effective if prophylactic efficacy was restricted to those fully susceptible at the time of immunization, while assuming similar uptake as realized for HepB vaccine. In this conservative scenario, offering vaccine to ≤26-year-old, ≤40-year-old, or all MSM resulted in penile HPV16 prevalence reductions of only 5.9% (90% PI 2.5%–9.9%), 10.4% (90% PI 3.9%–16.9%), and 11.1% (90% PI 4.2%–18.1%), respectively (Table 3). The corresponding reductions in anal HPV16 prevalence were 6.1% (90% PI 2.6%–9.5%), 11.0% (90% PI 4.8%–17.4%), and 11.7% (90% PI 5.1%–18.6%), respectively (S15 Fig). In these scenarios, HPV16 prevalence reductions remained below vaccine coverage projections (S11 Fig). The reductions induced by preadolescent boys’ vaccination were more robust, being only marginally reduced when vaccination provided “leaky” protection against infection.
Models with neither natural immunity nor latency predicted the highest reductions in anogenital HPV16 infections among MSM, whereas models with latency predicted the lowest reductions from a targeted immunization campaign (Figs 5 and S16). Likewise, the effectiveness of preadolescent boys’ vaccination in reducing HPV16 prevalence among MSM was weakest in models that assumed latency in combination with natural immunity. These differences were maintained under improved vaccine uptake, but reductions from targeted immunization became less dependent on latency assumptions when prophylactic efficacy was restricted to those fully susceptible (S17 Fig). Estimated reductions in HPV16 prevalence among MSM increased with sexual contact heterogeneity and decreased with assortative mixing, irrespective of vaccination scenario (Fig 5). The findings were similar with improved vaccine uptake and with restricted efficacy (S17 Fig).
This study explored the potential effectiveness of HPV vaccination for MSM in the Netherlands. Based on predictions from a range of penile–anal HPV16 transmission models, we estimated that around 30% of anogenital infections might be prevented after 40 years if uptake similar to that of HepB vaccine among MSM throughout the Netherlands were realized. This figure increased to 75% after 60 years when targeted vaccination was combined with sex-neutral vaccination in preadolescence, assuming 40% uptake among 12-year-old boys. HPV16 prevalence reductions among MSM mostly exceeded vaccine coverage projections, illustrating the efficiency of prophylactic immunization even when HPV vaccine is given after sexual debut.
Our work suggests that HPV vaccination could be effective when delivered to MSM utilizing the infrastructure available for targeted HepB vaccination. Our analysis also shows that, while vaccinating young MSM is important, inclusion of older MSM is likely needed to achieve substantial vaccine coverage and impact. The predicted reductions improved to around 50% with doubled vaccination uptake rates and no upper age for eligibility. This scenario represents a vaccine coverage projection that resembles the estimated HepB vaccination coverage in Amsterdam [26], where most MSM were actively recruited from specialist sexual health services and outreach locations such as saunas and gay bars. The scope for improved prevention is thus considerable, offering key opportunities for raising awareness about HPV-related cancer and promoting HPV vaccine acceptance among MSM [27,28].
Several caveats should be taken into consideration when assessing the potential impact of HPV vaccination of MSM. First, our work indicates that it may take several decades before reductions in HPV infection level are fully achieved, and another 15–30 years before the full impact on cancer incidence is reached [29]. The models may have overestimated the time scale at which the effects of vaccination become apparent if assortative mixing with respect to age is strong. However, data suggest that for MSM partnerships, age-assortative mixing is much less present than for heterosexual partnerships [30–34]. The modest and slow reductions in HPV16 prevalence are partly due to the assumption that uptake of HPV vaccine among MSM would resemble that of HepB vaccine, where it took more than a decade before the effects on transmission and incidence could be demonstrated [26]. In addition, HPV16 is characterized by a relatively high reproduction potential as compared to other vaccine-protected HPV types [35]. Consequently, one should expect modest herd effects from vaccinating against HPV16 in comparison to other HPV types, as demonstrated in a community-randomized HPV vaccine trial [36]. An Australian study also predicted a long duration for targeted vaccination effects to become fully apparent in vaccine-type HPV prevalence among MSM [14]. This long duration may influence willingness to participate in a selective vaccination program, and will also negatively affect the cost-effectiveness profile of targeted vaccination. Both issues may be alleviated by vaccine inclusion of low-risk HPV types associated with anogenital warts, as vaccination has been shown to induce rapid declines in wart incidence [6,7]. Presumably, the favorable cost-effectiveness profile of selectively vaccinating MSM in the UK was driven by the inclusion of anogenital wart prevention, as the favorable profile only applied to the use of quadrivalent HPV vaccine (including low-risk types 6 and 11), and not to HPV16/18 vaccination [13].
Second, the validity of the base-case scenarios strongly depends on the assumption of 85% efficacy against future HPV16 infections, irrespective of HPV16 infection status at the time of immunization. Thus, we assumed that prophylactic efficacy would also apply to recipients already infected with (or immune to) HPV16 at the time of immunization. While conceivable, this assumption has yet to be tested empirically. The effectiveness of targeted vaccination is profoundly reduced if prophylactic efficacy applies only when vaccine recipients are fully susceptible at the time of immunization, as in per-protocol analyses of vaccine trials [15,16]. Yet, HPV vaccine has demonstrated high efficacy and immunogenicity in adult women 24–45 years of age, regardless of previous exposure to HPV vaccine type [37], likely making this latter scenario overly conservative.
Third, the effectiveness of targeted vaccination varied considerably between the models included in the analysis, with lower reductions predicted in models that assumed naturally acquired immunity or reactivation of latent infections. As the mechanisms of immunity and latency become better understood [38,39], models will need to be revised to adequately capture the interactions of vaccine-induced protection and naturally acquired immunity or latency. Likewise, more data on the occurrence, acquisition, and duration of anogenital HPV infections in MSM would help to narrow down the range of transmission models compatible with data, and increase the precision of model-averaged prediction. In the meantime, well-calibrated dynamic models should be equipped to simulate and explore various assumptions around age specificity in HPV16 infection prevalence. A multi-modeling approach [24], as employed in this analysis, is valuable when evaluating the impact of an intervention in light of many structural model uncertainties.
Our dynamic model is the first to our knowledge to explicitly incorporate site-specific HPV infection and transmission among MSM. Previous models of HPV transmission in MSM remained ambiguous about the routes of transmission being considered, and circumvented the need to explicate site-specific transmissibility by considering general transmission probabilities in same-sex partnerships [13,14]. While such an approach greatly reduces model complexity, it goes at the expense of essential detail as the risks of penile and anal HPV infections are mediated by different behaviors (i.e., insertive versus receptive anal intercourse). Moreover, it has been suggested that probabilities of HPV transmission from the penis to the anus are likely to be significantly higher than those from the anus to the penis [40]. This supposition is borne out by our analysis. The overall reproduction number of penile–anal transmission is a composite of both site-specific reproduction numbers, analogous to transmission from men to women and back to men [41]. Therefore, leaving site-specific transmissibility unspecified could lead to biased predictions about the prevention of anogenital HPV16 infections in MSM through prophylactic immunization.
We did not consider condom use in our HPV16 transmission models. While some studies have found that consistent condom use may provide some degree of protection against HPV infection among high-risk men [42], systematic reviews provide no consistent evidence that condom use reduces the risk of becoming HPV DNA-positive [43,44]. Likewise, we found no clear relation in the H2M study between condom use during anal sex in the preceding 6 months and transition from uninfected to infected HPV16 or HPV18 states [45]. Nevertheless, our estimates of site-specific transmissibility could be biased as they are not adjusted for condom use. Moreover, our projections are sensitive to changes in sexual risk behavior, including the possibility of decreased condom use, e.g., in response to the introduction of HIV pre-exposure prophylaxis in the Netherlands [46]. Indirect effects on HPV transmission due to factors unrelated to HPV vaccination were beyond the scope of this study, but should be considered before planning selective vaccination of MSM.
We also did not incorporate transmission to the penile or anal site via oral sex. A recent study concluded that traditional heterosexual HPV transmission models (concerned with cervicovaginal and penile infections) may underestimate the population-level effectiveness of vaccination if a high proportion of genital infections originate from extragenital sites, whereas vaccination effectiveness may be overestimated if natural immunity to genital infections can occur following extragenital infections [47]. Currently, there is no strong evidence that oral infections are a reservoir for anogenital infections in MSM, or that clearance of oral infections can induce systemic immunity [48]. Moreover, the prevalence of oral HPV16 infection was below 2% in the HIV-negative H2M study participants, with a 5-fold lower incidence of oral HPV16 infection compared to anogenital HPV16 infection [49]. Hence, a substantial bias due to using only penile–anal transmission models to assess the potential effectiveness of HPV16 vaccination of MSM is not likely.
Finally, we did not incorporate the direct effect of HIV on HPV transmission dynamics. HIV is a strong and independent determinant of penile and anal HPV infections [3,20,49], but likely has a limited role in driving HPV transmission given the ubiquity of HPV infections among MSM in the Netherlands and the comparatively low prevalence of HIV. Nevertheless, HIV is one of the strongest risk factors for anal cancer, suggesting an important role of HIV in disease progression [50]. The policy in the UK of vaccinating HIV-positive MSM up to age 45 years against HPV [51], informed by a health economic study [13], is challenged by recent empirical data on the lack of efficacy in HIV-positive MSM aged 27 years or older [52]. Our modeling results apply primarily to HIV-negative MSM, for whom the HPV vaccine has shown prophylactic efficacy in preventing type-specific infections [16], genital warts [53], and lesion recurrence [54]. More research is needed on the role of HIV in HPV-induced neoplasia before more detailed predictions can be made with regard to anal cancer prevention. In any case, inclusion of MSM at high risk for HIV is paramount to achieve meaningful impact in selective vaccination programs.
The comparison of targeted vaccination with sex-neutral vaccination in preadolescence serves to illustrate the temporal benefit derived from a targeted immunization campaign among MSM. In addition, the comparison also serves as a benchmark to judge the potential effectiveness of targeted vaccination. The result, that vaccinating preadolescent boys would ultimately be more effective than offering HPV vaccine to MSM, partly depends on the supposedly moderate uptake of HPV vaccine among MSM in the Netherlands. However, even a targeted campaign that could reach a similar proportion of MSM as preadolescent vaccination would not be as effective, given the reduced efficacy of HPV vaccine when given after sexual debut. Therefore, the benefit of targeting interventions to MSM lies in a faster effectiveness of HPV vaccination regarding cancer prevention in males, and—possibly—in a lower number needed to vaccinate to prevent disease in males [9]. Besides, a sizeable proportion of the Dutch MSM population is not born in the Netherlands and may be missed by preadolescent vaccination.
Our work suggests that selective vaccination of MSM would be especially effective when combined with sex-neutral vaccination in preadolescence. Even a moderate uptake among preadolescent boys would already generate a substantially increased yield when combined with targeted vaccination, as it would render MSM immune against HPV16 upon entrance into the sexually active population. Conversely, a combined strategy could also safeguard against disappointing or unstable uptake in preadolescent vaccination programs, in the sense that it may be more feasible and sustainable to reach 40% of 12-year-old boys and MSM through a targeted campaign than it is to achieve 80% uptake among 12-year-old boys. While vaccinating 80% of preadolescent boys would be needed to achieve near elimination of HPV16 among MSM, our findings suggest stronger effectiveness from a combined strategy of vaccinating 40% of boys in preadolescence together with a targeted campaign among MSM for the first 30 years of vaccination.
In conclusion, this study suggests that a targeted immunization campaign among MSM in the Netherlands may generate considerable reductions in anogenital HPV16 infections in a high-risk population. Sex-neutral vaccination in preadolescence is likely needed to eliminate HPV-related diseases as a public health problem in men and women alike, but targeted vaccination deserves consideration, at least temporarily, to protect currently active MSM at high risk for anal cancer.
|
10.1371/journal.pgen.1007947 | TDP-43 induces mitochondrial damage and activates the mitochondrial unfolded protein response | Mutations in or dys-regulation of the TDP-43 gene have been associated with TDP-43 proteinopathy, a spectrum of neurodegenerative diseases including Frontotemporal Lobar Degeneration (FTLD) and Amyotrophic Lateral Sclerosis (ALS). The underlying molecular and cellular defects, however, remain unclear. Here, we report a systematic study combining analyses of patient brain samples with cellular and animal models for TDP-43 proteinopathy. Electron microscopy (EM) analyses of patient samples revealed prominent mitochondrial impairment, including abnormal cristae and a loss of cristae; these ultrastructural changes were consistently observed in both cellular and animal models of TDP-43 proteinopathy. In these models, increased TDP-43 expression induced mitochondrial dysfunction, including decreased mitochondrial membrane potential and elevated production of reactive oxygen species (ROS). TDP-43 expression suppressed mitochondrial complex I activity and reduced mitochondrial ATP synthesis. Importantly, TDP-43 activated the mitochondrial unfolded protein response (UPRmt) in both cellular and animal models. Down-regulating mitochondrial protease LonP1 increased mitochondrial TDP-43 levels and exacerbated TDP-43-induced mitochondrial damage as well as neurodegeneration. Together, our results demonstrate that TDP-43 induced mitochondrial impairment is a critical aspect in TDP-43 proteinopathy. Our work has not only uncovered a previously unknown role of LonP1 in regulating mitochondrial TDP-43 levels, but also advanced our understanding of the pathogenic mechanisms for TDP-43 proteinopathy. Our study suggests that blocking or reversing mitochondrial damage may provide a potential therapeutic approach to these devastating diseases.
| TDP-43 proteinopathy is a group of fatal neurological diseases. Here, we report a systematic examination of the role of mitochondrial damage in TDP-43 proteinopathy using patient brain tissues, as well as cellular and animal models. Our data show that TDP-43 induces severe mitochondrial damage, accompanied by activation of UPRmt in both cellular and animal models of TDP-43 proteinopathy. LonP1, one of the key mitochondrial proteases in UPRmt, protects against TDP-43 induced cytotoxicity and neurodegeneration. Our study uncovers LonP1 as a modifier gene for TDP-43 proteinopathy and suggests protecting against or reversing mitochondrial damage as a potential therapeutic approach to these neurodegenerative disorders.
| TDP-43 proteinopathy is characterized by the presence of TDP-43 immunoreactive inclusion bodies in the affected tissues. Clinically, TDP-43 proteinopathy manifests as a spectrum of different neurodegenerative diseases, ranging from dementia (especially fronto-temporal lobar degeneration, FTLD) and motor neuron disease (MND) to traumatic brain injuries [1–4]. FTLD is a prevalent form of dementia with progressive atrophy of the frontal and/or temporal cortices [5–7]. Amyotrophic Lateral Sclerosis (ALS), a common form of MND, is characterized by a progressive loss of upper and lower motor neurons [8–10]. TDP-43 associated neurodegenerative diseases are clinically and genetically heterogeneous. A significant fraction of ALS patients exhibit cognitive impairment [11,12]; and ~15% of FTLD patients also show locomotor defects and meet the diagnostic criteria for ALS [12,13]. TDP-43-positive lesions are the most frequently identified pathology among FTLD and ALS cases and also present in ~50% AD samples [14–16]. However, the pathogenic mechanisms underlying TDP-43 proteinopathy remain unclear.
Mitochondrial damage is associated with a range of neurodegenerative diseases, including Alzheimer’s disease (AD), Parkinson’s disease (PD) and MNDs [17–19]. Mitochondrial changes have been detected in cellular and animal models for TDP-43 proteinopathy [16,20–27]. It was recently reported that suppressing mitochondrial localization of TDP-43 blocked TDP-43 neurotoxicity [28]. However, mitochondrial morphological changes have not yet been characterized in patient samples, and the effects of TDP-43 on mitochondrial function remain controversial [27–29].
To maintain mitochondrial homeostasis, cells sense and respond to mitochondrial damage by activating a program known as the mitochondrial unfolded protein response (UPRmt), which includes induction of mitochondrial chaperones assisting in proper protein folding, and of proteases promoting clearance of misfolded proteins [30–32]. Recent studies suggest a role of UPRmt in Alzheimer’s disease, Parkinson’s disease and ALS-SOD [33–35]. However, the role of UPRmt in TDP-43 proteinopathy has not been reported.
Here, we present a systematic study of TDP-43 proteinopathy combining cellular and animal models with patient samples. Analyses using electron microscopy (EM) reveal prominent mitochondrial damage in brain tissues from TDP-43 proteinopathy patients. These mitochondrial impairments include swollen and degenerated cristae or a complete loss of cristae. Similar mitochondrial cristae changes are detected in our cellular and animal models. Consistently, mitochondrial functional impairments are observed, including decreased mitochondrial membrane potential, reduced mitochondrial ATP synthesis and elevated mitochondrial ROS production. Our data show that mitochondrial impairment induced by TDP-43 is an early event, preceding cell death. Furthermore, induced TDP-43 expression leads to the activation of UPRmt in both cellular and fly models for TDP-43 proteinopathy. LonP1, one of the key mitochondrial proteases in UPRmt, plays an important role in the degradation of mitochondrial TDP-43. Consistent with the mRNA changes of LonP1 in cellular and fly models, LonP1 protein levels are increased in a fraction of the brain samples of patients affected by FTLD-TDP. Importantly, down-regulation of LonP1 in TDP-43 expressing flies not only induces more severe mitochondrial damage, but also advances disease onset and exacerbates the neurodegeneration phenotype in the animal model. These results suggest that LonP1 plays a protective role against TDP-43-induced neurotoxicity, especially at an early stage of the disease. Together, our data demonstrate that mitochondrial damage is a critical feature of TDP-43 proteinopathy and suggest that protecting mitochondria may have therapeutic potential.
To investigate the role of mitochondria in TDP-43 proteinopathy, we examined mitochondrial morphology in brain samples from patients using transmission electron microscopy (TEM) and immuno-electron microscopy (IEM). Following resin-embedding to obtain clear images of mitochondria, we analyzed brain samples from five patients with the pathological diagnosis of either FTLD-TDP or ALS-FTLD-TDP, together with the samples from three control subjects without any TDP-43 pathology (for details, see S1 Table).
The majority of mitochondria in the control brain tissues showed normal morphology, with intact mitochondrial membrane and well-organized cristae (left panels in Fig 1A). In contrast, more than 80% of mitochondria in the patient brains exhibited significant mitochondrial damage, especially abnormal cristae structure (Fig 1). Abnormal mitochondrial cristae presented as either a “vesicular” type with swollen cristae (marked by the arrows in the middle panels of Fig 1A; as “Swollen” in Fig 1A and Fig 1B) or a “degenerated” type with a partial to complete loss of cristae (the right panels of Fig 1A; as “Degenerated” in Fig 1A and Fig 1B). Damaged mitochondria were significantly increased in all 5 FTLD-TDP brains as compared with the control brains (Fig 1B).
IEM analyses of the brain tissues using a specific anti-TDP-43 antibody revealed that TDP-43 immunostaining signals were clearly detected inside mitochondria in the brain samples of both control and FTLD-TDP patients (marked by arrows in Fig 1C; with enlarged views in insets), demonstrating that the endogenous TDP-43 protein is localized inside mitochondria, consistent with a recent report [28]. Interestingly, electron-dense TDP-43 positive protein aggregates were detected inside ~1% of mitochondria in FTLD-TDP patient samples (arrowheads in Fig 1D), but were not detected in any control samples. These EM analyses demonstrate that mitochondrial damage is a prominent feature in the pathology of brain tissues of TDP-43 proteinopathy patients.
To investigate the effects of TDP-43 on mitochondrial morphology and function in living cells, we established tetracycline (Tet) inducible HEK293 cell lines, expressing either wild type (Wt) or an ALS-associated TDP-43 mutant (A315T). Following Tet-induction for 24 hr, total cell lysates, cytoplasmic fractions and purified mitochondrial preparations were examined by Western blotting. The purity of the mitochondrial preparation was confirmed by the detection of mitochondrial protein TOM20 and the absence of the cytoplasmic GAPDH protein. Consistent with the IEM data from the human brain samples, the endogenous TDP-43 as well as the exogenously expressed Wt or ALS-mutant (A315T) TDP-43 were detected in purified mitochondria (Fig 2A; for a longer exposure, see S1A Fig), supporting the mitochondrial localization of the TDP-43 protein. Consistent with previous studies [36,37], expression of the exogenous TDP-43 suppressed expression of the endogenous TDP-43 (marked by “Endo” in Fig 2A).
We next performed EM analyses of HEK293 cells expressing TDP-43 to characterize mitochondrial changes. In control cells, the vast majority of mitochondria exhibited normal morphology, with well-organized cristae (Fig 2B). However, in cells expressing the A315T-mutant TDP-43, severe mitochondrial damage was detected, with significantly reduced mitochondrial sizes and impaired mitochondrial cristae 24 hr post-induction. When Wt TDP-43 was expressed, similar mitochondrial damage was also detected, although to a lesser extent (Fig 2B; see S1B and S1C Fig). These data indicate that expression of Wt or ALS-mutant TDP-43 protein leads to mitochondrial damage in cultured cells.
To examine the temporal relationship between TDP-43-induced mitochondrial damage and cell death, we carried out a series of experiments using the Tet-inducible cells expressing Wt- or A315T-mutant TDP-43 proteins at different time points (0, 24 or 36 hr) following induction of TDP-43 expression. We first measured mitochondrial membrane potential, ROS production and ATP synthesis (Fig 2C–2E). Cells were stained with JC1 (a mitochondrial membrane potential indicator), or mitoSOX red fluorescent dye (a mitochondrial ROS indicator), and then analyzed by flow cytometry. Mitochondrial membrane potential began to show a reduction at 24 hr post-induction in cells expressing A315T-mutant TDP-43; and by 36 hr post-induction, mitochondrial membrane potential reduction was detected in cells expressing either Wt or A315T-mutant TDP-43 (Fig 2C). By 36 hr following the induction of expression of Wt or A315T-mutant TDP-43, the mitochondrial ROS level was significantly increased (Fig 2D). Total cellular ATP levels and mitochondrial ATP synthesis were measured following published protocols [38,39]. Thirty-six hr following induction of TDP-43 expression (Fig 2E), total cellular ATP level did not change (S2A Fig). However, mitochondrial ATP synthesis at this time point was significantly reduced in cells expressing either Wt or A315T-mutant TDP-43 (with ~20% and ~25% decrease in the Wt and A315T-groups respectively), as compared with the control group (Fig 2E).
To understand the mechanism by which increased TDP-43 expression suppressed mitochondrial ATP synthesis, we examined which mitochondrial complexes (complex I through V) in oxidative phosphorylation were affected. Interestingly, complex I activity was significantly reduced by 24 hr following induction of either Wt or A315T mutant TDP-43 (Fig 2F); complex IV activity was also reduced by the expression of A315T-mutant TDP-43 (Fig 2I). In contrast, the activities of complexes II, III (Fig 2G, Fig 2H) and complex V (Fig 2J) were unaffected. These data indicate that increased TDP-43 expression impairs mitochondrial ATP synthesis, possibly by suppression of mitochondrial complex I. TDP-43-induced reduction in the complex I activity was not likely the result of overall suppression of complex I genes by TDP-43, because quantitative PCR analyses of a number of complex I genes did not show a general reduction in the expression of these genes (see S2B Fig). Future experiments are necessary to elucidate the mechanism by which TDP-43 suppresses the activity of complex I.
To examine cell death, cells were stained with an Annexin V-FITC/PI (propidium idodide) kit followed by flow cytometry analyses (Fig 3). Annexin V-positive/PI-negative, Annexin V-negative/PI-positive or Annexin V-positive/PI-positive staining indicates apoptosis, necroptosis or late apoptosis/necroptosis, respectively. Up to 36 hr post-induction, Annexin V-negative/PI-positive or PI/Annexin V double-postive cell populations did not show significant changes in TDP-43 expressing cells compared to the control group. Cells expressing A315T-mutant TDP-43 showed significantly increased cell death only after 36 hr post-induction of TDP-43 expression (~2% cells showing Annexin V-positive/PI-negative staining; compared with ~0.5% in the control cells); whereas cells expressing wild type TDP-43 showed a less dramatic increase in cell death, also only after 36 hr post-induction (Fig 3A, Fig 3B). It should be noted that at this time point only a small fraction (<5%; estimated by biochemical fractionation) of the total TDP-43 was detected in purified mitochondria (possibly due to the efficient degradation of mitochondrial TDP-43 before the disruption of the balanced mitochondrial proteostasis). Because mitochondrial dysfunction was observed at the 24 hr time point, these results demonstrate that TDP-43-induced mitochondrial dysfunction is an early event preceding cell death, suggesting that mitochondrial impairment may contribute to TDP-43 cytotoxicity.
To investigate TDP-43-induced mitochondrial damage in vivo, we examined transgenic flies expressing either Wt or A315T-mutant TDP-43 reported in our previous studies [40–42]. Transmission EM analyses of control fly eyes in 3-day old adult animals revealed intact ommatidial structures with seven rhabdomeres, whereas expression of either Wt or A315T-mutant TDP-43 in fly eyes led to severe ommatidial defects, often with a complete loss of rhabdomeres (Fig 4A). Mitochondria in fly eyes expressing either Wt or A315T-mutant TDP-43 showed a significant decrease in size when compared with control flies (Fig 4). Importantly, more than 85% of mitochondria in the photoreceptors expressing Wt or ALS-mutant TDP-43 exhibited swollen or vesicular cristae, whereas only ~5% of mitochondria in the control group showed damage (Fig 4A and Fig 4C). In this setting, TDP-43 was expressed in photoreceptors under a strong GMR-Gal4 driver from an early stage, leading to rapid and severe mitochondrial damage. By the time of EM examination, >85% mitochondria showed damage in both Wt and A315T-mutant groups, not allowing us to detect differences between the two groups. It is remarkable that mitochondria in fly photoreceptors expressing either Wt or A315T-mutant TDP-43 showed similar mitochondrial cristae damage as those detected in the brain tissues of TDP-43 proteinopathy patients (see Fig 1A). To examine whether the results observed were due to developmental defect(s), we used a system in which TDP-43 expression was induced only in adulthood using a temperature-sensitive tubulin-Gal80ts promoter with the GMR-Gal4 photoreceptor-specific driver or the Elav-Gal4 pan-neuronal driver (see S3 Fig). In this system, flies expressing A315T-mutant TDP-43 in photoreceptors following heat shock induction at the adult stage indeed exhibited progressive mitochondrial damage and retinal degeneration (S3 Fig).
The mitochondrion is a major source for the production of reactive oxygen species (ROS) [43]. Mitochondrial dysfunction can lead to the accumulation of ROS [44]. Furthermore, excessive ROS production affects neuronal survival and function [45,46]. We therefore examined whether TDP-43 expression affected mitochondrial ROS production in vivo using transgenic flies expressing TDP-43 in motor neurons. A fly line expressing mito-roGFP-Grx1, an in vivo mitochondrial ROS reporter [47], was crossed with either control RFP or TDP-43-RFP expressing flies. Ratiometric fluorescence confocal imaging was carried out to measure mitochondrial ROS levels in motor neurons expressing control (RFP) or TDP-43-RFP using a previously published protocol [47]. Significantly elevated mitochondrial ROS levels were detected in motor neurons expressing either Wt- or A315T-mutant TDP-43 as compared with the control group (see S4 Fig), indicating that TDP-43 expression in motor neurons resulted in mitochondrial dysfunction.
Our results presented above showed that increased TDP-43 expression led to mitochondrial cristae damage, reduced activities of mitochondrial OXPHOS complex I and IV, as well as decreased mitochondrial ATP synthesis. In addition, TDP-43 immuno-reactive aggregates were detected inside mitochondria of FTLD-TDP patient brain samples. These observations prompted us to examine if TDP-43 activated the mitochondrial unfolded protein response (UPRmt).
Using our inducible HEK293 cells expressing Wt or A315T-mutant TDP-43, we examined mRNA levels of known genes critical for UPRmt, including ATF5, HSPA9 (mtHSP70), HSP60 and LonP1. Quantitative RT-PCR analyses revealed that by 48 hr post-induction of TDP-43 expression, mRNA levels of ATF5 and LonP1 were increased, and that by 72 hr post-induction, mRNA levels of ATF5, HSPA9, HSP60 and LonP1 were all increased in cells expressing either Wt- or A315T-mutant TDP-43 (Fig 5A). To investigate whether TDP-43 expression activated UPRmt in vivo, we induced TDP-43 expression in transgenic flies at the adult stage by heat shock using Elav-Gal4 pan-neuronal driver containing a temperature-sensitive tubulin-Gal80ts element, Elav-Gal4/tubulin-Gal80ts driver [48] (see S3A Fig). At day 15 and day 30 after induction of TDP-43 expression, fly heads were collected for qRT-PCR analyses (Fig 5B). In female flies, by day 15 post-induction, HSP60A mRNA level was significantly increased in A315T-mutant expressing flies; and by day 30 post-induction, mRNA levels of HSP60A, Hsc-70-5, CG5045 (encoding ClpP) and two isoforms of Lon (the Drosophila ortholog of mammalian LonP1) were increased in TDP-43 expressing flies, especially those expressing A315T-mutant TDP-43. In male flies, the mRNA levels of all four genes were increased in flies expressing A315T-mutant TDP-43, and to a lesser extent in flies expressing Wt TDP-43, at 15 day post-induction. However, increased expression of only HSP60A, but not other three genes, was detected by day 30 post-induction of TDP-43 expression (Fig 5B). These data support that UPRmt is activated by TDP-43 expression in the fly model for TDP-43 proteinopathy. Future studies are necessary to understand the significance of and mechanisms underlying the gender different responses observed in TDP-43 flies.
We next examined if protein levels of these UPRmt genes are altered in TDP-43 proteinopathy patient samples using a panel of brain samples characterized previously [40]. Western blotting analyses indicate that the average level of LonP1 protein in TDP-43 proteinopathy patient brains was higher than that in the control brains (Fig 5C, Fig 5D). This is consistent with the possibility that UPRmt may be activated in a subset of TDP-43 proteinopathy patient brains. There was no significant difference between patient and control samples in the protein levels of either HSPA9 or HSP60. Together, these results support the notion that UPRmt is activated in cellular and animal models of TDP-43 proteinopathy as well as a subset of FTLD-TDP patient brains.
We further examined the relationship between LonP1 and TDP-43. LonP1 is a major mitochondrial matrix protease and a member of the evolutionarily conserved superfamily of AAA+ ATPases. LonP1 plays a critical role in mitochondrial protein quality control by preferentially degrading misfolded or oxidized proteins [49]. We first tested whether TDP-43 interacted with LonP1 in a co-immunoprecipitation assay using an anti-Myc antibody in cells expressing Myc-tagged TDP-43. LonP1 was detected among immunoprecipitated proteins from cell lysates expressing either Wt or A315T-mutant TDP-43, but not the control lysates (Fig 6A), suggesting that LonP1 interacted with TDP-43. Further co-immunoprecipitation experiments using a specific TDP-43 antibody showed that the endogenous TDP-43 and LonP1 proteins interacted with each other (Fig 6B). To examine if TDP-43 protein co-localized with LonP1 inside mitochondria, we performed immuno-electron microscopy (IEM) using FTLD-TDP brain samples. In these brain samples, TDP-43 immuno-staining signals (6-nm gold particles) were detected in close proximity to LonP1 immuno-staining signals (15-nm gold particles) (marked by the arrowheads in Fig 6C).
A number of studies suggest the roles of proteasome and autophagy in degradation of TDP-43 [50–57]. We then tested the effects of a proteasome inhibitor (MG132, MG) and an autophagy inhibitor (3-methyladenine, MA), and compared them with that of a LonP1 inhibitor [2-cyano-3,12-dioxooleana-1,9-dien-28-oicacid, CDDO (CD) [58] ] in the inducible TDP-43 expressing cells. Interestingly, neither the proteasome inhibitor (MG) nor the autophagy inhibitor (MA) had an effect on cell viability following induction of TDP-43 expression, whereas the LonP1 inhibitor (CD) specifically reduced the viability of cells expressing either Wt or A315T-mutant TDP-43 and enhanced TDP-43 cytotoxicity (see S5A and S5B Fig). At the concentrations used, none of these drugs affected viability of the control cells, indicating that the effect of the LonP1 inhibitor was specifically associated with TDP-43 expression (Fig 7A; S5A and S5B Fig).
We next examined whether increasing LonP1 expression suppressed TDP-43 cytotoxicity. Control (Ctr) or TDP-43 expressing cells were transfected with a vector control (-) or a LonP1-expressing plasmid (+) 24hr before Tet-induction; and cells were examined 36 hr post-induction. Increased LonP1 expression suppressed TDP-43 induced cytotoxicity (Fig 7B). Quantification of Western blotting (WB) signals showed a ~2-fold increase in LonP1 expression, as normalized by actin levels. The total TDP-43 levels did not show significant changes (see S5C Fig), which is not unexpected because TDP-43 protein is predominantly nuclear, although it is the cytoplasmic/mitochondrial levels of TDP-43 that are correlated with neurotoxicity, as shown by published studies including ours [28,59].
We further tested whether down-regulating LonP1 altered TDP-43 induced cytotoxicity. TDP-43 inducible stable cells were transduced with a vector control virus (Ctr) or a lentivirus expressing shRNA specifically targeting LonP1 (KD) that reduced the LonP1 protein level by ~50%. LonP1 knockdown (KD) significantly reduced the viability in cells expressing TDP-43 (Fig 7C). Fractionation experiments demonstrated that LonP1 down-regulation led to an increase in mitochondrial TDP-43 protein level in these cells although the cytosolic levels of TDP-43 were not dramatically affected (Fig 7D), indicating that LonP1 decreases the mitochondrial TDP-43 protein level. To test whether TDP-43 could be directly degraded by LonP1, we established an in vitro protein degradation assay using purified recombinant LonP1 protein. Our data demonstrated that purified Wt or A315T TDP-43 protein was degraded by the purified recombinant LonP1 protein in a manner dependent on LonP1 concentrations (Fig 7E) and dependent on ATP (see S5D Fig).
A number of other mitochondrial proteases are involved in mitochondrial proteostasis. The mRNA level of CG5045, the Drosophila homolog of ClpP, was also increased in transgenic TDP-43 flies (see Fig 5B). We thus examined if TDP-43 also interacted with ClpP. However, no detectable interaction between ClpP and TDP-43 was observed in a co-immunoprecipitation assay (supplemental S6A Fig). Consistently, down-regulation of ClpP did not affect the mitochondrial TDP-43 level, as shown by WB analyses of purified mitochondria from cells following ClpP knockdown (supplemental S6B Fig). Together, these data show that LonP1 reduces TDP-43-induced cytotoxicity, possibly by degrading mitochondrial TDP-43 protein.
To investigate whether altering Lon expression in vivo would modify neurodegeneration induced by TDP-43, we obtained fly lines over-expressing the Drosophila LonP1 ortholog, Lon, or expressing specific siRNA against Lon. Only one fly line overexpressing Lon was available, and it showed ~2-fold increase in Lon mRNA expression compared with control flies when the Elav-Gal4 driver was used (see S7A Fig). However, over-expressing Lon by itself in control flies led to retinal degeneration. This prevented us from testing the effect of over-expressing Lon in TDP-43 flies.
On the other hand, two siLon fly lines were obtained, #1 and #2, which reduced Lon expression to ~30% and ~60%, respectively, of that in the control flies (see S7A Fig). Down-regulating Lon expression by itself in control flies did not cause detectable phenotypes. The siLon#1 fly line showed more robust down-regulation efficiency and was thus used in subsequent experiments. We then crossed siLon flies with TDP-43 transgenic flies and examined retinal degeneration and locomotor function in adult flies expressing TDP-43 in photoreceptors or in all neurons respectively. Using the GMR-Gal4/tubulin-Gal80ts driver, we monitored the progression of retinal degeneration during the adult stage following induction of TDP-43 expression by pulses of heat shock. Retinal degeneration was examined using TEM. By day 20 following TDP-43 induction, flies expressing TDP-43 exhibited profound retinal degeneration. The control flies showed normal photoreceptor organization, and heat shock per se did not affect photoreceptor development or maintenance as previously reported [60]. In contrast, retinae in flies expressing TDP-43 showed ommatidial disorganization with a clear reduction in rhabdomere numbers. The average number of rhabdomeres in flies expressing Wt or A315T-mutant TDP-43 was 6 or 5 respectively, as compared with 7 in the control flies (Fig 8A, Fig 8B). In flies expressing Wt or A315T-mutant TDP-43, down-regulating Lon expression exacerbated retinal degeneration, reducing the average rhabdomere number to 5 (Wt; siLon) or 4 (A315T; siLon), respectively (Fig 8A and 8B).
Biochemical fractionation experiments indicate that down-regulating Lon in these flies led to an increase in the mitochondrial TDP-43 level, although there was no significant increase of the TDP-43 levels in the total cell lysates or in the cytosol (see Fig 8E, Fig 8F; S7B and S7C Fig). We further examined solubility of mitochondrial TDP-43 in these flies following sequential extraction using NP-40, SDS and urea. Down-regulating Lon increased the mitochondrial TDP43 protein level, especially the NP-40 soluble fraction in the Wt TDP-43 group and SDS-resistant/Urea-soluble fraction (in the urea lanes) in the A315T-mutant TDP43 group (see S8 Fig). Importantly, Lon knockdown in TDP-43 expressing flies exacerbated mitochondrial damage, with a further reduction in mitochondrial size and an increase in the percentage of damaged mitochondria in the retinae (Fig 8C, Fig 8D), although knock-down Lon by itself in the control flies did not affect rhabdomere or mitochondrial morphology (see S9 Fig). These results demonstrate that mitochondrial TDP-43 accumulation correlates with TDP-43-induced mitochondrial damage and neurodegeneration. Intriguingly, electron-dense aggregate-like structures were detected inside mitochondria in A315T; siLon flies (Fig 8A, marked by a black arrow in the lower panel in the “A315T; siLon” panel; also see S10 Fig, marked by an arrow). These electron-dense aggregate-like structures were not detected in any other groups of flies. Molecular characterization of these electron-dense aggregate-like structures awaits further studies in the future.
We also examined the effects of down-regulating Lon on the locomotor function of the flies expressing TDP-43 under the Elav-Gal4/tubulin-Gal80ts driver. Flies expressing TDP-43 showed progressive locomotor defects following induction of TDP-43 expression, with flies expressing A315T-mutant TDP-43 showing more severe defects. Down-regulating Lon expression in flies expressing Wt or A315T-mutant TDP-43 exacerbated the locomotor defects induced by TDP-43 (Fig 8G). In flies expressing Wt TDP-43, Lon knockdown significantly reduced locomotor function by day 15 onward in males and day 30 onward in females. In flies expressing A315T-mutant TDP-43, Lon knockdown significantly reduced locomotor function by day 10 onward in males and day 25 onward in females. The exacerbating effect of Lon down-regulation appeared more pronounced in males than in females. The onset of locomotor deficits was advanced in both females and males expressing Wt TDP-43. By day 40 post-induction of TDP-43 expression, in flies expressing Wt TDP-43 the locomotor index was >30, whereas down-regulating Lon in Wt TDP-43 flies led to a complete loss of locomotor function in both females and males (Fig 8G). These results show that Lon plays a protective role against TDP-43 induced neurodegeneration in these flies, especially during the early stage of the disease. Together, our data indicate that mitochondrial damage contributes to TDP-43-induced neurodegeneration.
TDP-43 is a multi-functional RNA/DNA binding protein involved in multiple processes of gene regulation, from chromatin remodeling, DNA stability to RNA processing, including microRNA biogenesis, transcriptional and splicing regulation, mRNA trafficking as well as mRNA stability regulation [3,4,61]. Over a decade ago, TDP-43 was identified as a characteristic protein in the inclusion bodies of tissues from patients affected by TDP-43 proteinopathy, including ALS-TDP and FTLD-TDP [1,62]. Since then, a large number of mutations in the TDP-43 gene have been identified in ALS patients, whereas dysregulation of TDP-43 gene expression or its function has been found in patients affected by FTLD and other neurodegenerative disorders [4,63,64].
Several groups have reported mitochondrial abnormalities in different models for TDP-43 proteinopathy, including abnormal mitochondrial clustering [24,26], and a shift in dynamics toward mitochondrial fragmentation [20,22,23,25]. A recent study reported the accumulation of TDP-43 in mitochondria in TDP-43 proteinopathy brain samples [28]. Of these studies, only one reported ultrastructural changes of mitochondria in mice expressing A315T-mutant TDP-43 [20]. However, it was not clear how widespread this damage was. There has not been, to our knowledge, a systematic morphological characterization of mitochondria in patient samples nor in TDP-43 proteinopathy model systems. Our study builds on these previous results by systematically and quantitatively examining TDP-43 induced mitochondrial damage using EM and other methods across different model systems and in patient samples. Our EM analyses clearly show that mitochondria frequently exhibited severe morphological impairment in TDP-43 proteinopathy patient samples and that such mitochondrial morphological changes are consistently detected across cellular and animal models of TDP-43 proteinopathy (Fig 1, Fig 2, Fig 4 and Fig 8). Interestingly, swollen mitochondrial cristae detected in the TDP-43 expressing cells and animals, and in patient samples, are reminiscent of the mitochondrial abnormality in mice expressing SOD1 mutant [65].
Recent studies indicate that cristae morphology determines the assembly and stability of respiratory chain super-complexes, and affects mitochondrial function [66,67]. It is not surprising that mitochondrial cristae are affected in a range of diseases, including neurodegenerative disorders. It has been reported that mitochondrial cristae are disrupted in Alzheimer's disease, showing concentric or parallel stacks [68,69]. A previous study from our group revealed that mitochondria in FTLD-FUS brain tissues showed a marked loss or disruption of cristae, with frequent detection of mitochondria in an “onion-like” deformed shape [70]. Data presented in this study demonstrate that vesicular or swollen mitochondrial cristae are a prominent feature not only in our cellular or animal models, but also in patient samples of TDP-43 proteinopathy (Fig 1, Fig 2 and Fig 4). Our results together with previous studies support the notion that mitochondrial impairment is a common pathogenic contributor to neurodegenerative diseases, and that distinct ultrastructural changes in mitochondria may reflect different mechanisms leading to mitochondrial damage.
Consistent with the morphological changes that we observed, mitochondrial membrane potential and mitochondrial ATP synthesis were reduced upon induction of TDP-43 expression (Fig 2). Interestingly, TDP-43 expression suppressed the activity of mitochondrial complex I, and to a lesser extent, complex IV, without affecting complexes II, III or V (Fig 2). The effect of TDP-43 on ATP synthesis and respiratory complexes has been examined in previous studies, but with discrepant results [23,27–29,71]. Onesto and colleagues observed no change in the total ATP level and reduced mitochondrial membrane potential in fibroblasts from ALS-TDP patients (carrying the A382T mutation), consistent with our results; however, they observed no differences in mitochondrial complex activities. Kawamata and colleagues, on the other hand, reported that there were no mitochondrial bioenergetic defects in fibroblasts or transgenic mice expressing TDP-43 mutants, although mitochondrial calcium handling seemed to be affected [29]. In contrast, Wang and colleagues observed a decrease in ATP synthesis and a decrease in relative levels and activity in complex I from fibroblasts from ALS-TDP patients and HEK293 cells transiently overexpressing wild-type or three ALS-mutants of TDP-43; however, they did not observe changes in the other complexes. Two groups provided evidence for mitochondrial dysfunction, including reduced mitochondrial respiration and ATP synthesis, in NSC-34 cells expressing ALS-mutant TDP-43 [27,71]. Further studies are necessary to resolve the discrepancy in these studies.
Our data presented here show that TDP-43 increases mitochondrial ROS production both in vitro and in vivo (Fig 2; S4 Fig). Mitochondrion is a major site for ROS production, and excessive ROS accumulation can further damage mitochondria [43,72,73]. Although there were no detectable effects of TDP-43 on ROS production in cultured fibroblasts in the previous study [23], data from our cellular model show a clear increase in mitochondrial ROS production induced by TDP-43 (Fig 2). Furthermore, TDP-43 expression in fly motor neurons significantly increased mitochondrial ROS levels in vivo (S4 Fig).
It is interesting to note that the electron-dense TDP-43 positive aggregates detected inside mitochondria in TDP-43 proteinopathy patient brain samples (Fig 1D) are reminiscent of the EM findings in lymphoblasts expressing LonP1 mutations of patients affected by cerebral, ocular, dental, auricular, skeletal (CODAS) syndrome [74]. The mitochondrial abnormalities reported in these CODAS patients are similar to those detected in our TDP-43 proteinopathy patient samples, including swollen intra- or intercristal compartments, swollen or vesicular cristae and intra-mitochondrial aggregate-like structures (see Fig 1) [74]. Intriguingly, similar intra-mitochondrial aggregates were detected in flies expressing A315T-mutant TDP-43 only when Drosophila LonP1 homolog, Lon, was down-regulated (see S9 Fig). Given that LonP1 is an ATP-dependent mitochondrial protease [49,74], and that mitochondrial ATP synthesis is suppressed by TDP-43, it is possible that reduced mitochondrial ATP synthesis might affect proteolytic activity of LonP1, resulting in further TDP-43 accumulation within mitochondria as the disease progresses and eventually leading to irreversible mitochondrial damage and the demise of affected neurons.
Our data from both mammalian cells and transgenic flies show that TDP-43 expression elicits UPRmt, a program that is evolutionarily conserved from nematodes to mammals. UPRmt induces expression of mitochondrial chaperones to assist in proper protein folding and proteases to promote clearance of misfolded proteins [30–32,75]. A variety of mitochondrial stresses induce UPRmt, including accumulation of misfolded proteins, depletion of mitochondrial DNA, ROS overload, perturbation of OXPHOS or mitochondrial translation, and disruption of the balance between mitochondrial- and nuclear-encoded proteins [30–32,76,77]. UPRmt has been reported in Parkinson's disease, Alzheimer’s disease and ALS-SOD1 [33–35]. UPRmt activation detected in our cellular and animal models for TDP-43 proteinopathy could be the result of the combined effects of TDP-43, including mitochondrial accumulation of TDP-43 protein, increased ROS production, decreased membrane potential, impaired respiratory chain function and decreased mitochondrial ATP synthesis. To our knowledge, there were no previous reports of UPRmt in TDP-43 proteinopathy.
Consistent with qPCR results from cellular and fly models, the LonP1 protein level was up-regulated in a fraction of patients affected by TDP-43 proteinopathy (Fig 5). Our data show that LonP1 interacts with TDP-43 and that purified LonP1 degrades TDP-43 (Fig 6 and Fig 7). More importantly, inhibition or down-regulation of Lon led to increased mitochondrial TDP-43 accumulation and exacerbated mitochondrial damage and neurodegeneration phenotype in vivo (Fig 8). It is conceivable that balanced protein synthesis and degradation of TDP-43 is critical for ensuring proper function of TDP-43 in the nucleus, cytosol and mitochondria. Recently, a new mechanism of mitochondria-mediated proteolysis, known as “mitochondria as guardian in cytosol (MAGIC)”, was reported for degrading mis-foled proteins [78]. By MAGIC, cytosolic proteins prone to aggregation can be imported into mitochondria for degradation by mitochondria proteases in yeast and human cells, and PIM1 (encoding yeast Lon protease) is a major player in this process [78]. The complete machinery for MAGIC remains to be defined. Further studies are necessary to determine whether MAGIC is a major mechanism in mammalian proteostasis.
Together, our data led to a working model for the role of mitochondrial degradation of TDP-43 in the pathogenesis of TDP-43 proteinopathy (Fig 9). Under physiological conditions, TDP-43 is predominantly nuclear, although it shuttles between the nucleus and cytoplasm, with a small amount of TDP-43 transported into mitochondria. When TDP-43 mutations occur, or under certain cellular stresses, the mitochondrial TDP-43 level is increased. Excessive mitochondrial TDP-43 accumulation results in mitochondrial impairment, manifesting as mitochondrial membrane potential loss, mitochondrial ROS increase, and reduced mitochondrial ATP synthesis. Such TDP-43-induced mitochondrial damage triggers UPRmt, allowing the cell to initiate a series of responses to regain mitochondrial proteostasis by up-regulating mitochondrial proteases, including LonP1. It is likely at this early stage, before mitochondrial damage becomes irreparable, that mitochondrial stress responses enable the cell to reverse mitochondrial dysfunction. However, as the disease progresses, chronic cellular stresses lead to the excessive accumulation of TDP-43 in mitochondria, inducing irreversible mitochondrial damage. For example, persistent increase in the ROS level and severe reduction in ATP synthesis may result in a vicious cycle of suppression of LonP1 proteolytic activity and further accumulation of mitochondrial TDP-43 in spite of an increased protein level of LonP1, culminating in activation of cell death program(s). Data from our animal model and patient samples, together with our in vitro findings, support the notion that LonP1 may provide a protective mechanism against TDP-43 mediated neurotoxicity. It is noted that the time courses of TDP-43-induced UPRmt gene activation showed differences in male and female flies (Fig 5B). Intriguingly, the exacerbation of locomotor deficits by Lon knockdown appeared to be more pronounced in male flies (Fig 8G). This is consistent with a previous report that expression patterns of Lon protein isoforms were different between male and female flies and that Lon was required for gender-specific responses to oxidative stress [79]. The mechanisms underlying such gender-specific stress responses remain to be elucidated. Further work is necessary to determine whether the gender-specific response(s) play a significant role in humans against neurodegeneration.
Since the discovery of TDP-43-containing inclusion bodies in ALS and FTLD patient samples, intense efforts have been made to identify proteases capable of degrading TDP-43. A number of elegant studies have proposed possible involvement of different proteases in degrading TDP-43, including caspases, calpain and asparaginyl endopeptidase [56,80–86]. None of the previously identified proteases have been shown to protect against TDP-43 induced neurotoxicity in vivo. Our biochemical experiments show that the endogenous TDP-43 and LonP1 interact with each other and that TDP-43 is degraded by the purified recombinant LonP1. Down-regulating LonP1 drosophila homolog, Lon, exacerbates TDP-43 induced mitochondrial damage and neurodegeneration. Together, these data provide previously unknown evidence that the mitochondrial protease LonP1 can protect against TDP-43 induced neurodegeneration in vivo. It will be interesting to investigate in the future whether genetic or epigenetic alterations that affect the expression or function of the human LonP1 gene may influence the onset or progression of TDP-43 proteinopathy. Our study suggests that improving mitochondrial function and reducing mitochondrial damage may provide therapeutic potential for patients affected by TDP-43 proteinopathy.
De-identified postmortem human brain samples from autopsied tissues at the Neuropathology Core of the Cognitive Neurology & Alzheimer's Disease Center at Northwestern University were used following NIH and institutional guidelines. There was no research involving human subjects in this study. All animal studies were performed in accordance with national and institutional guidelines.
HEK293 cells were cultured (37°C, 5% CO2) in DMEM (Gibco), supplemented with 10% FBS (Gibco) and transfected as previously described [70]. HEK293-based T-Rex293 cells (Invitrogen) were transfected with pcDNA4 TO/myc-His plasmids (Invitrogen) expressing either Wt, or A315T-mutant TDP-43 following the manufacturer’s manual. Control cells were transfected with an empty pcDNA4 vector. Individual clones of cells stably expressing TDP-43 were obtained following selection in zeocin (400 μg/mL). To induce TDP-43 expression, tetracycline (0.5μg/mL; unless specified otherwise) was added to the culture medium, and cells were cultured for different periods of time at 37°C until harvesting. Western blotting was used to confirm induction of TDP-43 protein expression.
Transgenic flies expressing the human TDP-43 (Wt or A315T-mutant) were described previously [40,41,87]. GMR-Gal4, OK371-Gal4, Elav-Gal4 and UAS-Lon-RNAi lines were obtained from the Bloomington Drosophila Stock Center (BDSC). Another UAS-Lon-RNAi fly line was obtained from the Vienna Drosophila Resource Center (VDRC). UAS-dLonOE was from the Kyoto Stock Center. The Tubulin-Gal80ts (Tub-Gal80ts) line was kindly provided by Dr. A. Guo (IBP, CAS) [48]. The UAS-mito-roGFP2-Grx1 fly lines were kindly provided by Dr. T. Dick [47].
For flies under the Elav-Gal4/Tub-Gal80ts-driver or GMR-Gal4/Tub-Gal80ts-driver, parental flies were crossed and cultured at 18°C, young flies after eclosion were transferred to 28°C for 4 hr every day to induce TDP-43 expression. Other flies were all cultured at 25°C. All flies were raised in standard fly food, 50% relative humidity, and 12hr-12hr light-dark cycles as described previously [41,70,87].
Antibodies used in this study include polyclonal rabbit-antibodies against TDP-43, ATP5A1, LonP1, HSPA9, ClpP, TOM20 and IMMT (ProteinTech Group Inc), as well as mouse monoclonal antibodies, anti-actin (ProteinTech Group Inc), anti-HSP60 (BD Biosciences) and anti-GAPDH (CWBIO). Rat-anti-dElav antibody is a kind gift from Dr. A. Guo.
Brain samples were evaluated for atrophy and for pathology by hematoxylin-eosin staining and immunostaining using corresponding antibodies, as previously described [40]. The brain tissue samples were fixed in 2.5% glutaraldehyde (GA, Electron Microscopy Sciences) for 2–3 hr at room temperature, after washing with PBS and fixation in 1% OsO4 buffer for 2 hr, the samples were dehydrated with graded ethanol solutions, and then embedded in Epon812 resin (SPI). Ultrathin sections (70 nm) were stained with 2% uranyl acetate for 30 minutes and then lead citrate for 10 minutes before imaging using an electron microscope (TecnaiTM Spirit, FEI).
For fly EM samples, fly heads were collected at day 3, fixed in 4% paraformaldehyde (PFA, Electron Microscopy Sciences) and 2.5% GA overnight at 4°C. For HEK293 cells, cells were rinsed with PBS and then fixed in 2.5%GA overnight at 4°C. TEM sections were prepared following protocols as described previously [88]. Fly heads and cells were then treated in the same manner as the brain tissues described above and sectioned on a Leica EM UC6/FC6 Ultramicrotome. After sections were transferred to copper grids, counter staining was performed with uranyl acetate and lead acetate before EM imaging.
Immuno-EM was carried out following our published protocol [70]. Briefly, samples were fixed in 2% PFA and 0.2% GA overnight. After rinsing with PBS, samples were embedded in 12% gelatin, dehydrated in 2.3M sucrose, subjected to ultrathin sectioning (70 nm) and then mounted on copper grids. After an additional rinse with PBS (with 1% BSA and 0.15% Glycine), samples were blocked in 5% goat serum (Electron Microscopy Sciences, EMS) for 30 minutes. Immunostaining was performed, incubating with primary antibodies for 2 hr followed by immunogold labeled secondary antibodies (EMS) for 1.5 hr. Following rinses with PBS, samples were re-fixed with 2.5% GA for 10 minutes and stained with 4% Uranyl acetate for 5 minutes, and imaged under a FEI TECNAI SPIRIT electron microscope.
Mitochondrial membrane potential was measured in inducible TDP-43 cell lines using the mitochondrial dye JC1 (Invitrogen) following a published protocol [89]. Briefly, 48 hr before assay, inducible stable cells expressing the control vector or TDP-43 were seeded in 6-well plates. Tetracycline (1μg/mL) was added to induce TDP-43 expression for 0, 24 or 36 hr. Cells were detached using Trypsin-EDTA, rinsed in cold PBS and then stained using JC1 (5uM) for 20 minutes at 37°C. Following staining, cells were measured using flow cytometry (BD FACS Calibur) and were analyzed by FlowJo software. Data were obtained from four independent experiments. More than 20,000 cells were measured per group in each experiment.
Image acquisition and analyses of mitochondrial ROS of larval VNC motor neurons were performed according to published protocols with slight modifications [47]. Briefly, OK371-Gal4/UAS-mito-roGFP2-Grx1 flies were crossed with female control or TDP-43 transgenic flies. Third instar wandering larvae were dissected in PBS containing 20mM N-ethyl maleimide (NEM) (Sigma-Aldrich), and incubated for 10 minutes. Larvae were then rinsed with PBS and then fixed with 4% PFA before mounting. Fixed larval ventral nerve chord (VNC) samples were imaged with a Leica SP8 confocal microscope equipped with a 40X oil immersion objective. Probe fluorescence was excited sequentially at 405 nm (reduced roGFP) and 488 nm (oxidized roGFP) (frame by frame) and detected at 500–530 nm. A ratio image was created by dividing a 405-nm image by the corresponding 488-nm image pixel-by-pixel, resulting in the ratio of reduced to oxidized roGFP. Images were processed and quantified using ImageJ.
The total cellular ATP level was measured using a CellTiter-Glo Luminescent Assay (Promega) according to the manufacturer’s instruction. Briefly, 48 hr before assay, the control, Wt or ALS-mutant TDP-43 stale HEK293 cells were seeded in 96-well plates. One μg/mL tetracycline was added to induce TDP-43 expression for 0, 12, 24, or 36 hr. Following removal of the culture media and cell lysis, reaction mixtures were transferred to another opaque 96-well plate to measure luminescence. Luminescent signal values were normalized by the protein amount in each group to determine the total cellular ATP levels.
Mitochondrial isolation was performed according to published protocols with minor modifications [38,70]. Briefly, stable TDP-43-expressing HEK293 cells were suspended in isolation butter [0.22M mannitol, 0.07M sucrose, 20mM HEPES (pH 7.2), 1mM EGTA], homogenized with a Glass/Teflon Potter Elvehjem homogenizer (Bellco Glass Inc) and then fractionated by sequential centrifugation. Pellets (the mitochondrial fraction) were washed twice with wash buffer (0.25M sucrose, 50mM HEPES, 1mM EGTA, pH7.4) and were then resuspended in the same buffer. The protein amount was determined by the BCA protein assay (Pierce).
Fly mitochondrial purification was performed according to a published protocol with minor changes [90]. Sixty fly heads were collected under a microscope and were transferred into a Glass-Teflon Dounce homogenizer containing 500 μL of cold isolation buffer (225 mM Mannitol, 75 mM Sucrose, 10 mM MOPS and 1 mM EDTA, 2.5 mg/mL BSA) and homogenized on ice for 20 strokes. The homogenate was transferred to a 1.5 ml tube for centrifugation at 600 g for 10 min at 4°C. The supernatant was centrifuged at 8,000 g for 10 min at 4°C to enrich for mitochondria. Mitochondrial pellet was washed with 0.5 ml wash buffer (225 mM Mannitol, 75 mM Sucrose, 10 mM KCl, 10 mM Tris-HCl and 5 mM KH2PO4) and were then resuspended in the same buffer.
Mitochondrial ATP synthesis was measured using a published protocol with minor modifications [39]. Briefly, equal amounts (30μg) of purified mitochondria were incubated with reaction substrates (0.15mM P1, P5-di (adenosine) pentaphosphate; 2mM malate; 2mM pyruvate; 0.1mM ADP) with or without oligomycin at 37°C for 5 minutes. Reaction mixtures were stopped by adding boiling stop buffer (100mM Tris-HCl, 4mM EDTA, pH 7.4) and then an equal amount of CellTiter-Glo reagent (Promega) was added to measure ATP using a microplate reader. Mitochondrial ATP synthesis was quantified by subtracting the ATP content in the presence of oligomycin from the ATP content in the absence of oligomycin of the corresponding group.
Stable inducible HEK293 cells expressing either the vector control or TDP-43 (Wt or A315T-mutant) were established as described above. Mitochondria were purified from these cells 24h following induction with tetracycline (1μg/mL) using a published protocol [70]. Briefly, mitochondria were collected from the boundary between 23% and 40% percoll of gradient centrifugation. Mitochondrial respiratory chain complex activities were measured following the published protocols [39,91]. Briefly, 10 μg of mitochondria were applied to a 100μl reaction mixture containing 30 mM KPO4 pH7.2, 5mM MgCl2, 2.5 mg/mL BSA, 0.3 mM KCN, 0.13 mM NADH, 2 μg/mL antimycin A and 97.5 μM ubiquinone-1. The complex I specific activity was determined by the subtraction of the nonspecific activity in the presence of rotenone from the total NADH oxidase activity in the absence of rotenone. Complex II activity was measured in reaction mixture containing 30 mM KPO4 (pH7.2), 5 mM MgCl2, 2.5 mg/mL BSA, 0.3 mM KCN, 50 μM DCPIP, 20mM succinate, 2 μg/mL antimycin A and 65 μM decylubiquinone. The complex II specific activity was determined by subtracting the nonspecific activity in the presence of malonate from the total ubiquinone reductase activity in the absence of malonate. Complex III and IV activities were measured by reduction and oxidation of cytochrome C, respectively, monitoring OD550 respectively, as described previously [91]. Complex V activity was measured by subtracting non-specific activity in the presence of oligomycin following the published protocol [39].
Mitochondrial ROS level was measured as described previously [70]. Briefly, 48 hr before the assay, inducible stable cells expressing the control or TDP-43 were seeded in 6-well plates. Tetracycline (1μg/mL) was added to induce TDP-43 expression for 0, 24, 36hr, respectively. Cells were detached using Trypsin-EDTA, rinsed in cold PBS and then stained with mitoSOX-Red for 20 min at 37°C. After washes, cells were fixed with 4% paraformaldehyde for 20 minutes at room temperature. Cells were measured using flow cytometry (BD FACS ArialI) within 1 hr with analyses using the FlowJo software. Data were obtained from four independent experiments, with more than 20,000 cells were measured per group in each experiment.
Cell death was measured using an Annexin V-FITC Apoptosis Detection Kit I (BD) according to the manufacturer’s instructions. Briefly, 48 hr before the assay, inducible stable cells expressing the control or TDP-43 were seeded in 6-well plates. Tetracycline (1μg/mL) was added to induce TDP-43 expression for 0, 24 or 36 hr. Cells were detached by Trypsin-EDTA, rinsed in cold PBS and then stained with Annexin V-FITC and propidium iodide (PI) followed by immediate analyses (within 1 hr) using flow cytometry (BD FACS Calibur). Data were obtained from four independent experiments, and more than 20,000 cells were measured per group in each experiment.
Cell viability and cytotoxicity were determined using a CytoTox-ONE Homogeneous Membrane Integrity kit following the manufacturer’s instructions (Promega). Briefly, the activity of lactate dehydrogenase (LDH) results in the generation of the fluorescent resorufin product, which was measured using a SPECTRAmax GEMINI XS (Molecular Device; excitation at 560 nm and emission at 590 nm). The cellular LDH activity quantifies the number of viable cells (cell viability); and the activity of LDH released in the culture media quantifies the number of non-viable cells that have lost membrane integrity (cytotoxicity).
A cDNA encoding the human LonP1 protein (amino acid residues 115–959) was cloned into vector pET32M3C [a modified version of the pET32a vector (Novagen, 69015–3)], expressed as an N-terminal thioredoxin and 6XHis-tagged protein and purified from E. coli (Rosetta strain, Novagen) following the published protocol [92]. Purified human LonP1 was analyzed by SDS-PAGE followed by Coomassie Brilliant Blue staining and by immunoblotting using an anti-LonP1 antibody. Following Tet-induction (1μg/mL tetracycline) of the inducible HEK293 cells for 36hr, MycHis-tagged Wt or A315T-mutant TDP-43 protein was purified using Ni-Sepharose (GE Healthcare). Purified TDP-43 protein was incubated in a 30 μL in vitro degradation reaction system [20 mM Tris-HCl (pH8.0), 20 mM NaCl, 10 mM MgCl2, 1 mM DTT, 5 mM ATP] with different concentrations of purified LonP1 protein for 90 min at 37°C. The reaction products were analyzed by Western blotting using the corresponding specific antibodies to detect TDP-43 and LonP1 proteins.
Total RNA was isolated from HEK293 cells or fly heads using TRizol reagent (Invitrogen) as described previously [70]. cDNA synthesis and qPCR were performed as described [30,70,79] using the corresponding primers (see S2 Table). HPRT-1 and Actin5C were used as reference genes for mammalian cells and fly tissues, respectively.
The adult fly locomotor assay was carried out as described previously with minor modifications [41]. Briefly, flies were examined every 5 days with their locomotor index measured as the percentage of flies climbing above a 6-cm line in 15 seconds after they were tapped to the bottom of an empty vial. The experiment was repeated 10 times for each group.
The protein solubility was examined as described previously with minor modifications [40]. Briefly, 100 fly heads were collected for mitochondrial purification. 100 μg of the mitochondrial fractions were resuspended in 200 μL RIPA lysis buffer containing 0.5% NP-40, extracted for 20 minutes on ice and then centrifuged at 12,000 g to collect the supernatant as the NP-40-soluble fraction and the pellet. The NP-40-insoluble pellet was then resuspended and extracted in 200 μL RIPA buffer containing 2% SDS for 20 minutes on ice. Following centrifugation at 12,000 g, the supernatant was collected as the SDS-soluble fraction. The SDS-insoluble pellet was then resuspended and extracted in 100 μL RIPA buffer containing 8 M urea for 20 minutes on ice. Following centrifugation at 12,000 g, the supernatant was collected as the urea-soluble fraction. All fractions were then subjected to Western blotting analysis.
Data were collected in Excel (Microsoft) and analyzed using GraphPad Prism 6 unless specified otherwise. Differences between two groups were analyzed using a Student’s t-test. Multiple group comparisons were performed using a one-way or two-way analysis of variance (ANOVA) followed by post-hoc tests. The bar graphs with error bars represent mean ± standard error of the mean (SEM). Significance is indicated by asterisks: *, P < 0.05; **, P< 0.01; ***, P< 0.001.
|
10.1371/journal.ppat.1003214 | Fates of Retroviral Core Components during Unrestricted and TRIM5-Restricted Infection | TRIM5 proteins can restrict retroviral infection soon after delivery of the viral core into the cytoplasm. However, the molecular mechanisms by which TRIM5α inhibits infection have been elusive, in part due to the difficulty of developing and executing biochemical assays that examine this stage of the retroviral life cycle. Prevailing models suggest that TRIM5α causes premature disassembly of retroviral capsids and/or degradation of capsids by proteasomes, but whether one of these events leads to the other is unclear. Furthermore, how TRIM5α affects the essential components of the viral core, other than capsid, is unknown. To address these questions, we devised a biochemical assay in which the fate of multiple components of retroviral cores during infection can be determined. We utilized cells that can be efficiently infected by VSV-G-pseudotyped retroviruses, and fractionated the cytosolic proteins on linear gradients following synchronized infection. The fates of capsid and integrase proteins, as well as viral genomic RNA and reverse transcription products were then monitored. We found that components of MLV and HIV-1 cores formed a large complex under non-restrictive conditions. In contrast, when MLV infection was restricted by human TRIM5α, the integrase protein and reverse transcription products were lost from infected cells, while capsid and viral RNA were both solubilized. Similarly, when HIV-1 infection was restricted by rhesus TRIM5α or owl monkey TRIMCyp, the integrase protein and reverse transcription products were lost. However, viral RNA was also lost, and high levels of preexisting soluble CA prevented the determination of whether CA was solubilized. Notably, proteasome inhibition blocked all of the aforementioned biochemical consequences of TRIM5α-mediated restriction but had no effect on its antiviral potency. Together, our results show how TRIM5α affects various retroviral core components and indicate that proteasomes are required for TRIM5α-induced core disruption but not for TRIM5α-induced restriction.
| The TRIM5 proteins found in primates are inhibitors of retroviral infection that act soon after delivery of the viral core into the cytoplasm. It has been difficult to elucidate how TRIM5 proteins work, because techniques that can be applied to this step of the viral life cycle are cumbersome. We developed an experimental approach in which we can monitor TRIM5-induced changes in the viral core at early times after infection, when TRIM5 exerts its effects. Specifically, we monitored the fate of the viral capsid protein, the integrase enzyme and the viral genome. We show that TRIM5 induces disassembly of each of these core components, and while some core components simply dissociate, others are degraded. These dissociation and degradation events all appear to be dependent on the activity of the proteasome. However, we also find that each of these TRIM5-induced effects events are not necessary for inhibition. The assay developed herein provides important insight into the mechanism of TRIM5α restriction and can, in principle, be applied to other important processes that occur at this point in the retrovirus life cycle.
| Primates express a range of restriction factors that inhibit retroviral infection, and variation in restriction factors is an important determinant of retroviral tropism [1]–[3]. TRIM5α is one such factor [4], and is a member of the large family of tripartite motif (TRIM) proteins that share a common N-terminus composed of a RING domain that functions as an E3 ubiquitin ligase, one or two B-box domains required for higher-order assembly and a coiled-coil dimerization domain (RBCC) [5]–[8]. TRIM5α also encodes a variable C-terminal B30.2/SPRY domain that recognizes incoming retroviruses [4], [9]–[13] and the consequence of this recognition is that infection is inhibited soon after viral entry [14], before reverse-transcription is completed. The viral capsid (CA) protein is the direct target of TRIM5α proteins [15]–[17], and is recognized by TRIM5α multimers only in the context of assembled viral cores, but not as monomers [15], [16], [18], [19]. The RING domain of TRIM5α exhibits E3 ubiquitin ligase activity, and its removal, or mutation of key cysteine residues that are required for this activity reduces the potency of TRIM5α-mediated restriction [4], [10], [20], [21].
TRIM5α proteins with distinct spectra of antiretroviral activity are present in most, perhaps all, primate species. For example, the prototypic rhesus macaque TRIM5α (rhTRIM5α) is a potent inhibitor of HIV-1 infection but does not efficiently restrict simian immunodeficiency viruses of rhesus macaques (SIVmac) [4]. Human TRIM5α (huTRIM5α) and African green monkey TRIM5α (AGM TRIM5α) also exhibit antiretroviral activity [22]–[24] and although AGM TRIM5α restricts a broad range of retroviruses, huTRIM5α is known to restrict only equine infectious anemia virus (EIAV), and N-tropic MLV (N-MLV) [22]–[25]. Thus, the antiretroviral activity of TRIM5α appears to be quite plastic. Underscoring this point, in two different primate lineages (macaques and owl monkeys), independent retrotransposition events have placed a cyclophilin A (CypA) cDNA into the TRIM5 locus, generating a fusion gene with utterly different antiretroviral specificity, wherein the B30.2/SPRY domain is replaced by CypA [26]–[29].
Although various domains of TRIM5α that are required for restriction have been well defined [7], [8], [30], the precise mechanism by which TRIM5α acts on the incoming viral cores to disrupt infection has been enigmatic. The presence of a restricting TRIM5α protein causes a decrease in the yield of pelletable CA protein following infection and, in the case of huTRIM5α restriction of N-MLV, the loss of particulate CA protein is accompanied by an increase of soluble CA [31], [32]. These experiments prompted a model whereby TRIM5α accelerates the uncoating of retroviral cores. Consistent with this model, a chimeric rhTRIM5α protein, containing the RING domain of TRIM21, lead to the shortening of capsid-nucleocapsid tubes assembled in vitro [33], [34].
A second aspect of TRIM5α-induced restriction is the role played by proteasomes. While inhibition of proteasomes does not rescue infection of restricting cells, it does rescue the formation of an integration-competent reverse-transcription complex, and appears to stabilize capsids in the cytoplasm of restricting cells [14], [25], [35]–[39]. One interpretation of these data is that TRIM5α causes a two-phase block to infection, in which passage of viral DNA to the nucleus is blocked, and then TRIM5 induces the viral core is disassembled by proteasomes. In other studies, however, inhibition of proteasomes was shown to cause a general increase in cytosolic particulate capsid independent of TRIM5 restriction [40], [41]. It has also been proposed that TRIM5α accelerates degradation of CA, by a proteasome-independent pathway [36]. In addition to acting directly on the viral core, TRIM5α has been recently shown to promote innate immune signaling, an activity that is stimulated by and may contribute to restriction of retroviral infection [42]. Overall, it is unclear what the sequence of events is during restriction, and which events are necessary or superfluous for antiviral activity.
As most studies of TRIM5-mediated restriction have focused on CA, the fate of other components of the viral core during restriction is unknown. Given that inhibition of proteasomes in restricting cells can rescue the formation of an integration competent reverse transcription complex [35], [37], one idea is that degradation of core-associated CA leads to the liberation of viral RNA and other core proteins e.g. enzymes. Thus, the physical separation of viral genomes and enzymes could lead to a block in reverse transcription. Alternatively, proteasomes may directly be involved in degradation of other core components. The lack of clarity in current pictures of how TRIM5α works is at least partly due to the difficulty of analyzing retroviral cores in infected cells using biochemical assays. This problem was partly overcome by the development of a “fate-of-capsid” assay, in which viral cores in cytosolic extracts prepared from infected cells are pelleted through a sucrose cushion [31], [32]. This approach has been utilized in a number or studies of retroviral restriction by TRIM proteins [31], [32], [39]–[41], [43]–[46] and capsid stability in infected cells [46]–[48]. Although this assay is very informative and essentially the only widely used assay for the biochemical analysis of post-entry events [31], [32], it does have limitations. First, only a fraction of the input material is actually analyzed - the endocytosed CA, which is thought to constitute the majority of the internalized material, is excluded. In addition, this approach has been applied only for the analysis of CA. Moreover, although restriction by TRIM5α likely occurs at early times after infection (i.e. 1–2 hours) [14], most studies employing this assay analyze events that take place at later stages in infection. Finally, it has been debated whether the CA analyzed during TRIM5α restriction represents viral cores in the infectious pathway [32], [36], [38], [40], as a large fraction of internalized retroviral particles are thought to be nonproductively trapped and degraded in endosomes and lysosomes [49].
In order to overcome these problems, we developed a biochemical assay by which we can monitor the effects of TRIM5α on various components of retroviral cores at early times in infected cells. The approach we took was, essentially, to elaborate existing “fate of capsid” assays. Specifically, we utilized Chinese hamster ovary K1 (CHO-K1)-derived pgsA745 cells (pgsA) which lack surface glycosaminoglycans and, perhaps as a consequence, can be very efficiently infected by VSV-G-pseudotyped viruses. Cytosolic proteins isolated from infected pgsA cells and its derivatives stably expressing various TRIM5α proteins were fractionated on linear sucrose gradients. This approach enabled the fates of CA, integrase (IN), viral genomic RNA and reverse transcription products to be monitored. Using this assay we could show that the aforementioned viral components cosediment in a dense fraction. Moreover, we found that various components of retroviral cores have different fates during TRIM5α-mediated restriction, and can be degraded or disassembled. All of these effects on retroviral cores could be at least partially blocked by proteasome inhibition, but this manipulation did not rescue infectivity. These findings suggest that events that occur prior to core disassembly, rather than core disassembly itself or the action of proteasomes, is crucial for TRIM5α-mediated restriction.
To facilitate analyses of TRIM5α-mediated restriction, we developed a biochemical assay in which we monitored various components of retroviral cores in newly infected cells. We used a CHO-derived cell line (pgsA), because it can be very efficiently infected by VSV-G pseudotyped retroviruses and does not express a TRIM5 protein that restricts MLV or HIV-1 infection [50]. Virions were bound to cells at 4°C, the inoculum was removed, and cells were either harvested immediately (T = 0 hr) or incubated at 37°C to allow infection to proceed for two hours (T = 2 hr). Our previous observations indicate that events critical for TRIM5α restriction take place during this time [14]. Extracts from infected cells were separated on linear sucrose gradients and the presence of various core components in gradient fractions assessed.
Initially, we focused on N-MLV infections and characterized the effects of huTRIM5α restriction on the CA protein. As indicated in Fig. 1A, N-MLV was efficiently restricted in the pgsA-huTRIM5α cell line, as compared to unmodified pgsA cells. When cells were harvested immediately after the virion-binding step (T = 0 hr), CA was present throughout the gradient but enriched in fractions 5 to 8. This distribution is likely a consequence of virions being bound to plasma membrane fragments of varying sizes (Fig. 1B). As expected, the amount of virions bound to pgsA and pgsA-huTRIM5α cells was similar (Fig. 1B). When cells were harvested after a 2 hour incubation at 37°C following virion binding (T = 2 hr) two distinct populations of CA molecules were present in unmodified pgsA cells. One concentration of CA molecules was present at the very top of the gradient (fractions 1 and 2), and presumably represented non-particulate material. A second concentration of CA molecules that were presumably part of a larger complex was evident in fractions 6 to 8, towards the bottom of the gradient (Fig. 1C). Strikingly, the dense peak of CA protein was absent when pgsA-huTRIM5α cells were used as targets (Fig. 1C). Moreover, a clear increase of CA concentration in soluble fractions was observed (Fig. 1C). These results are consistent with prior findings using the established “fate of capsid” assay [31], [32] and imply that TRIM5α may lead to the disassembly of the capsid during the time at which TRIM5 proteins are known to exert their effects.
Although prior data [31], [32], and the findings in Fig. 1 illuminate what happens to CA protein during TRIM5α-induced restriction, the fate of other core components under restrictive conditions was unknown. Therefore, we next asked how the behavior of other components of the N-MLV cores, namely IN, viral RNA and reverse transcription products, are affected by huTRIM5α. To determine the distribution of MLV IN, we inserted a 3×HA epitope tag at its C-terminus in the context of a Gag-Pol expression plasmid. MLV particles generated using this modified Gag-Pol expression plasmid were highly infectious. Notably, when cells were harvested and subjected to analysis immediately after virion binding (T = 0 hr), IN protein was detected primarily in fractions 4 to 8 (Fig. 2A, Fig. S1A), the same fractions in which CA protein was enriched after virion binding (Fig. 1B). As expected, there was no major difference in the amount and migration pattern of IN when pgsA or pgsA-huTRIM5α cells were used. At two hours after infection (T = 2 hr) (Fig. 2B), a dense complex containing IN was detected in unmodified pgsA cells, and virtually all of the IN protein co-migrated in the gradient with the large CA containing complex identified in Fig. 1. Strikingly, the presence of huTRIM5α appeared to induce complete loss of IN at 2 h after infection (Fig. 2B). Note that a protein band detected in fractions 1–3 migrates slightly more slowly than IN and is also detected in uninfected cells (Fig. 2B) as well as when the T = 0 hr blots subjected to a longer exposure (Fig. S1A), indicating that it is a nonspecifically cross-reacting species. In contrast to the CA protein (Fig. 1C), IN was not enriched in soluble fractions under restrictive conditions (Fig. 2B), rather it appeared to be removed from cells.
We next determined the fate of viral genomic RNA during TRIM5α mediated restriction by performing quantitative RT-PCR on the gradient fractions. As was the case with the IN protein, the viral RNA was found primarily in fraction 4 to 8 of the gradient after virion binding to pgsA cells (Fig. 2C). After 2 h of infection in pgsA cells, viral RNA was found mostly in a large complex that comigrated with IN and the large CA containing complex (Fig. 2D). The co-migration of CA, IN and viral RNA suggested that they were part of the same complex, perhaps representing intact, or nearly intact viral cores. Consistent with this notion, the migration of viral RNA through the gradient (peaking at fraction 7) was very different to the migration of a cellular RNA encoding GAPDH, which was localized to fraction 3 (Fig. 2E). Notably, the presence of huTRIM5α in target cells caused a loss of viral RNA from the large complex (Fig. 2D). This huTRIM5α-induced loss of viral RNA from the large complex was accompanied by the appearance of a peak of viral RNA in fraction 3 where cellular GAPDH RNA is present (Fig. 2D, E). In other words, huTRIM5α appeared to liberate viral RNA from a sub-viral complex, causing it to adopt the behavior of a generic cellular mRNA.
Although the bulk of reverse transcription is likely not completed at T = 2 hr [14], we could easily detect reverse-transcription products in infected cells at this time point. These viral DNA species co-migrated with other components of the viral core under non-restrictive conditions (Fig. 2F). However, as expected, reverse transcription was blocked in cells expressing huTRIM5α and little viral DNA was detected anywhere on the gradient (Fig. 2F).
Given that all components analyzed that are predicted to be components of the viral core, co-fractionated with each other, it is likely that the sub-viral complexes detected herein represent functional complexes in which reverse transcription is taking place. The fact that TRIM5α clearly affected the fates of each of the viral components that were present in the dense fraction indicates that they were present in the cytoplasm, as they should not be affected by TRIM5α if they were in any other cellular location (e.g. endosomes). Moreover, these results suggest that TRIM5α-mediated restriction involves both disassembly and degradation, with the differing ultimate fates of various core components.
We next performed three control experiments to verify that the effects that we observed in Figs. 1 and 2 are truly relevant to restriction. Because it is thought that a significant fraction of internalized virions remain trapped in endosomes, we first asked whether the different gradient-migration patterns of core components under restrictive and nonrestrictive conditions was dependent on viral entry into the cytoplasm. To that end, cells were infected with VSV-G pseudotyped N-MLV for two hours in the presence of ammonium chloride (NH4Cl), which prevents endosome acidification and blocks VSV-G mediated entry. After 2 h of infection in the presence of NH4Cl, CA was distributed throughout the gradient (Fig. 3A), while IN (Fig. 3B) and viral RNA (Fig. 3C) were found primarily in fraction 4–8. This pattern was similar to that observed when cells were harvested immediately after virion binding, and quite different to that observed when infection was allowed to proceed for 2 h in the absence of NH4Cl (Fig. 1, 2). Importantly, the migration profile of the core components in the presence of NH4Cl was not affected by huTRIM5α. As an additional control experiment, we infected the non-restricting pgsA cells with either VSV-G-pseudotyped N- MLV (Env (+)) or N-MLV VLPs without VSV-G (Env (−)) for two hours. We could not detect CA (Fig. S2A) or IN (Fig. S2B) in the gradients prepared from cells incubated with Env (−) VLPs, This suggested that the Env(−) particles either did not efficiently bind to the target cells, or were degraded in endosomes without entering the cytoplasm. As expected, Env (−) particles were completely non-infectious (Fig. S2C) and, importantly, the initial virus inoculum contained equal amounts of Env (+) and Env (−) particles (Fig. S2D). Thus, the changes in the behavior of core components induced by huTRIM5α was dependent on VSV-G-mediated binding and entry.
Next, to determine whether the effects of huTRIM5α on N-MLV cores is a result of restriction activity, we performed similar experiments to those described above with B-MLV, which is insensitive to huTRIM5α restriction [22]–[25]. When viral cores were harvested after synchronization (T = 0 hr), B-MLV CA, IN (Fig. 4A, Fig. S1B) and viral RNA (Fig. 4B) co-fractionated, primarily in fractions 5 to 8, although CA was also detectable in other fractions (as was observed with N-MLV (Fig. 1–3)). All core components migrated in a similar pattern irrespective of the presence of huTRIM5α (Fig. 4A, 4B). At two hours post-infection, CA, IN (Fig. 4C, Fig. S1C) and viral RNA (Fig. 4D) were all observed as components of large complexes regardless of the presence of huTRIM5α. Moreover, unlike N-MLV, huTRIM5α did not lead to any observable increase of B-MLV CA (Fig. 4C) or viral RNA (Fig. 4D) in soluble fractions (1 to 3). As expected, the level of reverse transcription at this time point was not affected by huTRIM5α and viral cDNA co-fractionated with other core components (Fig. 4E). Collectively these results validate our assay and support the notion that changes in the behavior of viral core components are induced by a restricting TRIM5α protein.
The role of proteasomes during TRIM5α restriction has been a matter of debate, with several studies showing that inhibition of proteasomes does not restore infectivity that is restricted by TRIM5 proteins [14], [31], [32], [35], [37]. Moreover, in studies that employed the fate of capsid assay, huTRIM5α was reported to retain a significant ability to induce solubilization of MLV CA in the presence of proteasome inhibitor MG115 [41]. Some experiments have shown that proteasome inhibition causes a general increase in particulate HIV-1 and MLV capsids in both restricting and non-restricting cells [40], [41]. However, it has been demonstrated that proteasome inhibition can restore reverse transcription, and the formation of a functional preintegration complex in the presence of TRIM5α [35], [37]. As such, it is somewhat unclear whether proteasome inhibitor-restored reverse transcription complexes lack other core components or whether they are indistinguishable from unrestricted cores. Indeed, previous studies have indicated that viral DNA that is synthesized under TRIM5α-restricted, but proteasome inhibitor-restored conditions cannot enter the nucleus and become integrated [35], [37].
Therefore, we asked whether inhibition of proteasomes in cells expressing huTRIM5α could restore the presence of large N-MLV sub-viral complexes containing CA, IN and viral RNA. To that end, we infected pgsA-huTRIM5α cells in the presence of MG132, a proteasome inhibitor. As observed above in Fig. 1 and Fig. 2, when huTRIM5α expressing cells were infected in the absence of MG132, large complexes containing CA (Fig. 5A) and viral RNA (Fig. 5B) were lost and there was a concomitant increase in the levels of CA and viral RNA in soluble fractions. Strikingly, MG132 treatment restored large subviral complexes containing CA (Fig. 5C) and viral RNA (Fig. 5D) as well as the formation of reverse transcription products (Fig. 5E). Importantly, in contrast to a previous study [41], we did not observe a non-specific increase in dense N-MLV capsid in cells in the presence of MG132. This may be either due to the fact that a different proteasome inhibitor was used in the study by Diaz-Griffero et al. [41] or that the indirect effects of proteasome inhibition our assays is minimized, because we analyzed an earlier time point in infection. Nonetheless, as previously reported, proteasome inhibition did not restore N-MLV infectivity in huTRIM5α cells (Fig. 5F). These results suggest that although proteasomes play an important role in mediating the observed biochemical changes on viral cores induced by TRIM5α in our assays, they are not central to TRIM5α restriction.
Recent findings have suggested the possibility that the uncoating (loss of CA protein) of HIV-1 viral cores early after infection is stimulated by reverse transcription [46], [51]. In addition, reverse transcription was suggested to be required for rhTRIM5α-mediated disassembly of the HIV-1 core using the fate of capsid assay [46]. It was possible therefore that the initiation of reverse transcription might facilitate, or even be required for, the apparent disassembly and destruction of core components that we observed. Therefore, we repeated the above experiments in the presence of the reverse transcriptase inhibitor, AZT. Importantly the doses of AZT used were sufficient to block infection (Fig. S3A), and the synthesis of reverse transcripts (Fig. S3B) under non-restricting conditions. Notably, treatment of pgsA cells with AZT during the 2 h infection assay did not affect the distribution of CA and IN in sucrose gradients: CA was present in both sets of fractions containing soluble proteins and large complexes while IN localized primarily to fractions containing large complexes (Fig. 6A). The presence of huTRIM5α in target cells led to complete disappearance of both CA and IN from large complexes, with an accompanying increase of CA in soluble fractions under these conditions (Fig. 6A). Similar to CA, even in the presence of AZT, huTRIM5α lead to the release of viral genomic RNA from the large complex (Fig. 6B). Notably, the peak of viral RNA in the presence of huTRIM5α was lower than that in its absence (Fig. 6B), suggesting that the viral RNA that is released from the core may be targeted for degradation (discussed in detail below). Inhibition of proteasomes under restricting conditions, when reverse transcription was blocked substantially restored the presence of CA, IN (Fig. 6A) and viral RNA (Fig. 6C) in large complexes. These results confirm our previous findings and suggest that huTRIM5α action involves both disassembly and proteasome-mediated degradation of viral core components, and that these events occur independently of reverse transcription.
We next sought to extend these observations and asked whether HIV-1 cores are similarly affected by TRIM5α restriction. To this end, we generated pgsA cells that stably express rhTRIM5α, which potently restricts HIV-1 infection (Fig. S4A). When cells expressing hu- or rhTRIM5α were harvested immediately after synchronization, CA was detected primarily in the top two fractions and in fractions 5 to 7 (Fig. 7A), whereas IN (Fig. 7A) and viral RNA (Fig. 7B) were more distinctly localized in fractions 5 to 7. As expected, there was no difference in the behavior and amounts of HIV-1 core components harvested from huTRIM5α and rhTRIM5α cells at T = 0 h (Fig. 7A, 7B). At T = 2 h post-infection, the CA protein in huTRIM5α cells was present as two distinct species with distinct migration properties in the gradient. A predominant species was present at the top of the gradient, likely corresponding to soluble proteins while a second species was present in denser sucrose fractions, likely representing viral reverse transcription complexes (Fig. 7C). The overall profile of the behavior of HIV-1 CA molecules in the sucrose gradient was quite similar to that of MLV, but the relative abundance of the soluble CA protein was greater in the case of HIV-1, suggesting the possibility that HIV-1 cores are either uncoated more rapidly following infection, or are inherently less stable in sucrose gradients. As was the case with MLV, the larger CA containing complex was lost in the presence of a restrictive TRIM5α protein (in this case rhTRIM5α, Fig. 7C). However, a corresponding increase of CA in soluble fractions was not observed, perhaps because soluble CA was already quite abundant under non-restricting conditions (Fig. 7B). Similarly, as was the case with MLV, rhTRIM5α restriction also led to the disappearance of HIV-1 IN from dense fractions, without any concurrent increase in soluble protein containing fractions (Fig. 7C).
Although rhTRIM5α restriction appeared to induce a decrease in the levels of viral RNA in dense fractions, this was not accompanied by an increase in the absolute levels of soluble RNA, although the relative amounts of soluble RNA vs. large-complex-associated RNA were increased in the presence of rhTRIM5α (Fig. 7D). This was unlike our observations with N-MLV, and makes the analysis of HIV-1 RNA profiles difficult to interpret (see discussion). Of note, under non-restricting conditions, the peak of viral RNA in dense fractions did not perfectly overlap with that of CA and IN (Fig. 7D). This could possibly be a consequence of instability of HIV-1 cores in cells or on sucrose gradients, as was suggested by the relative abundance of soluble CA versus complex-associated CA (Fig. 7C). In contrast, the products of reverse transcription co-fractionated nearly precisely with CA and IN under non-restricting condition and, as expected, were substantially reduced in rhTRIM5α cells (Fig. 7E) suggesting that the large complexes containing CA, IN and viral DNA are, or are derived from, functional HIV-1 reverse transcription complexes.
To overcome any potential impact of reverse transcription on uncoating [46], [51], we repeated the above experiments in the presence of the reverse transcriptase inhibitor nevirapine. Importantly the doses of nevirapine used were sufficient to block infection (Fig. S4B), and reverse transcription (Fig. S4C) under non-restricting conditions. As was found with MLV, inhibition of reverse transcription in restrictive or non-restrictive cells did not affect the behavior of viral CA and IN proteins, neither of which were present in large complexes in the presence of rhTRIM5α (Fig. 8A). These results contrast with recent findings which suggest that inhibition of reverse transcription blocks the rhTRIM5α-mediated disassembly of the HIV-1 cores [46]. Notably however, nevirapine treatment of huTRIM5α cells substantially increased the level of viral RNA in dense fractions, and caused it to co-fractionate with CA and IN (Fig. 8B). This finding suggests that the poor co-fractionation of CA, IN and viral RNA observed in Fig. 7D is a consequence of reverse transcription, or RNaseH activity, rather than misbehavior of HIV-1 cores on sucrose gradients. Notably, under restricting conditions, in the presence of nevirapine, viral RNA was lost from the large complexes, with no accompanying increase in soluble fractions (Fig. 8B). This contrasts with our findings with MLV, where restriction led to an increase in the levels of soluble viral RNA.
Finally, we determined whether proteasome inhibition restored the presence of HIV-1 cores under restricting conditions. When rhTRIM5α-expressing, HIV-1 infected cells were treated with MG132 and nevirapine, CA, IN (Fig. 8A), viral RNA (Fig. 8C) and reverse transcription products (Fig. 8D) were all significantly restored in dense fractions. However, as was the case with restricted MLV infection, proteasome inhibition did not restore virus infectivity (Fig. 8E). It is important to note that, unlike a previous study [40], we did not observe a non-specific increase in particulate capsid in cells in the presence of MG132. This may be either due to the fact that our assays are performed at much earlier time points post-infection, which may minimize indirect effects of proteasome inhibition, or that a different proteasome inhibitor was used in the study by Diaz-Griffero et al. [40]. Nonetheless, these results suggest that rhTRIM5α modifies the HIV-1 cores in a way that likely leads to the degradation of both IN and viral RNA. Although this process is sensitive to proteasome inhibition, but is not required for the antiviral activity of TRIM5α to be manifested.
We then sought to confirm there observations by asking whether similar changes on HIV-1 cores can be induced by a different restrictive TRIM protein, namely owl monkey TRIMCyp (omkTRIMCyp) [27], which was shown previously to reduce the amount of pelletable capsid in a fate of capsid assay [32], [40], [43]. This experimental system is better internally controlled, as restriction by omkTRIMCyp protein can be overcome by treatment of cells by cyclosporin A (CsA), which prevents its binding to viral CA protein [27]. Since restricted and non-restricted HIV-1 core components were more reliably compared in the presence of reverse transcriptase inhibitor nevirapine (Fig. 8), we performed similar experiments in pgsA-omkTRIMCyp cells in its presence. As expected, although the treatment of pgsA-omkTRIMCyp cells with CsA restored infectivity (Fig. 9A) and reverse transcription (Fig. 9B), the doses of nevirapine used in these experiments were sufficient to block both of these processes (Fig. 9A, B). In omkTRIMCyp-expressing cells treated with nevirapine alone, viral CA (Fig. 9C), IN (Fig. 9D) and viral RNA (Fig. 9E) were absent in dense fractions, without any notable increase in soluble fractions, similar to our observations with rhTRIM5α. As expected, large sub-viral complexes containing CA (Fig. 9C), IN (Fig. 9D) and viral RNA (Fig. 9E) were restored in the presence of CsA. Importantly, when pgsA-omkTRIMCyp cells were treated with MG132 and nevirapine, CA (Fig. 9C), IN (Fig. 9D) and viral genomic RNA (Fig. 9E) were all restored in dense fractions to almost the same level as observed under CsA treatment. However, as it was the case with restricted MLV and HIV-1 infections, proteasome inhibition did not restore virus infectivity (Fig. 9A). These results together show that omkTRIMCyp disrupts HIV-1 cores in a similar way to that of rhTRIM5α and leads to degradation of at least some core components. Likewise, although this process is sensitive to proteasome inhibition, proteasomes are not required for the antiviral activity of omkTRIMCyp.
We formulated an experimental approach in which the fates of multiple viral core components can be tracked in infected cells, with the aim of understanding how TRIM5α restricts retroviral infection. The approach is similar in principle to the “fate of capsid” assay [31], [32], in which the putative separation of viral cores from infected cell lysates on sucrose gradients enables the analysis of their composition. However, our assay is more elaborate, and perhaps more effective, in several aspects. First, we monitored TRIM5α- and TRIMCyp-induced changes not only for CA, but also for IN, viral RNA and reverse transcription products in the same fractionation experiment. Second, in our assay, all of the input cellular material is analyzed, without the need for exclusion of putatively endocytosed virions. Although it is generally held that the majority of retroviral particles become trapped in endosomes of target cells, complicating analysis of early events in infection, this did not seem to be a major problem in our experiments. Indeed, the nearly complete disappearance of IN at T = 2 h specifically from restricting cells, argues that there is very little virus associated with the cells that had not reached the cytoplasm by this time point. Although the reasons for this are not clear, possibilities include highly efficient VSV-G-mediated entry in pgsA cells, particular instability of endocytosed virions in pgsA cells, or the fairly low MOIs used in these experiments. Third, infections are fully synchronized and the unbound input virus is removed before infection, which could limit the number of virions that are nonspecifically endocytosed. Fourth, analysis is carried out at an early time (2 h) after infection when events relevant to TRIM5 restriction occur [14]. Fifth, we have incorporated quantitative aspects in our experimental system: Q-PCR analysis of viral RNA has proven to be an accurate and quantitative indicator of the fate of the viral core undergoing TRIM5α restriction. Overall our findings suggest that of all of the above components are present in a large complex comprising all or part of the virion core that is a functional intermediate in the infection pathway.
Our findings provide insight into events that take place during TRIM5α restriction (Fig. 10). In parallel with previous findings [31], [32], we observed that N-MLV CA was redistributed from large complexes to soluble fractions in cells expressing huTRIM5α (Fig. 1C). We expanded these observations and show that viral RNA was released from large complexes as a result of huTRIM5α restriction (Fig. 2D). In contrast, MLV IN was not retained in a soluble form following its loss from dense fractions, and appeared to be degraded (Fig. 2B, Fig. 10).
HIV-1 differed from MLV in that neither HIV-1 CA nor viral RNA was apparently increased in soluble fractions concurrent with their loss from large complexes (Fig. 7B, 7D, 8C, 8E). However, the comparative pre-existing abundance of CA in soluble fractions may have masked any redistribution of CA protein to those fractions. Possible reasons for the discrepant fate of MLV and HIV-1 RNA under restricting conditions are discussed below. In the case of HIV-1 IN, the protein was lost from cells under restricting conditions in much the same way as was observed for MLV. Collectively these results indicate that TRIM5α causes both disassembly and degradation of viral components with similarities and differences in the fates of individual core components across retroviral genera (Fig. 10).
Recent findings have suggested the possibility that the uncoating of retroviral cores early after infection is stimulated by reverse transcription [46], [51] and that rhTRIM5α-mediated disassembly of HIV-1 cores requires reverse transcription activity [46]. Although in some experiments reverse transcriptase inhibitors modestly increased the amount of capsid detected by western blotting, we did not observe any effect of RT inhibitors on TRIM5-mediated disassembly/degradation of cores in this study. The reasons underlying the discrepancy between our results and the study by Yang et al. [46] are not clear. However, one would predict that reverse transcription is not required for restriction by TRIM5, based on the fact that TRIM5 acts rapidly after entry [14], before majority of reverse transcription is completed.
The precise role of proteasomes in TRIM5-mediated restriction has been difficult to unambiguously determine. As previously demonstrated [35], [37], inhibition of proteasomes in restricting cells restored MLV and HIV-1 reverse transcription (Fig. 5E, 8D). Importantly, we found that proteasome inhibition restored a core complex that is biochemically indistinguishable from unrestricted viral cores, and contained CA, IN and viral RNA (Fig. 5, 6, 8). As such, it is unlikely that TRIM5α mediates the complete disassembly of cores without the aid of proteasomes. Nevertheless, it is clear that proteasomes are not required for restriction by TRIM5α, as MG132 treatment of restricting cells does not restore virus infectivity ([14], [31], [32], [35], [37] and Fig. 5F, 8E, 9A). Recent findings suggest that TRIM21/TRIM5α chimeras have the propensity to form hexameric lattices on the HIV-1 core, and it is possible that this phenomenon, in itself, constitutes the underlying mechanistic basis for restriction [52]. The assembly of such a lattice on the core may block the targeting of viral reverse-transcription or pre-integration complexes to the nucleus, because circular viral DNA forms are not generated during restricted HIV-1 infection under conditions of proteasome inhibition [35], [37]. However, because HIV-1 and MLV apparently have different underlying mechanisms of entering the nucleus, it is possible that the other mechanisms that sequester viral DNA (e.g. failure to uncoat) may underlie the inability of HIV-1 or MLV to productively infect restrictive cells under conditions of proteasome inhibition.
It is intriguing that some N-MLV and HIV-1 core components, notably viral RNA (and perhaps CA), have somewhat different fates under restrictive conditions (Fig. 10). A possible explanation for this difference is that N-MLV core components are intrinsically more stable and as such, are degraded at a slower rate after TRIM5α-induced disassembly. Alternatively, rhTRIM5α and omkTRIMCyp may either specifically recruit a cofactor that more efficiently degrades the core components or simply disassemble HIV-1 cores at a faster rate. The loss of both N-MLV and HIV-1 IN in dense fractions without any apparent increase in soluble fractions may reflect the previously reported intrinsic instability of these proteins [53].
We did not detect obvious ubiquitinylation of any core proteins undergoing restriction in our assays. It is conceivable that ubiquitin-independent degradation or disassembly by proteasomes may be important for the observed effects on the cores [54]–[59]. Alternatively, if TRIM5 is responsible for ubiquitin modification of only a small fraction of core-associated proteins (e.g. CA), we would not be able to detect this modification yet it could be responsible for core disassembly.
The most striking difference between HIV-1 and MLV restriction is the fate of the viral RNA following its release from the core. It appears that MLV RNA is largely preserved, in a soluble form, whereas HIV-1 RNA is lost. We speculate that the mechanism by which HIV-1 viral RNA is lost during restriction is related to its nucleotide composition. It has long been known that the high AU content destabilizes the HIV-1 genome [60]–[66]. It is therefore conceivable that once the HIV-1 genome is exposed in the cytosol as a result of restriction, AU-rich elements may lead to the degradation of the genome, in the same way as has been observed with several RNAs coding for oncoproteins and growth factors [67]. Alternatively, proteasomes themselves, which have been suggested to comprise RNase activity, or other putative TRIM5α associated RNase activities may lead to selective degradation of AU-rich viral RNA molecules [68]. Nevertheless, it is unlikely that RNA degradation is critical for TRIM5 restriction as TRIMCypA chimeras containing the RBCC domain from other TRIM proteins, certain RING domain mutants of TRIM5α and squirrel monkey TRIM5α can restrict HIV-1 and SIVmac infection, respectively, after reverse transcription is completed [39], [69], [70].
TRIM5α mediated restriction serves as a useful model on which to base investigations of post-entry events. As such, the assay developed here could also be utilized to study restriction-independent events in newly infected cells. For example, it has been suggested that retroviral cores are optimally stable, and changes in CA stability in vitro can lead to defects in reverse transcription [71]. The assay developed here could identify the effects of such changes on multiple viral core components in infected cells. However, a caveat of our assay is that the precise nature of the ‘large complexes’ to which we refer is not known. For instance, it is not known whether the large complexes containing CA and the cofractionating core components actually represent intact conical viral cores. Previous investigations of cores isolated from extracellular virions and infected cells revealed notable differences in the density of N-MLV, B-MLV and HIV-1 ‘cores’ [31], [71]–[74]. We did not observe such differences in our assays, as separation of cytosolic extracts in our experimental system is based on size, rather than density. Therefore, it is plausible that MLV and HIV-1 cores of different densities migrate almost identically on the sucrose gradients as they have similar sizes. Notably, even under non-restrictive conditions, a significant fraction of CA is present in soluble fractions. A similar phenomenon has been previously observed by others during isolation and biochemical characterization of HIV-1, and to a lesser extent MLV, reverse transcription complexes in infected cells [75], [76] This could be a consequence of the disassembly of some fraction of CA immediately upon infection or of the fact that only a proportion of the virion CA protein is actually assembled into cores in mature virions [77]. It is unlikely that the soluble CA represents complete disintegration of a fraction of viral cores in dense sucrose gradients [71], [73], as neither viral RNA nor IN is solubilized under non-restrictive conditions.
Overall, we devised a novel experimental approach in which events that take place during TRIM5α restriction can be analyzed, and that can be applied generically to the study of early events in retrovirus replication cycle. Our results indicate that viral core components have distinct fates during TRIM5α restriction and are either disassembled or degraded. Importantly, in line with the two-step mechanism previously proposed [35], [37], [39], [69], although the TRIM5α-induced biochemical changes on the viral cores in our assays are sensitive to proteasome inhibition, proteasomal degradation is clearly not required for restriction. Future studies will address by which mechanism TRIM5α can restrict retrovirus infection as well as the mechanistic details of how different core components are affected by restriction.
CHO K1-derived pgsA-745 cells (CRL-2242, ATCC) and all of its derivatives were maintained in Ham's F12 media (Life technologies, 11765-054) supplemented with 10% fetal bovine serum and 1 mM L-glutamine. HEK 293T cells were obtained from ATCC (CRL-11268) and maintained in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum. VSV-G pseudotyped viruses were produced by transfection of 293T cells with plasmids expressing HIV-1 or MLV Gag-Pol, a packagable vector genome (see below) carrying GFP [15], [78] and VSV-G at a ratio of 5∶5∶1, respectively, using polyethyleneimine (PolySciences, Warrington, Pennsylvania, United States) as described previously [79].
Sequences encoding huTRIM5α, rhTRIM5α and omkTRIMCyp were inserted into LNCX retroviral vector plasmid (Clontech), which were subsequently used to generate cloned pgsA-745 cell lines stably expressing huTRIM5α and rhTRIM5α. MLV and HIV-1 vector genome plasmids, CNCG and CCGW, respectively, encode a GFP reporter under the control of CMV promoter [15], [78]. NL4-3 derived HIV-1 Gag-Pol sequence were inserted into the pCRV-1 plasmid [80] and carry a hemagglutinin (HA) tag at the C-terminus of integrase (pNL-GP IN-HA). Sequences encoding B-MLV and N-MLV Gag-Pol inserts carrying a single copy or three copies of HA-tag at the C-terminus of integrase were inserted into pCAGGS plasmid [81]. Further details of plasmids are available upon request.
PgsA745 cells, or derivatives thereof, (4×106) were plated on 10-cm cell culture dishes one-day before infection. For each treatment and time point, two such 10-cm dishes were used. In parallel, 2.5×104 PgsA745 cells were plated in 24-well plates to determine virus infectivity in each experiment. The corresponding MOI on 10-cm dishes was ∼0.025 for MLV infections and ∼0.01 for HIV-1 infections. Cell culture supernatants containing VSV-G pseudotyped viruses were filtered and treated with RNase free DNaseI (Roche) at a concentration of 1 unit/ml for 1 hour at 37°C in the presence of 6 mM MgCl2. Cells were washed with ice-cold phosphate-buffered saline (PBS) and 6–7 ml of chilled virus (adjusted to contain 20 mM HEPES) was added to the cells. After allowing virus binding to cells at 4°C for 30 minutes, the inoculum was removed and cells were washed three times with PBS. Parallel infections were carried to determine the infectious titer in a given experiment. Cells were either harvested immediately (T = 0 hr) or incubated at 37°C for 2 hours (T = 2 hr) in complete cell culture media. In some experiments, cyclosporine A, proteasome and reverse transcriptase inhibitors were included during virion binding and during incubation at 37°C. Cells were collected in 1X PBS-EDTA, pelleted and resuspended in 1 ml of hypotonic buffer (10 mM Tris-Cl pH 8.0, 10 mM KCl, 1 mM EDTA supplemented with complete protease inhibitors (Roche) and SuperaseIN (Life technologies)). After incubation on ice for 15 minutes, cell suspension was dounce homogenized by 50 strokes, using pestle B. The disruption of cells and the integrity of nuclei were monitored by Trypan blue staining of cells and nuclei (Fig. S5). Nuclear material was pelleted by centrifugation at 1000×g for 5 minutes and post-nuclear supernatant was layered on top of a 10–50% (w/v) linear sucrose gradient prepared in 1X STE buffer (100 mM NaCl, 10 mM Tris-Cl (pH 8.0), 1 mM EDTA). Samples were ultracentrifuged on a SW50.1 rotor at 30000 rpm for 1 hour. Ten 500 µl fractions from top of the gradient were collected, and proteins, RNA and DNA in each fraction was analyzed as described below.
Proteins in each sucrose fraction were precipitated by trichloroacetic acid as described previously [82]. Protein pellets were resuspended in 50 µl of 1X protein sample buffer and analyzed by western blotting. The primary antibodies used were: mouse monoclonal anti-HA (HA.11 Covance), mouse monoclonal anti-HIV-1 p24CA (183-H12-5C NIH), mouse monoclonal anti-HIV-1 IN (a gift from Michael Malim) and goat polyclonal anti-MLV p30 (a gift from Stephen Goff).
For analysis of DNA and RNA, 50 µl of each fraction of the sucrose gradient was digested with proteinase K, phenol∶chloroform extracted and precipitated using sodium acetate/ethanol as described previously [82]. For analysis of RNA, samples were further treated with DNase I, extracted again and reverse-transcribed using the ImProm-II reverse transcription kit (Promega). The resulting cDNA and DNA samples were used as template for quantitative real-time PCR (qPCR) using FastStart Universal SYBR Green Master Mix (Roche) and ABI 7500 Fast PCR system. PCR primers were designed within the GFP region of the vector genome. The primer pairs used in this study are as follows: GFP: Forward: 5′ AAGTTCATCTGCACCACCGGCAA Reverse: 5′ TGCACGCCGTAGGTCAGG; GAPDH: Forward: 5′ AGG TGA AGG TCG GAG TCA ACG, Reverse: 5′ GGT CAT TGA TGG CAA CAA TAT CCA CTT TAC.
|
10.1371/journal.pntd.0007188 | A combined experimental-computational approach for spatial protection efficacy assessment of controlled release devices against mosquitoes (Anopheles) | This work describes the use of entomological studies combined with in silico models (computer simulations derived from numerical models) to assess the efficacy of a novel device for controlled release of spatial repellents. Controlled Release Devices (CRDs) were tested with different concentrations of metofluthrin and tested against An. quadrimaculatus mosquitoes using arm-in cage, semi-field, and outdoor studies. Arm-in-cage trials showed an approximate mean values for mosquito knockdown of 40% and mosquito bite reduction of 80% for the optimal metofluthrin formulation for a 15-minute trial. Semi-field outdoor studies showed a mean mortality of a 50% for 24 hour trial and 75% for a 48 hour trial for optimal concentrations. Outdoors studies showed an approximate mean mortality rate of 50% for a 24 hour trial for optimal concentrations. Numerical simulations based on Computational Fluid Dynamics (CFD) were performed in order to obtain spatial concentration profiles for 24 hour and 48 hour periods. Experimental results were correlated with simulation results in order to obtain a functional model that linked mosquito mortality with the estimated spatial concentration for a given period of time. Such correlation provides a powerful insight in predicting the effectiveness of the CRDs as a vector-control tool. While CRDs represent an alternative to current spatial repellent delivery methods, such as coils, candles, electric repellents, and passive emanators based on impregnated strips, the presented method can be applied to any spatial vector control treatment by correlating entomological endpoints, i.e. mortality, with in-silico simulations to predict overall efficacy. The presented work therefore presents a new methodology for improving design, development and deployment of vector-control tools to reduce transmission of vector-borne diseases, including malaria and dengue.
| Spatial Repellents (SRs) represent another tool to fight vector-borne diseases, such as malaria and dengue. Newly developed active ingredients were designed to repel or kill vectors in space, creating a shield effect, unlike topical repellents, such as DEET, that rely on vectors to be near or in physical contact with protected target. Metofluthrin and transfluthrin are examples of active ingredients (AIs) designed as SRs against mosquitoes. The efficacy of SRs heavily depends on the delivery method. Currently, there is a lack of fundamental understanding of effectiveness of SR delivery methods. Current delivery modalities of SR do not rely on quantitative models to estimate targeted efficacy, making the end-user overshoot or undershoot the dosage required for protection. Optimizing the dosage over time is critical to obtain protection for a given space in a giving period of time, as well to prevent AI resistance in the long run. The key is therefore to deliver just enough dosage needed to repel or kill the vectors. The presented work provides a novel approach to predict performance of SRs based on an experimental-computational methodology to quantify effectiveness of controlled release devices as a function of AI physical properties (e.g. volatility), device design parameters combined with physical variables, and environmental conditions (e.g. temperature, wind velocity). This work therefore provides the groundwork for estimating quantitative effectiveness of SR delivery methods against mosquitoes.
| Vector-borne diseases represent a global public health threat. More than 1 million people die annually of vector-borne diseases. Malaria alone is responsible for 400,000 deaths a year, and most cases are children under five years of age [1–3]. Current vector control techniques can be divided into contact repellent and spatial repellent products [4–7]. Current strategies against mosquitoes include products applied to the skin, insecticide treated nets (ITNs) and indoor residual spraying (IRS). Some topical repellents suffer from undesired odor and texture, resulting in poor user acceptability. Topical repellents presence lowers in time due to skin washing, sweating, clothe rubbing or skin absorption, requiring several applications per day, and thus are limited in duration and rely on user compliance [8,9]. ITNs and IRS are inexpensive, effective indoors and when individuals remain in the confines of the treated areas (building), however neither are an option for outdoors or active users outside of the residence. [10].
Spatial repellent products include electrical repellents, passive emanators based on impregnated strips, candles, and, coils [11]. Electrical repellents, including motorized fans, require power to operate, making their widespread use limited. In addition, many of these devices require frequent cartridge replacement. Passive emanators using impregnated strips are limited to a single compound showing limited efficacy only for short periods of time [12,13]. Candles are based on essential oils, whose efficacy has been questioned [14]. Furthermore, candles and coils require an open flame or ember to operate, representing a fire hazard, and are limited in duration, requiring frequent replacement. Coils are the most widely used products for low income countries, where electricity is not widespread. Currently, the number of coils sold worldwide is estimated at 30 billion yearly at estimated price range of US$ 0.025–0.1 per unit [15]. Their competitive cost together with their ease of use provides compelling reasons to be considered as one of the most widely accepted consumer products for vector control. Coils exhibit a number of disadvantages, including: the Active Ingredient (AI) represents less than 1% of the overall device volume or mass; it requires an open flame starter; it is a fire hazard during operation; it poses respiratory-related health risks, and it lacks continuous release that limits their use to 4–12 hours per unit [16,17].
In addition to the available armamentarium of spatial repellent devices, active devices based on Micro-Electro-Mechanical Systems (MEMS), have been reported as a potential tool against mosquitoes [18]. In spite of their remarkable advantages, including low cost, batch fabrication, small form factor and the ability to actively control the delivery process, MEMS require power to operate, and they are limited in duration due to their limited payload. Therefore, there is a current need for a flameless, passive, cost effective, multi-chemistry delivery method for spatial repellency that provides sustained and safe protection over weeks in order to increase adherence of use, and ultimately further reduce risk of vector-borne diseases. Controlled Release Devices (CRDs) were developed to overcome some of the limitations of current spatial repellent (SR) delivery systems.
Present CRD design uses a set of wells where formulated AIs are placed, each well is capped by a lid that forms a chamber. The chamber lids integrate a set of tightly calibrated pores to control the release of the formulated AI. The CRD also includes an exothermic process that is activated when the device is open and exposed to air via inlets to provide an increase release rate. Fig 1A and 1B show an image of a CRD and its internal components.
The advantages of the CRD are: no electrical power, no battery; no open flame, making it safer than coils and candles; manufacturable with biodegradable materials, making it environmentally friendly; it can last up to two weeks; it can store multiple AIs. Present CRD design can be used in indoors or outdoors applications, and provides a cost-effective solution that can be mass produced and easily deployed.
To maximize CRDs performance, high volatility is required in the AI to be used while it has to be effective against Mosquitoes. As pyrethroids compounds are known to have high efficacy in protection against mosquitoes, metofluthrin was chosen [19]. Metofluthrin is commonly referred to as a Spatial Repellent (SR), but it is in fact considered an insecticide by the Environmental Protection Agency (EPA), which has oversight to approve new AIs in the US.
Herein we describe the systematic approach used in the CRD development, summarized in Fig 2. The first step was the physical characterization of several metofluthrin formulations to the evaporation rate of the AI. This data is then fed into a Computational Fluid Dynamics (CFD) model to predict estimated concentrations of AI in the space around the device. First generation CRDs were manufactured using 3D printing Stereo Lithography Apparatus (SLA) technology for rapid prototyping and eventually second generation CRDs were subsequently manufactured using micro-injection molding, a cost effective technique for mass production. CRDs were tested in an arm-in-cage study, performed to determine the optimal metofluthrin concentrations in a 15 minute trial. Knockdown and bite inhibition were used as entomological endpoints. Once the optimal metofluthrin formulation was determined, CRDs were tested with such formulation for semi-field studies in order to correlate the mortality with the AI concentration in space for 24 and 48 hour periods. Mortality was selected as the preferred entomological endpoint since it can be used to correlate estimated concentration with highly specific spatial locations for a given time point. The established correlation between mortality and estimated metofluthrin concentration can be used as a tool to further tailor the CRD design and form of deployment, e.g. number and distribution of CRDs, in order to target a protective volume with a desired mortality. Lastly, outdoor studies were performed to evaluate the performance of CRDs.
Fig 2 shows a summary of the process flow detailing the used methodology.
Physical characterization of metofluthrin solutions was performed in order to characterize metofluthrin release rates. 50 mL falcon tubes containing metofluthrin with isopropyl alcohol (IPA) as a solvent and at the following concentrations: 1%, 5%, 10%, 30%, 50% (v/v), were left open for evaporation rate determination at room temperature conditions. IPA at 99% v/v, was used as solvent to increase metofluthrin volatility. IPA was chosen as solvent due to metofluthrin high solubility in it (313.2 gr/L), high volatility and low toxicity [19]. The total mass and total volume of remaining metofluthrin solution in each sample tube for 12, 24, 48 and 96 hours was measured using analytical balance and calibrated volumetric measurements under laboratory controlled temperature conditions (20–25°C). Changes in relative concentrations of metofluthrin and IPA were determined via Mass Spectroscopy (MS), estimating concentration of active isomers of metofluthrin E and Z. Using the same approach, the release rate of metofluthrin was determined. Water was estimated in the solution due to the hygroscopic nature of metofluthrin and IPA. Estimated metofluthrin release rates as a function of concentration were calculated from these studies. The rates were then scaled to the device geometry, calculated by multiplying by the ratio of the surface areas of the CRD to the falcon tube, in order to serve as input for the in-silico model to model spatial concentration evolution with time.
CRDs preliminary versions were 3D printed, then manufactured using micro-injection molding in Zytel st801. CRDs design consisted of an exothermic reactor bottom containing a 50 mL reservoir to store an exothermic material (based on iron oxides) that acted as a heat source to increase volatilization of AI after activation and hence distribute AI in the volume to be protected faster. The device also included a middle reservoir array that included seven 0.5 mL reservoirs to store the AI. The device also included a top layer comprising a membrane with 30 inlet pores, each 1 mm in diameter, to allow oxygen diffusion into the exothermic reactor, and 1,000 outlet pores, each with 200 μm diameter pores for controlled release of the AI.
The exothermic reactor relied on oxygen to start the exothermic reaction, leading to a local increase in temperature of up to 55°C for eight hours and thereby enhancing the volatilization of the AI. Once cooled down, device mass rate was high enough to keep the protective volume until AI depletion.
Controlled release relies on evaporation of the AI formulation through the membrane pores until the AI reservoirs are depleted. The release rate of AI is dependent on the open surface area. Metofluthrin, an effective spatial repellent, was selected due its high efficacy, low toxicity profile and relatively high vapor pressure at room temperature [20]. CRDs were vacuum sealed, and an oxygen absorber was incorporated inside the package to further reduce the presence of any oxygen to prevent the activation of the exothermic reaction.
An in-silico model was developed to estimate metofluthrin spatial concentration distribution in a domain to match the observed mosquito mortality found during the entomological tests. The semi-field studies in tents configuration was chosen to reduce the air movement uncertainty that arises in an open domain. The domain consisted of an exterior volume where wind actual measured average speed and direction conditions were set. The tent was placed inside the domain with open doors to allow air to enter the interior domain. Fans present during the test were modelled inserting a speed jump at the fan location enforcing fan flow. Mass point sources were placed at the proper location with the measured device mass rate. A device is considered a point source due to its relative small size compared to tent volume. The model was developed using Computational Fluid Dynamics techniques in ANSYS [21] with k-ε turbulence for flow simulation and an implicit scalar transport scheme to tackle metofluthrin advection and diffusion in a transient solution. Changes in temperature do not affect significantly AI distribution, since natural convection dominates the mass transfer. From a fluid mechanics perspective, the estimated low AI concentrations will not lead to any buoyancy effect.
To characterize metofluthrin release rates for each of the tested formulations, as shown in Fig 3A, metofluthrin (AI), IPA (solvent), and water mass were integrated in the calculations to determine the fraction of each of these components as a function of time. Fig 3B shows the mass of IPA as a function of time, Fig 3C shows the mass of water as a function of time, and Fig 3D shows the mass of metofluthrin as a function of time. Water content was studied together with IPA mass and metofluthrin mass due to its hygroscopicity.
IPA evaporation rate was similar for the proposed formulations, while metofluthrin increased its evaporation rate with the IPA fraction, which suggests that IPA increases the metofluthrin evaporation rate. Metofluthrin evaporation rate was calculated as the change in mass over time, shown in Fig 3E. In this plot, it is possible to observe that all curves converge towards the same rate level, which represents the 100% metofluthrin (beyond 48 hours).
Selection of the optimal concentration to be used in the CRDs required the Arm-in-Cage studies as the first step. Experimental setup is shown in Fig 4A and 4B. Fig 4C shows the average percentage knockdown and Fig 4D shows the average percentage of mosquito bites as a function of concentration, ranging from 1% to 100% metofluthrin. It is possible to observe that knockdown increases with metofluthrin concentration while bites are reduced, with more data dispersion in knockdown than in bites.
A test with a combination of the two best performing concentrations was performed. One device with 30% metofluthrin and other one with 100% metofluthrin were tested. It was found that such combination performed even better than the best result obtained with two identical CRDs. This improvement may be attributed to the complementary of the different release kinetic profiles. One possible interpretation is that the first period of high evaporation was mainly driven by the 30% metofluthrin formulation, complemented by a second period of evaporation primarily driven by the 100% metofluthrin.
Semi-field studies were carried out in tents after the optimization process performed in arm-in-cage studies. Fig 5A summarizes the experimental setup. Experimental results are shown in Fig 5B for 24 hours and in Fig 5C for 48 hours (N = 4). For each tent location a bar plot showing mosquito mortality at each of the three defined heights is shown, together with their average mortality per tent location (yellow) and the average control defined as a 100% IPA loaded CRD (green). A uniform mortality distribution is observed though locations and heights. A minimum mortality of 50% can be observed in the first 24 hours, reaching 75% in 48 hours, with no significant differences between locations and heights. From the CFD model, simulated metofluthrin concentrations at each location and height are plotted for 24 hours and 48 hours in Fig 5D and 5E, respectively.
Furthermore, the estimated spatial concentrations obtained from the CFD simulations versus mortality for every location are plotted for 24 hours and 48 hours in Fig 6A and 6B, respectively. When analyzing this correlation plot, it was found that the lower and middle level pouches seem to require less concentration to reach the same mortality as the higher level pouches. Less concentration was required to be effective in 48 hours than in 24 hours, which could be attributed to the longer exposure of the mosquitoes to the metofluthrin.
Based on these plots, a linear regression analysis was performed to establish the correlation between mosquito mortality and metofluthrin concentration. The metofluthrin concentration for 100% mortality can therefore be calculated, resulting in 0.234 ppm for 24 hours and 0.097 ppm for 48 hours. Fig 6C shows the estimated iso-surface plots plotted for 24 and 48 hours. These iso-surfaces provide the spatial limit concentration found such that inside these convex surfaces a targeted mosquito mortality of 100% would be guaranteed.
Finally, Fig 7A and 7B shows the experimental set up for the outdoor experiments and Fig 7C the spatial location of the pouches. The bar chart plotted in Fig 7D shows the average mosquito mortality per pouch for the distances and heights evaluated from the CRDs. Results were plotted to show average mosquito mortality per level for a given distance from the source. Controls were also shown. A slight distance effect is observed showing a decrease in mortality with almost no dependence on pouch height. Mortality dispersion in not significant across directions, even considering the unrestricted air movement.
In silico modeling provides a powerful multi-dimensional tool to estimate AI concentration as a function of targeted mortality, for a desired 3D space in a given period of time and environmental conditions, which include 3D boundary conditions, temperature, and wind velocity. A correlation was established between simulated average concentrations and mosquito mortality obtained from semi-field experiments. Additional tests to address repellency and bite inhibition in open spaces, such as Human Landed Catches, could be performed in the future. Both parameters are associated with protective surfaces what makes them more complex to correlate with a spatial distribution.
The ability to correlate metofluthrin concentration with mosquito mortality in a 3D space as a function of time could potentially allow to customize SR delivery based on a target mortality and defined environmental conditions over a defined region. It is therefore possible to obtain predictive behavior of devices in terms of efficacy over a 3D space, defining a bubble of protection, over a period of time, while monitoring toxicity thresholds.
Arm-In-Cage trials demonstrated that formulations of metofluthrin with IPA at 30% and 100% provided the highest percentage of knockdown in the 40–50% range and bite inhibition in the 70–90% range. Semi-field studies were performed to show the performance of CRDs in a semi-outdoor environment. High mosquito mortality rates in the range of 60–90% relative to an independent control over a period of 24 and 48 hours validated the use of CRDs for protection against mosquito bites. It was also possible to plot the mortality (Fig 5B and 5C) per pouch, as well as simulations results for spatial concentration per pouch (Fig 5D and 5E) once the release rate for each device was empirically estimated. These projected concentrations allowed to estimate the performance of device as these values were correlated with the morality values. The established correlation provides a powerful tool for projecting device performance. Outdoor experiments provided another insight in the use of CRDs as a SR protection tool. Experimental results showed mosquito mortality in pouches in the range of 40–60% relative to an independent control over a period of 24 hours for distances of up to 2.5 m from the devices.
The presented design of CRD can accommodate volumes of up to 20 mL of AI, which represents approximately about half of the device total volume. Such a CRD version was designed to have an effective persistence (duration) of up to 2 weeks of continuous usage. The proposed cost of the device will be in the range of US$0.25–0.5 per unit for volumes of 1–10 million units. CRDs do not require electrical power, and do not constitute a fire hazard. Moreover, CRDs future material selections could include Mirel, a biodegradable polymer, or other selection of advanced polyhydroxyalkanoate polymers (PHAs), or even paper-based versions as an environmental friendly solution. Future device designs will include a small transparent indicator with a dye to show remaining device capacity.
The development and testing of a novel type of SR delivery system was introduced starting from idea conceptualization to formulation development, followed by in silico model, device design and manufacturing and entomological studies including arm-in-cage, semi-field and field experiments. CRDs were designed with a novel methodology that integrated combination of entomological endpoints and computational models to target efficacy and spatial protection as part of the design requirements. This multidisciplinary method allowed for a quantitative approach to device development and optimized performance. The design method allows for tailoring release kinetic profiles and field distributions for public health interventions in buildings, facilities and open areas. Spatial coverage can be achieved by the number and distribution of devices deployed. The duration can be tailored by the device capacity.
Experimental results have shown the potential use of CRDs as the next generation SR device. CRDs represent a simple and cost-effective solution for enhanced protection against vector-borne diseases.
|
10.1371/journal.pntd.0004957 | A Unified Framework for the Infection Dynamics of Zoonotic Spillover and Spread | A considerable amount of disease is transmitted from animals to humans and many of these zoonoses are neglected tropical diseases. As outbreaks of SARS, avian influenza and Ebola have demonstrated, however, zoonotic diseases are serious threats to global public health and are not just problems confined to remote regions. There are two fundamental, and poorly studied, stages of zoonotic disease emergence: ‘spillover’, i.e. transmission of pathogens from animals to humans, and ‘stuttering transmission’, i.e. when limited human-to-human infections occur, leading to self-limiting chains of transmission. We developed a transparent, theoretical framework, based on a generalization of Poisson processes with memory of past human infections, that unifies these stages. Once we have quantified pathogen dynamics in the reservoir, with some knowledge of the mechanism of contact, the approach provides a tool to estimate the likelihood of spillover events. Comparisons with independent agent-based models demonstrates the ability of the framework to correctly estimate the relative contributions of human-to-human vs animal transmission. As an illustrative example, we applied our model to Lassa fever, a rodent-borne, viral haemorrhagic disease common in West Africa, for which data on human outbreaks were available. The approach developed here is general and applicable to a range of zoonoses. This kind of methodology is of crucial importance for the scientific, medical and public health communities working at the interface between animal and human diseases to assess the risk associated with the disease and to plan intervention and appropriate control measures. The Lassa case study revealed important knowledge gaps, and opportunities, arising from limited knowledge of the temporal patterns in reporting, abundance of and infection prevalence in, the host reservoir.
| Many dangerous diseases emerge via spillover from animals, with limited human-to-human infection (stuttering-transmission) often being the first stage of human disease spread. Understanding the conditions (biological, environmental and socio-economic factors) that regulate spillover and disease spread is key to its mitigation. Here we are interested in questions such as: If we have quantified pathogen dynamics in the reservoir, with some knowledge of the mechanism of contact, can we estimate the likelihood of spillover events? Can we tease apart how much the disease is transmitted by animals and how much by humans? We developed a unified mathematical framework, based on Poisson processes with memory of past events, to understand the dynamics of spillover and stuttering-transmission. This framework, which can be applied across the disease transmission spectrum, allows the teasing apart of the disease burden attributed to animal-human and human-human transmission. Using this model, we can infer human disease risk based on knowledge of infection patterns in the animal reservoir host and the contact mechanisms required for transmission to humans.
| An important class of pathogens are those transmitted from animals to humans (zoonosis). The dangers associated with zoonotic pathogens are twofold. Firstly, the pathogen can adapt to the new human host and acquire the ability to transmit sustainably from human-to-human without the need for continued seeding from the animal reservoir. The pathogens involved occasionally transmit rapidly amongst its immunologically naïve new host causing devastating health impacts as demonstrated by the global SARS outbreak, the swine influenza pandemic and the recent Ebola epidemic, which probably originated from one zoonotic spillover event. Perhaps, however, HIV-1 is the most spectacular case of a recent zoonotic emergence, originating from an endemic infection of chimpanzees in Central Africa. Zoonotic infections are the origin of the majority of established human pathogens [1] of which influenza, measles, smallpox and diphtheria are examples [2]. Secondly, zoonotic pathogens can spill over from animal reservoirs continually and cause a heavy burden of disease. Human rabies from domestic dogs is an important and preventable example.
The origins of major human infectious diseases can be conceptualised as a continuous transition across different epidemiologic stages [3]. The first stage is when a pathogen exclusively infects animals (‘reservoir dynamics’). The second is when the pathogen occasionally jumps to the dead-end-host human population (‘spillover’). This is followed by a third stage, when human-to-human transmission becomes possible but leads only to self-limiting chains of transmission (‘stuttering transmission’). The final stage is when a pathogen gains the ability to transmit effectively between humans and no longer requires zoonotic transmission [3]. An additional scenario is when the pathogen infects both animals and humans in a sustainable manner.
Measuring and predicting cross-species transmission is extremely difficult. This is because spillovers are often, but not always (as the situation for Lassa Fever demonstrates), rare events driven by the complex interactions of multiple causes, including ecological factors (e.g. presence of hosts with differing degrees of susceptibility and periodicity in their abundance), epidemiological and genetic factors (e.g. a broad set of pathogen life histories and periodicity of infection prevalence), and anthropogenic activities (e.g. land-use and behavioural changes affecting direct and indirect interactions with reservoir hosts) [4]. Particularly challenging are zoonoses with stuttering transmission, as separating the contribution of animal-to-human from human-to-human transmission is extremely difficult.
Not surprisingly, theoretical [5–7] and experimental studies able to disentangle the many complex aspects of transmission at the animal-human interface are scarce [3, 8]. An increasing body of research recognises the need for a new paradigm integrating biological, social and environmental sciences with mathematical modelling to explain the mechanisms and impacts of zoonotic emergence [9, 10].
Understanding zoonotic spillover and stuttering transmission are, therefore, two very important public health challenges. The scientific, public health and medical communities working at the interface of animal and human pathogens are challenged with many questions, such as: i) if we know the pathogen abundance and prevalence in the reservoir and have some knowledge of the mechanism of contact, can we estimate the likelihood of the next spillover event? ii) is there a signature in the patterns of disease occurrence that enables us to distinguish the spillover (animal-to-human) burden from the stuttering chain (human-to-human) burden? iii) how is zoonotic risk driven by specific social, economic, environmental and biological factors?
Mathematical modelling has been used before to estimate the relative contributions of zoonotic spillover and human-to-human transmission, [11]. This approach was based on rarely available information of nosocomial and extra-nosocomial outbreaks that were known to be instances of pure human-to-human chains. More general methods are needed. Here, we developed a unified mathematical framework for spillover and stuttering chain processes. These are conceptually similar mechanisms; they are both arrival processes. The key difference is that zoonotic spillovers are assumed to arise from random and independent contacts with the reservoir with no influence of past infections (assuming no depletion of susceptibles, i.e. the pool of people who can be infected by contact with the reservoir or humans). In contrast, a stuttering chain, which arises from human-to-human transmission, is affected by the number of past human infections as each infected person can also trigger a chain of new cases. Zoonotic spillovers are also affected by past events when depletion of susceptibles, through death or development of sterilising immunity, is important.
Mathematically, zoonotic spillovers are described by Poisson processes (Cox processes if stochasticity in the rate of infection becomes important) or by self-correcting (i.e. decreasing rate of infection) processes if depletion of susceptibles occurs, while stuttering chains are described by a combination of self-exciting (i.e. increasing rate of infection), due to previous human infections, and self-correcting due to depletion of susceptibles, processes (see Table S1 in S2 Text). We tested different models by comparing their predictions with the corresponding outputs from independently-simulated epidemics generated by an agent based model (ABM). As an illustrative example, we also applied the final model to Lassa Fever (LF), a zoonotic, viral haemorrhagic disease common in West Africa, for which data from Kenema Government Hospital (KGH) in Sierra Leone [12] are available. LF represents an important model for this kind of study (Fig 1). The disease reservoir is Mastomys natalensis [13], one of the most common African rodents, but an important proportion of the burden of disease is ascribable to human-to-human transmission; this is supported by the arguments presented in [11] and by a recent case of secondary transmission of locally acquired Lassa fever in Cologne, Germany [14]. This case study was particularly instructive, revealing important challenges in current knowledge of LF, thus informing the direction of future research.
The proposed method is general but for illustrative purposes the presentation is based on the LF system exemplified in Fig 1. For the LF case study, we used data abstracted from patient medical records and LF diagnostic tests for 1002 suspected LF cases who presented to the KGH Lassa Ward from 27th of April 2010 to the 31st of January 2012 [11, 12]. For the list of symbols see also Supporting Information, S1 Text.
The phenomenology of spillover events ought to be linked with disease dynamics in the reservoir and the mechanism of contact between species. We assume that LF is caused by independent random ‘contacts’ (mediated by contaminated food, fomites etc.) between humans and rodents. Thus the probability P that k events occur during a time τ (e.g. number of admissions to hospital in one week) can be described by a stochastic Poisson process:
P ( k ) = exp - λ τ ( λ τ ) k k ! (1)
where λ is a parameter (rate) representing the expected number of zoonotic spillovers per time unit. The parameter λ is expected to depend on other drivers [15]. In the simplest scenario the human population is uniformly subjected to random and independent contacts with the reservoir. Only a fraction of these contacts, equal to the infection prevalence of the reservoir, are a potential source of infection. Accordingly, we assume:
λ = N H P r R ( N R ) χ R η R ( N R ) (2)
where NH is the human population size, i.e. the total number of people in a suitable area A, e.g. a village; PrR(NR) is the prevalence of infected rodents; χR is a parameter combining two complex mechanisms: the ability of the reservoir to excrete a suitable dosage of the virus and the human response to it. We refer to this parameter as infection-response efficiency, and we formally define it as the product of the probability that the virus is excreted from the reservoir and the probability that humans acquire infection when challenged with the virus. ηR(NR) is a measure of exposure, given by the product η R ( N R ) = ξ ( N R ) A where ξ(NR) is the probability of a contact (direct or mediated) between a single member of the human population and the population of NR rodents per time unit and area unit. Both the pathogen prevalence, PrR, and the exposure, ηR, are expected to be functions of rodent abundance, NR, although a clear evidence of correlation between LASV prevalence and M. nataliensis abundance is lacking. The area A essentially depends on the dispersal range of the rodents and, in the presence of human-to-human transmission, on the mobility of people. Here we assumed that the area A used is suitable for considering the system closed (no change in the population apart from the disease induced mortality) and for assuming uniform mixing, i.e. each person is equally in contact with each other and with the rodent population. As in the current model the size A of the system is fixed, we consider the overall parameter ηR(NR). Here and throughout, we refer to the quantities NH, PrR(NR), χR, ηR(NR) (and also χN and ηR(NH) defined below) as constituent factors.
The assumption that the system is closed can be relaxed. The simplest approach would be capturing the phenomenology of births, deaths and migrations by allowing a time-dependent functional form for the human population size NH = NH(t). Alternatively, changes in the human population size can arise from implementing an appropriate population dynamics model for NH. The approach can be further extended to incorporate explicitly-spatial effects by building an interconnected meta-population model based on homogeneous regions and allowing immigration/emigration of individuals.
Quantities such as the rodent population size, NR, and infection prevalence, Pr, are often seasonal therefore the rate λ ought to be explicitly time-dependent resulting in a non-homogeneous Poisson process. Most importantly, all the terms in Eq (2), i.e. rodent population, NR, infection prevalence, Pr, human population size, NH, and the infection-response efficiency, χR, are in general, stochastic. Thus the parameter λ in Eq (1) should be replaced with a random variable leading to the so-called doubly stochastic or Cox process. When the rate λ is a gamma-distributed variable, the Cox process is described by a negative binomial distribution (S3 Text). After some algebra based on well-known properties of the negative binomial distribution, we can present further relationships between some parameters of the negative binomial distribution (including mean μ and variance σ2 that uniquely determine the distribution) and the mean μλ and variance σ λ 2 of the associated gamma-distribution for the rate λ (i.e. μ = μλ, σ 2 = σ λ 2 + μ λ, see Table S1 in S3 Text).
As is known, when σ λ 2 approaches zero, then the negative binomial approaches a standard Poisson distribution. The properties shown in Table S1 in S3 Text, however, have important implications for quantifying the risk of spillovers. To estimate the probability of a spillover, it is sufficient to know the value of the parameters μ and σ2. These, in turn, can be estimated from the mean and variance, μλ and σ λ 2, in the rate λ, which, ultimately depend on the constituent factors.
Based on the hypothesis posed in Eq (2), we show how to infer the mean and variances μλ and σ λ 2 directly from knowledge of the human population size, NH, the abundance of rodents, NR, and also the infection-response efficiency, χR. Since we expect that NR, NH and χR are independent random variables, the mean value of the rate λ is given by the product μ λ = μ N H μ η R P r R μ χ R, where μ N H and μ χ R are the mean values associated with the size of the human population, NH, and the infection-response efficiency, χR; μ η R P r R is the mean value of the random variable arising from the product ηR(NR)PrR(NR), i.e. the exposure to the infected reservoir only (while ηR(NR) is the ‘exposure to the reservoir’, irrespective of this being infected or not). Similarly, the variance σ λ 2 can be estimated as
σ λ 2 ≈ η R ( N R ) P r R ( N R ) χ R 2 σ N H 2 + N H χ R ∂ σ ( N R ) P r R ( N R ) ∂ N R 2 σ N R 2 + N H η R ( N R ) P r R ( N R ) 2 σ χ R 2 (3)
where we used the usual approximation:
σ f 2 ≈ ∂ f ∂ X 2 σ X 2 + ∂ f ∂ Y 2 σ Y 2 + 2 ∂ f ∂ X ∂ f ∂ Y cov X Y . (4)
for a function of two random variables X and Y where NR, NH and χR are independent. Of course, if σ N H 2 ≈ σ N R 2 ≈ σ χ R 2 ≈ 0 then the spillover process is well approximated by a standard Poisson process.
In some situations the explicit dependency of the quantity ηR(NR)PrR(NR) on the abundance of the reservoir is known or can be crudely estimated. Then, the mean and variance μλ and σ λ 2 can be evaluated directly from the NR as shown for a range of relevant situations in Table S1 in S4 Text (see also Davis et al. [15]).
The model above was derived with the assumption that the number of susceptibles is constant. In a small population, however, the depletion of susceptibles is expected to be an important effect that can result in a self-constraining epidemic. Following model Eq (1), we replaced the (fixed) size of the human population NH with the (variable) number of susceptibles, SH. Thus, the probability of observing k cases at any time tj during the interval [(j − 1)τ, jτ] (with tj ∈ [(j − 1)τ, jτ]) is the piecewise function defined on discrete intervals:
P ˜ ( k , t j ) = exp − λ ˜ j − 1 τ ( λ ˜ j − 1 τ ) k k ! with rate λ ˜ j = S H ( t j ) η R ( N R ) P r R ( N R ) χ R (5)
where the time-dependent terms at time tj are estimated at the end of the previous interval [(j − 1)τ, jτ]. Underlying this choice is the assumption that the time step τ is comparable to the transition time from the susceptible to non-susceptible category, and λj can be considered constant during this time interval. To estimate SH(tj), we considered the case of an initially susceptible population. For simplicity we assumed no external immigration and that the size of the human population at the initial time is SH(0) = NH. As soon as spillover events start, part of the human population becomes infected; some with resulting life-time immunity and others die. As we consider a closed human population, the number of susceptibles is:
S H ( t j ) = N H - C H ( t j ) if N H > C H ( t j ) 0 otherwise (6)
where CH(jτ) represents the cumulative number of people who had been infected at any past time during the interval [0, jτ], irrespective of if they recovered or died. This corresponds to:
C H ( t j ) = C H ( t j - 1 ) + E [ P ˜ ( k , t j ) ] (7)
where E [ P ˜ ( k , t j ) ] is the expected number of spillover events during the time-interval [(j − 1)τ, jτ], as E [ P ˜ ( k , t j ) ] = λ ˜ j - 1 τ, thus we have
C H ( t j ) = C H ( t j - 1 ) + S H ( t j - 1 ) η R ( N R ) P r R ( N R ) χ R τ (8)
The probability P ˜ ( k , t j ) at time tj in Eq (5) can be iteratively calculated by replacing the susceptible and cumulative infected, SH and CH, with their explicit expressions given in Eqs (6) and (8) estimated at the previous time tj−1. Of course, if the depletion of susceptibles is negligible then S(tj) ≈ NH and the model collapses to a standard Poisson process or Cox-process if we allow for stochasticity in the rate. Eq (5) is a particular case of a class of models known Hawkes point processes (see [16] and references therein). We refer to these processes as ‘zoonotic spillover with depletion of susceptibles’ (in mathematical parlance ‘Self-Correcting Poisson’).
Hawkes point processes introduced above suggest a natural extension of the current model to include human-to-human transmission. In this context each infection event at time tj, represented by IH(tj), has a certain probability of generating new events. Accordingly, the probability of observing k cases at any time tj ∈ [(j − 1)τ, jτ] is the piecewise function:
P ^ ( k , t j ) = exp − λ ^ j − 1 τ ( λ ^ j − 1 τ ) k k ! λ ^ j = S H ( t j ) η R ( N R ) P r R ( N R ) χ R ︷ zoonosis + S H ( t j ) η H ( N H ) P r H ( N H , t j ) χ H ︷ human - to - human P r H ( N H , t j ) = I H ( t j ) S H ( t j ) + I H ( t j ) + R H ( t j )
(9)
where ηH(NH) is the probability that a single person is in contact with any other member of the human population per time unit; χH is the human analogue of the reservoir infection-response efficiency, i.e. the product of the probability that the virus is excreted from a person and the probability that a person acquires infection when exposed to the virus; PrH(NH) is the infection prevalence in the human population, which is the proportion of infected members IH(tj) in relation to the total size of the current population, i.e. for an SIR-type of model SH(tj) + IH(tj) + RH(tj) where RH(tj) is the number of recovered individuals. SH(tj) is given by Eq (6) with
C H ( t j ) = C H ( t j - 1 ) + E [ P ^ ( k , t j ) ] (10)
where E [ P ^ ( k , t j ) ] is the expected number of spillover events during the time-interval [(j − 1)τ, jτ], as E [ P ^ ( k , t i ) ] = λ ^ i - 1 τ, thus we have
C H ( t j ) = C H ( t j − 1 ) + I H z o o n + I H h − h I H z o o n = [ N H − C H ( t j − 1 ) ] η R ( N R ) P r R ( N R ) χ R τ ︷ zoonosis + I H h − h = [ N H − C H ( t j − 1 ) ] η H ( N H ) I H ( t j − 1 ) S H ( t j − 1 ) + I H ( t j − 1 ) + R H ( t j − 1 ) χ H τ ︷ human - to - human until N H ≥ C H ( t j − 1 ) (11)
C H z o o n ( t j ) = ∑ j I H z o o n ( t j ) represents the cumulative number of infections up to time tj due to zoonotic spillover and C H h - h ( t j )= ∑ j I H h - h ( t j ) represents the cumulative number of infections up to time tj arising from human-to-human transmission. The model requires the further condition:
I H ( t j ) = C H ( t j ) − ∑ i [ R H ( t j ) + D H ( t j ) ] R H ( t j ) = R H [ t j − 1 ] + γ r I H [ t j − 1 ] τ D H ( t j ) = D H [ t j − 1 ] + γ d I H [ t j − 1 ] τ
(12)
where DH(tj) is the disease induced mortality, γr and γd are the recovery and mortality rates respectively.
The model Eqs (9)–12 can be interpreted as an immigration-birth process [16] where the immigrants, i.e. zoonotic spillovers, arrive according to a Poisson process with rate λ ^ ( t j ). Each immigrant produces ‘offspring’, which by analogy is really new infections from human-to-human transmission leading to a stuttering chain, according to a rate which is dependent on past events. The model is a mixture of a self-exciting process (new cases generate subsequent cases (offspring)) and a self-correcting process (due to depletion of susceptibles). We refer to this type of processes as ‘zoonotic spillover with human-to-human transmission’ (in mathematical terms ‘Poisson with Feedback’). We also considered the case when the rate λ ^ ( t j ) is drawn from a gamma-distribution, i.e. ‘zoonotic spillover with human-to-human transmission when random effect in the rate are important’ (mathematically ‘Poisson-Gamma Mixture with Feedback’, Table S1 in S2 Text).
The probability P ^ ( k , t j ) at time tj in Eq (9) can be iteratively calculated by replacing the susceptible and infected, SH and IH, with their explicit expressions given in Eqs (6), (10)–12 estimated at the previous time tj−1. The contribution of human-to-human transmission at any time tj, Q(tj), can be readily calculated by comparing the cumulative number of infections due to zoonotic transmission terms to those due to human-to-human transmission in Eq (11), for example by studying the quantity:
Q ( t j ) = C H h - h ( t j ) C H ( t j ) (13)
To simplify the notation, we use the symbols ζ = PrR(NR)χR ηR and κ = χH ηH for the overall unknown parameters, and refer to these as ‘zoonotic exposure’ (which incorporates the host infection prevalence) and ‘effective human exposure’ respectively. We also define the forces of infection from animal or human source as ΛR = NHumans PrR(NR)χR ηR and ΛH = NHumans PrH(NH)χH ηH respectively.
Variation in the population size NH was also considered by discussing how the analytical solutions for the cumulative number of infections scales with the population size and by analysing predictions for NH = 1000 and NH = 2000 (S7 and S10 Texts, for the value of the parameters used in the numerics see Table 1 and Table S1 in S6 Text).
We considered a set of NH agents. Each agent being in one of four possible categories: susceptible, infected, recovered or dead. At any time step, susceptible agents can transit to the infected category, while infected agents can either recover or die. This is essentially a Bernoulli trial, e.g. a random process with exactly two possible outcomes. The transition from the susceptible to the infected status is therefore mimicked by simulating, at any time tj, a number of Bernoulli trials (SH(tj) or NH if we assume no depletion of susceptibles) with probability given by the appropriate rate divided by the number of trials. For instance, if we considered spillover and human-to-human transmission with depletion of susceptibles, the probability is λ ^ ( t j ) τ / S H ( t j ). This choice ensures that, at any time tj, if the number SH(tj) is large, then the corresponding set of Bernoulli trials are well approximated by a Poisson process with rate λ ^ ( t j ) τ. Similarly, infected agents die or recover by simulating IH(tj) statistically independent Bernoulli trials with probabilities γdτ/IH(tj) and γrτ/IH(tj) respectively.
The importance of mathematical modelling to elucidate the complexity of infectious disease dynamics and to indicate new approaches to prevention and control is widely accepted (see e.g. [17]). The task is not free of challenges, especially for emerging diseases [18]. This is further complicated by abiotic factors such as land use change [4] and social difference demonstrating how risks are not generalisable [19]. Here, we start proposing some measures for the risk of zoonotic spillover and their link with drivers of transmission. Then we present predictions for the model compared with predictions from an independent ABM. Finally we apply the model to LF data, illustrating important challenges and knowledge gaps.
The mean time between two spillover events and the probability of observing k spillovers during a certain time τ are suitable measures for the risk of cross-species transmission that naturally arise from the present mathematical framework. Based on the findings above, the risk of a spillover event can be represented by a discrete probability distribution, which can be generally described by a negative binomial distribution. This is fully identified by the mean and variance, empirically inferred or calculated from the mean and variance associated with the rate of infection λ as displayed in Table S1 in S3 text.
In some situations, we know how the exposure to the reservoir and its infection prevalence depend on the abundance of the reservoir NR. For example, it is reasonable to expect the exposure ηR(NR) is proportional to the reservoir abundance NR. The dependence of the infection prevalence on NR can also be inferred for many regimes at the endemic equilibrium, e.g. frequency and density dependent Susceptible Infected Removed (SIR), Susceptible Exposed Infected Removed (SEIR), etc. models (see Table S1 in S4 Text). For these cases, calculation of the mean and variance μλ and σ λ 2 is straightforward. In Table S1 in S4 Text, we consider four illustrative scenarios. In many situations the mean risk of spillover increases with the size of the human population NH. The associated variance, however, increases with the square of NH. The dependency on the reservoir abundance NR is in general more complicated. For instance, in scenario 1 the variance in the risk of spillover, σ λ 2, increases with the square of the NR. In contrast, in scenario 2 the variance σ λ 2 is not affected by the abundance NR, while in scenario 4 it reaches an asymptotic value for large NR.
For pure zoonotic spillovers, there is no human-to-human transmission, therefore the rate of infection is not affected by the number of humans already infected. In some cases, variation in the number of susceptibles can be ignored, for example when the impact of immunity and/or mortality is negligible compared to the total population. In this case, every spillover event is independent of previous spillover events. Furthermore, the rate of infection is itself a stochastic quantity as random differences are expected from village to village and from time to time. If these stochastic differences are small, then the rate of infection can be well approximated by its mean value and the distribution of zoonotic spillover described by the well-known Poisson distribution. These stochastic fluctuations, however, can be important; in this case the distribution of zoonotic spillovers is better described by the so-called negative binomial distribution (Eq (S2) in S3 Text, which arises from simple Poisson processes after incorporating stochasticity in the rate of infection given that the distribution of the rates can be well approximated by a gamma-distribution). In this case, the variance of the number of zoonotic spillover events is always larger that their mean value, which is over-dispersion.
Accordingly, we ran the ABM to generate zoonotic infections by simulating NH random experiments (Bernoulli trials) with transition probability proportional to the force of infection from an animal source (i.e. ΛRτ/NH, see Table S1 in S5 Text). Fig 2 shows the cumulative number of zoonotic infections generated by the ABM compared with the corresponding theoretical model (expressed by Eq (1) or Eq (S2) in S3 Text, when random effects in the rate of infection become important). As expected, the profile for the cumulative number of occurrences averaged over the multiple stochastic realisations is linearly increasing with time with the slope given by the mean rate of infection. When the rate of infection is also stochastic, e.g. because the outbreaks occurred in different regions with different eco-epidemiological and socio-economic factors, larger deviations from the average profile are observed. This is the typical situation when the available data are aggregated at the national level without distinguishing the specific local factors.
In many situations, the contribution of net immigration, births and deaths (other than infection-induced) to the human population size is negligible, at least for short time-scales. Still, once a spillover occurs, the infected individual might either recover or die, but will never transit back to the susceptible category. Thus we considered the situation when the total number of individuals is fixed, but the number of susceptibles is decreasing due to the accumulation of spillover events resulting in immunity and/or mortality. As the number of infected increase, the pool of susceptibles decreases reducing the rate of new infections; in other words the process is ‘self-correcting’ (Eqs (5)–(8) or their generalization when random effects in the rate of infection become important).
Accordingly, we ran the ABM to generate zoonotic infections by simulating a number of Bernoulli trials, with number of trials being equal to the time-varying number of susceptibles, and transition probability proportional to the force of infection from animal an source (i.e. ΛRτ/SH, see Table S1 in S5 Text. Note the force of infection is time-dependent as the number of susceptibles is changing). Fig 3a shows the cumulative number of zoonotic infections generated by the ABM compared with the theoretical model (Eq (5), see also the analytical solution in S7 Text, for the particular case when the mortality and recovery rates are zero).
As expected, a key effect of incorporating depletion of susceptibles in the model is that the average cumulative number of occurrences always results in a concave (i.e. downward) function, provided that there is no birth/immigration of new susceptibles and no temporal variation of the exposure. This is because the rate at which spillover events occur decreases with time and the average size of the jumps in the sample path becomes smaller and smaller. Over time, the profile asymptotically approaches the size of the human population NH (here set to NH = 1000 unless stated otherwise). This is more pronounced for high values of the zoonotic force of infection ΛR.
The ability of a pathogen to transmit between people enables the generation of chains of infection. Fig 3b shows the cumulative number of infections due to only the human-to-human route of transmission. The infections are generated by the ABM by simulating NH Bernoulli trials with transition probability proportional to the force of infection from human source (i.e. ΛHτ/NH Table S1 in S5 Text).
The predictions are compared with the theoretical model (Eq (9)) with the conditions of no zoonotic spillover and no mortality or recovery. A human infection triggers new infections that, in turn, generate other new infections. In other words the process is ‘self-exciting’ and the cumulative number of infections increases exponentially with rate equal to the effective human exposure (κ, Eq (S10) in S7 Text). The presence of zoonotic spillover events leads to a qualitatively similar behaviour, resulting in convex (i.e. upward) average profiles for the cumulative number of infections with no upper bound (S8 Text). This because the rate of infection increases as the number of infections increase.
In general, both effects, self-correction due to depletion of susceptibles and self-excitation due to the impact of past infections on new chains of human-to-human transmission, are expected to play a role. The combined effects lead to an average profile for the cumulative number of occurrences that is initially convex until the depletion of susceptibles dominates the dynamics. This can be seen in Fig 3c which shows the cumulative number of infections for the combined ‘zoonotic and human-to-human’ model. The infections were generated by the ABM by simulating Bernoulli trials (with the number of trials equal to the time-varying number of susceptibles), with transition probability proportional to the force of infection from either animal or human source (i.e. ΛRτ/SH or by ΛHτ/SH, Table S1 in S5 Text). The predictions are compared with the corresponding theoretical model (Eq (9)). As expected, the cumulative number of infections increases as an S-shape function asymptotically approaching the human population size NH (exactly as a logistic function if there is no mortality or recovery, Eq (S10), in S7 Text).
Knowing the zoonotic exposure and effective human exposure, we can estimate the relative contributions of zoonotic spillover and human-to-human transmission, (more precisely, by substituting the values of the two exposures ζ and κ in Eq (13)). In general these exposures are not known, but can be estimated via common statistical techniques, such as Markov Chain Monte Carlo (MCMC). To validate the methodology, we ran the ABM for the combined zoonotic and human-to-human model as described in the above section and counted the number of infections arising from zoonotic transmission and those from human-to-human transmission. All the ABM-simulated infections (with no distinction of the route of transmission) were used as input into MCMC estimation [20, 21] of the zoonotic exposure rate and effective human exposure rate, which are otherwise unknown (i.e. the parameters ζ and κ). The MCMC-inferred parameters were used to calculate the cumulative number of infections due to zoonotic spillover and those due to human-to-human transmission (i.e. C H z o o n and C H h - h, according to Eq (10)) and compared with the corresponding cumulative number of infections generated from the ABM (Fig 4). There is a small discrepancy between the MCMC-inferred parameters and the ones imposed in the ABM (the medians of the two estimated parameters were respectively 0.055 and 0.008 vs 0.05 and 0.01). This is expected as the ABM simulates Bernoulli trials rather than Poisson processes and the discrepancy decreases with the number of simulated trials. The prediction improved for larger number of trials (S9 Text). Of course full agreement is expected when the number of trials approaches infinity. For the range of simulations considered here, the relative contributions of zoonotic spillover and human-to-human transmission are not affected by the human population size NH (Fig S1 in S10 Text).
It is instructive to show some challenges encountered from the application to LF. The key problem is the lack of information on the temporal dependency of the zoonotic exposure (i.e. the parameter ζ). We therefore considered two simple scenarios. Firstly we assumed a constant zoonotic exposure. Secondly, we allowed its variation in a piecewise linear (triangular) fashion (Fig 5) with the highest peak in March, corresponding to the hottest month in Sierra Leone. This choice is, perhaps, the simplest way to capture variation in the drivers of transmission such as temperature that might affect the abundance and prevalence of the rodents or human mobility. For simplicity we considered only two changes in the slope of this function during the time of the study; this is sufficient for the illustrative purposes of this exercise.
Fig 5 shows the cumulative number of occurrences for the the combined zoonotic and human-to-human model (see Fig S1 in S11 Text, for the corresponding model with a stochastic rate of infection) for constant and piecewise linear variation of the zoonotic exposure. In both situations, the unknown parameters were optimized with the KGH data using an MCMC approach in R [20, 21].
A qualitative inspection shows that both predictions are compatible with the empirical data. The one with constant parameters, however, requires an exceptionally large contribution of human-to-human transmission (≈ 90%) [11]. For the second scenario, agreement with the data requires a positive increase of zoonotic exposure ζ followed by a decrease after March 2011 resulting in approximately 22% of cases due to human-to-human transmission. This value, however, is no longer invariant by the human population size; further testing with different values of NH resulted in different proportions of human-to-human contribution.
A rigorous selection of the two models based on information criteria is problematic. In the current Bayesian context, the Watanabe-Akaike Information Criterion (WAIC) [22, 23] or the Deviance Information Criterion (DIC) [23, 24] appear as ideal tools, at least at first glance. Their suitability, however, is questioned in our specific situation. First, the time series of the number of Lassa cases violates the assumption of independence, which is an essential constraint for WAIC [23]. Lack of independence in the data appears to be a limitation for DIC too [23]. Furthermore, DIC is not appropriate for model selection with mixture models [23], which is the case here as the model comprise multiple populations, i.e. the set of individuals infected via a zoonotic route and those via human-to-human transmission at each time step. In addition, one of the parameters (the zoonotic exposure ζ) is time-dependent in one of the models (Fig 5). We are not aware of a rigorous, systematic assessment of the different information criteria in the presence of time-dependent parameters. Finally, an ideal information criterion should select the true model, when this is in the model set, and the closest one otherwise. This task does not appear to be always achieved by the many information criteria available in the literature. Here, we presented two exemplar models, but in the true model, the functional shape of the zoonotic exposure might be different from being constant or piecewise linear, also the effective human exposure might not be constant. For all these reasons, we think that it is more prudent to postpone any conclusion until more accurate data on exposure rates, rodent infection prevalence and reporting bias become available. Nevertheless, for indicative purposes we present the BIC scores for the two models (BIC = 30973.96 and BIC = 1228.164 respectively). Thus, given the above caveats, the model allowing a temporal variation in zoonotic exposure performs better that the one with constant exposure. This further suggests the importance of solid research in measuring and quantifying exposure.
Temporal variations in rodent abundance and LF virus prevalence were also introduced into the model according to the seasonal patterns observed in West Africa for M. natalensis captured inside houses and in the proximity of cultivation [25]. Using these data, the model predictions are not able to capture the initial convex shape in the cumulative number of spillovers (S12 Text).
Based on a systematic review of the literature, Lloyd-Smith et al. [3] pointed out that models incorporating spillover and stuttering transmission are rare. The authors found that only 2% of studies of directly-transmitted zoonoses included a mechanistic model of spillover transmission, while stuttering transmission was modelled only in approximately 4% of all studies. Here we developed a unified theoretical framework with the aim of filling this gap. Both zoonotic spillover and stuttering chains can be governed by arrival processes and modelled as generalized Poisson processes, with zoonotic spillover being a particular case of the general model when the probability of human-to-human transmission is null. Although the initial motivation of our work focused on spillovers and stuttering chains (basic reproductive number R0 < 1), there is no theoretical or practical impediment to use the model beyond the sub-critical regime (R0 > 1). Indeed the effective reproductive number for Lassa fever data is larger than one [11] (for the effective reproductive number for the simulation generated by the ABM see S13 Text). The theoretical unification of these processes is not a mere question of mathematical elegance, it is critically important for a meaningful comparison of the different stages of disease propagation.
Disentangling the contribution of animal-to-human from human-to-human transmission is of crucial importance to inform appropriate control measures. The shape of the cumulative number of occurrences can provide indications of the modes of transmission. A concave, saturating profile is an expected outcome due to depletion of susceptibles. In contrast, a convex region in the profile of cumulative number of occurrences suggests that human-to-human transmission plays an important role. Alternative explanations are possible. A convex shape in the cumulative number of occurrences might arise from temporal variations in the model parameters (e.g. probability of contact between humans and rodents, infection prevalence in rodents, infection-response efficiency) and/or in the human population size.
A fundamental gap in our current knowledge is the mechanisms governing the transition from spillover to stuttering chain to sustained transmission. Stuttering and established human-to-human transmission require the pathogen to have the ability to transmit from human-to-human resulting in a non-zero value for the parameter χH (the product of the probability that the virus is excreted from a person and the probability that a person acquires infection when exposed to the virus). Non-biological factors, however, may be involved in the shift from one stage to another. For example, in a sparse population with limited exposure to the reservoir, the disease can rapidly die out due to depletion of susceptibles and/or because the average time between two contacts is longer than the infectious period (stuttering chain scenario). If the size of the human population and/or the frequency of contacts increase, however, then uninterrupted chains of transmission are possible. Thus, the disease switches from a stuttering to a sustained chain of transmission. In the current framework, the conditions leading to this transition can be inferred and quantified by imposing no exposure to the reservoir and studying under which conditions in the parameter values the average solution of Eq (9) results in a fading or in an established non-zero time-series of events.
This work was inspired and guided by the One Health vision: a holistic approach that recognises the inter-connections among human health, animal health and the environment. Accordingly, the model was designed so that a wide range of environmental, biological, ecological, social, economic and political drivers could be readily incorporated. This was done by explicitly expressing the rate of transmission as a function of the constituent factors: the size of the human population, the prevalence of infection in the reservoir host, the probability per time unit of reservoir host-to-human and human-to-human contact, and/or the infection-response efficiency in the human when challenged with the pathogen. For example, complex social, economic and political drivers (e.g. demographic pressure, human mobility, etc.) could be translated and quantified in terms of their impact on the typical size of the human population exposed to the disease, i.e. the factor NH. Economic and behavioural drivers (e.g. the practice of burning fields after harvesting, driving M. natalensis towards villages, young boys catching rodents as a ludic activity, seasonal crowding of miners in dwellings) could, once these factors are researched, be expressed in terms of their effects on exposure to disease i.e. the factor ηR(NR) and ηR(NH). Ultimately, complex biological, physical, environmental and social factors can be expressed as factors that can be either measured or quantified via independent models and fed into the current, modular approach or integrated in a Bayesian hierarchical framework.
Being able to infer the likelihood of zoonotic spillover from basic information about the reservoir host and the exposed human population would help to address public health needs and also be of interest to the medical and scientific communities. Our approach addresses this by partitioning the rate of transmission into the product of the constituent factors: the effective human population size at risk, pathogen prevalence in, and human exposure to, the reservoir, and infection-response efficiency. Knowledge of these factors could be gathered, at least in principle, from data collection or other models. For example, despite practical challenges, novel tools for direct or indirect estimation of wildlife abundance/diversity are continuously being developed, e.g. remote sensing [26] and public engagement [27]. Also, understanding infection dynamics in reservoir hosts is key to understanding spillover dynamics. Despite logistical and financial challenges, there is an increasing body of research on infection dynamics in wildlife (e.g. hantavirus and rodents [28], viral pathogens in African lions [29], viruses in African bats [30], rabies in bats [31]). Quantifying the contact rate of people with the reservoir host and/or other humans is difficult and depends on the mode of transmission, but effective rates of exposure could be estimated from serological data. Of course these factors present a degree of stochasticity, explaining the over-dispersion in many ecological data and indicating that spillover events are governed by Cox processes rather than a simple Poisson process. The probability mass function of spillover events is well described by a negative binomial distribution, suggesting that the rate of transmission is, at least approximately, gamma-distributed (although alternative mechanisms might lead to a negative binomial, S14 Text). Under this assumption, the simple knowledge of mean and variance in the rate of transmission or in the constituent factors are sufficient to completely determine the probability of observing a certain number of spillover events in a particular time-window. Further simplifications are possible when rates of zoonotic exposure and pathogen prevalence in the reservoir host can be explicitly linked to host abundance, as shown for a range of relevant situations discussed in Table S1 in S4 Text. Summary statistics can be readily calculated as the mean rate of spillover events, the mean time between two spillover events and the associated variances. Further theoretical and empirical work in this broad area is essential to enable evidence based reduction of zoonotic disease burden.
Application to LF is an interesting example of un-identifiability [32], at least when the analysis was based on a visual inspection alone. Both assumptions (constant and piecewise linear trend in the zoonotic exposure) appear to be equally compatible with the empirical data and the effects of human-to-human transmission can be confounded with those due to temporal variations in the parameters. Nevertheless, the model based on a piecewise linear trend in the zoonotic exposure is the one selected by BIC, although we cannot rule out other models, such as temporally varying zoonotic exposure in a non-linear fashion. Such uncertainty is expected to be removed as soon as more accurate data on actual exposure rates, rodent infection prevalence, spatial distribution of human population size, and further information of reporting bias, become available. For instance, collecting longitudinal human and rodent serological data, at the same location, was the initial objective of the current consortium. However, it was unfortunately hampered by the recent Ebola outbreak. The accuracy of the spatial distribution of human population size is expected to increase and research in health-seeking behaviour is currently conducted in African settings to understand, among other questions, reporting bias. Measuring exposure is in general challenging, although some progress has been made (e.g. incidence of arthropod bites in England [33] can be used as proxy for exposure for a range of vector-borne diseases [33]). In general, we believe that inferring exposure requires the combined effort from different types of research, e.g. serological surveys, possibly compared with analogue infections from the same reservoir; studies on human behaviour and interaction with the reservoir; mechanistic models to mimic the exposure process. A parallel, interesting avenue of future research would be exploring how the functional form of the zoonotic and effective human exposure affect inference results. It is worth noticing that the proposed framework requires only a general knowledge of the functional form of these quantities (e.g. if the temporal profile of the zoonotic exposure is constant, periodic, linear) and not detailed measurements.
In a recent work Andersen et al. [34] generated a genomic catalogue of almost 200 LASV sequences from clinical and rodent reservoir samples. Sustained human-to-human transmission would cause a ladder-like genetic structure of the phylogenetic tree which was not observed in their study. Such structure, however, is not expected if most human-to-human transmissions are caused by super-spreaders as recently shown [11]. Understanding super-spreading events is perhaps one of the most compelling scientific challenges in the field of epidemiology. These include molecular techniques to detect them (e.g. sequencing the virus in persons known to be infected by a super spreader, quantifying possible differences in the viral load between super-spreaders and non super-spreaders), social science exercises to elucidate behaviour and contact patterns, biological and medical investigation to uncover the physiology of super-spreaders, and mathematical modelling to disentangle the complex interactions of these different factors.
When zoonotic exposure is time-dependent, the predictions are sensitive to the assumed size of the human population. This problem relates to difficult and long-standing questions of spatial scale, and how to enumerate the population at risk in models. In the current framework, the natural spatial scale is the typical spatial range of M. natalensis, and the human population size living in the region. Currently we have no detailed information on the location of patients. This is, however, changing as KGH has started to record more accurate information on the address of patients (rather than simply “Kenema” as done in most cases available to us). This information would allow a more realistic meta-population model based on small patches, corresponding to the range of activity of M. natalensis, where the human population size is known, the mass action assumption is expected to be more correct, and depletion of susceptibles more relevant. The meta-population model could be improved by allowing immigration/emigration of individuals.
Temporal variations in rodent abundance and LF virus prevalence were introduced in the model according to the seasonal patterns observed in West Africa [25]. The patterns of seasonality in exposure alone, however, cannot explain the particular shape of the cumulative number of cases in KGH data. Furthermore, preliminary social science and rodent ecology data collected by our consortium suggest increased dry season (December to April) exposure linked to intensive cultivation of wetlands for horticulture (see also [4]). The generality of our framework allows the incorporation of a variety of sources of temporal variation.
The impact of stochastic fluctuations on the risk of spillover within the eco-epidemiological systems was not fully studied here. Our work could be further extended to address specific questions like: i) are occasional, large, random bursts in reservoir host infection more likely to spillover into the human population than smaller, but highly correlated, fluctuations? ii) How do model predictions depend on the particular epidemiological model (e.g. SI, SEIR, inclusion of reservoir host carrying capacity)? Environmental stochasticity and external periodic drivers (e.g. seasonality in the reservoir host population or rainfall) can certainly resonate with the natural frequencies of the eco-system [35] with large effects on transmission dynamics in both reservoir host and the spillover populations. This non-trivial interaction between internal noise and external periodic drivers might explain why evidence of the trophic cascade hypothesis (e.g. large amounts of precipitation lead to increased resources, followed by increased rodent abundance and then to increased risk of epizootics and human cases) is so elusive [36].
For simplicity and to demonstrate a proof of concept, our predictions are based on the assumption of uniform mixing and, in most cases, a closed human population; i.e. each case from KGH is potentially in contact with each other case and with the rodent population, and no birth or immigration of individuals was allowed. Although this is generally a reasonable assumption when dealing with village communities, as there is high human mobility in Sierra Leone [37, 38], the model ought to be extended to include spatial variability, e.g. via a meta-population approach and/or linked on spatial environmental and habitat variables [39]. This is an important area where participatory modelling and ethnographic research [37, 40] is much needed to gather information on actual patterns of mobility and social networking, and hence potential contact patterns.
The impact of birth, death and mobility of individual is an an other important topic. Perhaps the simplest scenario is when there is a small immigration of infected into the community (so that the total number of individuals can be approximated as a constant). If this process is governed by a Poisson mechanism, then it is mathematically equivalent to spillover events. This will result in a mere replacement of the zoonotic spillover rate with an ‘effective’ one which incorporate both zoonotic spillover and immigration, with no qualitative change in the findings above. Other scenarios raise more intriguing questions. For example how would the shape of the cumulative number of cases and the proportion of human-to-human transmission change in the presence of a periodic immigration of batches of individual? How is this affected by the size of the batches and the number of infected in each batch? Is there a critical threshold in the flux of infected individuals and the typical time between two arrivals that can lead to persistence of the disease, even in a reservoir-free area? How is the spatial distribution of the disease affected by the particular patterns of human mobility? (e.g. by studying traces of bank notes it has been shown that trajectories in human mobility are described by Lèvy walks, see [41] and also [42–44]). These are crucial questions to be addressed in future research.
Measuring the ‘true’ incidence of disease, and therefore morbidity and mortality rates, is a common problem in epidemiology. This includes under-ascertainment arising when not all cases seek healthcare, under-reporting due to failure in the surveillance system, and reporting bias caused in the way the research was conducted [45]. Addressing this important issue is beyond the scope of the current work. We only highlight that in a situation when the zoonotic exposure and the probability of reporting are constant, then reporting bias can have a limited impact on inferring the proportion of human-to-human transmission as we have shown that this is not sensitive to the size of the population NH. Otherwise, marked Poisson processes [46], might be the natural extension of this framework to investigate the effect of reporting bias.
Of course, if KGH, and health authorities in general, could reduce reporting bias it would be highly beneficial to studies such as this. Such improvements include improving methods and increasing funding for contact tracing and rapid diagnostic test development. This is another area where social and ethnographic research is needed to elucidate people’s patterns of reporting and health seeking behaviour for LF, and the social, cultural and economic factors that affect this, so enabling a more accurate assessment of extent and sources of bias.
In conclusion, we developed a conceptually simple, rigorous and transparent framework unifying the fundamental mechanisms of zoonosis with the crucial advantage that, in general, the approach does not require intense numerical computations.
|
10.1371/journal.pcbi.1005321 | Using Chemical Reaction Kinetics to Predict Optimal Antibiotic Treatment Strategies | Identifying optimal dosing of antibiotics has proven challenging—some antibiotics are most effective when they are administered periodically at high doses, while others work best when minimizing concentration fluctuations. Mechanistic explanations for why antibiotics differ in their optimal dosing are lacking, limiting our ability to predict optimal therapy and leading to long and costly experiments. We use mathematical models that describe both bacterial growth and intracellular antibiotic-target binding to investigate the effects of fluctuating antibiotic concentrations on individual bacterial cells and bacterial populations. We show that physicochemical parameters, e.g. the rate of drug transmembrane diffusion and the antibiotic-target complex half-life are sufficient to explain which treatment strategy is most effective. If the drug-target complex dissociates rapidly, the antibiotic must be kept constantly at a concentration that prevents bacterial replication. If antibiotics cross bacterial cell envelopes slowly to reach their target, there is a delay in the onset of action that may be reduced by increasing initial antibiotic concentration. Finally, slow drug-target dissociation and slow diffusion out of cells act to prolong antibiotic effects, thereby allowing for less frequent dosing. Our model can be used as a tool in the rational design of treatment for bacterial infections. It is easily adaptable to other biological systems, e.g. HIV, malaria and cancer, where the effects of physiological fluctuations of drug concentration are also poorly understood.
| In this era of rising concerns about antibiotic resistance, the rational design of optimal antibiotic treatment regimens remains an important unrealized goal. At this time, the characteristics of antibiotic treatment regimens (e.g. dosing levels, treatment duration, route of administration) are determined largely based on costly in vivo experiments. The sheer number of possible dosing strategies that must be tested contributes to the delay and cost of the development of new drugs and may limit the feasibility of finding optimal regimen characteristics. Here, we demonstrate how modeling the chemical kinetics of drug-target binding can identify the best time-concentration profile of antibiotics. Using both analytical approaches and numerical simulations, we find that the physicochemical characteristics of drug-target binding are sufficient to explain the pharmacodynamics of commonly used antibiotics such as ampicillin, isoniazid and tetracycline. In practical terms, our models can be used as a tool in the rational design of treatment for bacterial infections. Because of the generality of drug-target binding kinetics, these approaches may also be adapted to other diseases where the effects of physiological fluctuations of drug concentration are also poorly understood, such as HIV, malaria and cancer.
| The rise of antibiotic resistance underlines the need for employing existing antibiotics prudently. Although antibiotic dosing regimens have been investigated for more than half a century [1], we do not yet have a sufficient understanding of the link between drug dosing and bacterial killing to design rational treatment strategies [2, 3]. Even for antibiotic regimens that have been standard of care, substantial improvements in dosing levels [4], treatment frequency [5] and treatment duration[6–8] have been made decades after their introduction. Most experimental and some clinical studies investigate antibiotic concentrations at a constant or at an average concentration. However, drug concentration at target tissues can fluctuate substantially over time. These fluctuations can influence the effectiveness of treatment, with the importance of such fluctuations differing substantially between classes of antibiotics [9].
Three alternative descriptions of effective antibiotic concentration are commonly used (so-called pharmacokinetic drivers): i) the total concentration integrated over a given time interval (area under the curve, AUC), ii) the peak concentration (Cmax) or iii) the time during which the concentration exceeds a specific threshold (time above MIC, TC>MIC, Fig 1A). For some drugs Cmax correlates best with bacterial clearance [10], for example in clinical trials with isoniazid [11]. Even once-weekly dosing was slightly superior to daily dosing for the novel TB drug bedaquiline when holding total drug administration constant [12]. For rifampicin [13] and quinolones, the total amount of drug [14] appears to be the best predictor of treatment success. For beta-lactams, the time above the minimal inhibitory concentration (MIC) correlates best with bacteriological response [2]. For some antibiotics, such as tetracycline, antibacterial action depends on both TC>MIC and AUC [10]. Each of these three measures of exposure (AUC, Cmax and TC>MIC) would be optimized by employing different dosing strategies, for example by using large intermittent doses to increase Cmax or by employing extended release formulations to increase TC>MIC [15].
A clear mechanistic understanding of antibiotic pharmacodynamics has not yet been achieved, and this lack of knowledge is a major obstacle for the design of rational treatment regimens. Treatment strategies for bacterial infections (e.g. dose levels, dosing frequency, and duration of therapy) are usually developed based on pharmacodynamic and pharmacokinetic data collected through expensive in vitro and in vivo studies [9, 16–18]. Specifically, the question of which pharmacokinetic driver governs antibiotic efficacy has to be determined experimentally with hollow-fiber systems or animal models [11, 19, 20]. This experimental information in turn can be incorporated into mathematical models [21], but to our knowledge there is no mathematical model that can guide these experiments.
Thus, the development of models that can inform optimal dosing strategies from data collected in early phases of antibiotic development could speed the drug development process and help to identify promising compounds that should be prioritized [22]. Here, we extend a modeling framework [23] that integrates bacterial population biology with the intracellular reaction kinetics of antibiotic-target binding to investigate how the kinetics of drug-target binding affect bacterial response to fluctuating antibiotic concentrations. We find that the physicochemical characteristics of drug action predict differences in antibiotic pharmacodynamics at fluctuating concentrations and correlate well with observed data.
Using three models that incorporate complexity and realism in a stepwise fashion (Fig 1), we consider how reaction kinetics govern the expected bacterial responses to antibiotics. First, we use a simple model that considers only drug-target binding to explore the general principles of antibiotic-target reaction kinetics. Then, we use more complex models to simulate the action of two specific antibiotics, ampicillin and tetracycline, under a range of different dosing strategies. We assess which physicochemical characteristics of these two drugs explain their distinct pharmacodynamic behavior and evaluate how an understanding of these physicochemical characteristics can inform more effective dosing regimens.
In this section, we will employ Model 1, which allows us to focus exclusively on the kinetics of antibiotic-target binding A + T ⇌ AT (Model 1, Eqs (1–5)). Using this framework, we first explore the relationship of the MIC with physicochemical parameters. We then investigate the factors that may cause a delay in the bacterial response to antibiotics after initial exposure and factors that may extend these responses after antibiotics are withdrawn.
Recommended antibiotic dosage varies widely depending on the employed antibiotic and the targeted pathogen. It is therefore difficult to compare antibiotic action in terms of absolute concentrations. Typically, all measures of antibiotic efficacy are defined relative to the MIC of the specific bacteria/drug pair (Cmax/MIC, AUC/MIC and TC>MIC) to circumvent this problem. To be able to use a modeling framework based on physicochemical characteristics of drug action, it is therefore useful to define the MIC based on physicochemical properties [23]. In the simplest case, when we assume a constant antibiotic concentration in this framework, the MIC depends on two parameters: the drug target affinity (KD) and the threshold of bound target (fc) at which the net growth of a bacterial population is zero (Eq (3)). Fig 2A illustrates the expected MIC according to Eq (3) which depends on drug target affinity KD and the critical threshold fc. The absolute concentration of antibiotic at MIC rises with the threshold occupied target required for bacterial suppression (fc). Given any threshold level of target occupancy, drugs with a higher binding affinity (lower KD) will require smaller concentrations to prevent bacterial growth (lower MIC).
Classical models of antibiotic pharmacodynamics typically assume that the antibiotic concentration at any time point determines the net bacterial growth rate at that same time point. This assumes both that the antibiotic acts instantaneously and that previous antibiotic exposure has no continuing influence on bacterial growth. In reality, however, there is typically a delay between initial exposure and antibiotic effect and there may also be post-antibiotic period in which bacterial growth remains suppressed even after the antibiotic is removed from the extracellular space. Here we use our modeling framework to understand how the onset and end of antibiotic action are affected by the physicochemical properties of drug-target binding. We use the reaction kinetics of drug target binding (Eq (4)) to show the dynamics when the antibiotic is applied at a concentration slightly above MIC (1.01 x MIC). We use this concentration for illustration purposes because at this concentration, the critical fraction of binding fc is reached in finite time, but never substantially exceeds this threshold. (The influence of higher concentrations is explored in Fig 3.)
The different scenarios in Fig 2B illustrate the time course of drug-target binding at the same concentration relative to the MIC, but different absolute antibiotic concentrations. From the limited number of studies in which antibiotic-target dissociation rates have been directly measured, we assume that these rates range between 10−3/s and 10−4/s [24, 25]. Slower turnover of drug-target binding (i.e. a longer half-life of the drug-target complex) is associated with a delayed onset of action (compare Fig 2B dotted to solid lines).
Surprisingly, we find that the system approaches equilibrium more quickly when fc is higher. This effect can be explained as follows: The absolute antibiotic concentration at MIC rises sharply with the threshold fc, and can map to very different absolute drug concentrations (see Fig 2A). Initially, only the forward reaction is relevant when a negligible amount of target is bound and proceeds with the rate kf[A][T0] (i.e. as the product of forward reaction rate, antibiotic concentration and target molecule concentration). Under conditions where the antibiotic concentration is held constant, the equilibrium fraction of bound target [A] is given by [AT]eq=[A][A]+KD and asymptotically approaches 1. Therefore, the velocity of the reaction increases more quickly than the fraction of bound target at equilibrium. This produces a paradoxical finding: when dosing antibiotics at the same levels relative to their respective MIC, those that require a high threshold of bound target to be effective are expected to act more quickly (Fig 2B).
The delay until an antibiotic is effective depends on many physiological and biochemical factors. Since this model focuses on the reaction kinetics alone (ignoring diffusion barriers and concentration gradients), Model 1 provides a lower bound for the expected delay until onset of antibiotic action. Even here, for reasonable parameter settings, we find that even this delay can extend for several hours. One potential approach for speeding antibiotic-target binding and reducing delay to onset of action is to increase antibiotic exposure through higher dosing. Lower thresholds and slower turnover are associated with delays until antibiotic action; these effects can be overcome by increasing the drug concentration (Fig 3). The light blue solid and dotted vertical lines in Fig 2B indicate when the fast and slow reactions reach the fc, i.e. the time of onset of the antibiotic action (tonset), and can be compared to the blue and green lines in Fig 3B at 1.01MIC. When the antibiotic-target reaction equilibrates slowly, a high dose of antibiotic is especially beneficial and minimizes the opportunity for additional bacterial replication events prior to onset of antibiotic action.
Bacterial growth often remains suppressed after the antibiotic concentration drops below the MIC (i.e. the post-antibiotic effect). This effect occurs because drug-target complex dissociation is not instantaneous. Therefore, high drug concentrations that saturate the target beyond the threshold required for antibiotic action fc may have additional benefits if they extend bacterial suppression beyond the time that the antibiotic concentration exceeds the MIC.
We use our model to identify the conditions in which high antibiotic concentrations are expected to prolong antibiotic action. For simplicity, for these simulations we assume that at the time of antibiotic withdrawal, 99.9% of the target is bound and that the antibiotic concentration both inside and outside of the bacterial cell immediately drops to zero. Under these assumptions, Eq (1) can be simplified and the unbinding of the antibiotic corresponds to a simple exponential decay. Fig 4 illustrates the expected dissociation of the drug-target complex for antibiotics with different half-lives tbound. When the threshold required for antibiotic action fc is very high, the antibiotic stops working very rapidly and the length of the post-antibiotic effect is brief and relatively insensitive to the half-life of the drug-target complex. Conversely, when there is both a low threshold and a slow turnover time of drug-target binding, the post-antibiotic period may last for several hours.
Next, we investigate the dynamics of drug target binding under different dosing regimens. We use ampicillin as an example, because a large body of literature describes the time-dependent action of beta-lactam antibiotics, both in experimental models as well as in patients [26–30]. In addition, the reaction kinetics of drug-target binding are relatively well established. To investigate the generality of our findings, we then simulate the reaction kinetics of isoniazid, a prodrug that accumulates in the bacterial cell.
Beta-lactams acetylate penicillin-binding proteins (PBPs, the target molecules), and thereby inhibit cell wall synthesis. The acetylation of PBPs consumes beta-lactams, and therefore the drug-target reaction is not reversible. However, PBPs are constantly de-acetylated and the effects of the antibiotic are therefore reversible. The kinetics of PBP acetylation and de-acetylation as well as target occupancy at MIC have been determined experimentally (Table 1). In single cell experiments, ampicillin has no detectable sub-MIC activity (S1 Fig) so we assume that antibiotic is effective only while the fraction of bound antibiotic exceeds fc.
To explore whether TC>MIC, AUC or Cmax are the best predictors of antibiotic efficacy, we model three simplified dosing strategies: i) an idealized bolus injection where the drug concentration immediately reaches its peak and then declines exponentially, ii) a hypothetical pharmacokinetic curve where the antibiotic concentration is maintained just above the MIC (1.01x MIC) for the same length of time >MIC as in i) and then falls instantaneously to 0, and iii) a curve of similar shape to ii) that retains the same area under the curve as i) (see Fig 5A). In ii), the time above MIC is identical to i) but we eliminate the excess binding that occurs because of the initial high peak in i). In iii), the AUC is the same as in i) but instead of the high peak concentration, there is a significantly prolonged TC>MIC. In other words, all graphs in the middle column of Fig 5 have the same time > MIC as those in the left column, and all graphs in the right column have the same AUC as those in the left column.
First, we investigate whether our modeling framework can reproduce the time-dependent action of beta-lactams based on the known physicochemical characteristics of the drug and its target. Fig 5A shows numerical simulations of Eq (6) using experimentally determined parameters (see Table 1). Note that this equation includes diffusion across the bacterial cell envelope. However, penicillin-binding proteins are either located in the cell envelope or in case of gram-negatives in the periplasm, and we therefore assume here that the diffusion barrier to the target is negligible. We adapted Eq (6) to describe the consumption of beta-lactams during target acetylation by dropping the backward reaction term kr[AT] for the differential equation describing the intercellular antibiotic [A]i. For all three dosing strategies, antibiotic action starts immediately after the drug concentration rises above MIC and stops as soon as the drug concentration falls below MIC, i.e. increasing the AUC alone without increasing TC>MIC does not change antibiotic action substantially (compare Fig 5A left and right panel). This is in accordance with the observation that the efficacy of beta-lactams strongly depends on the time above MIC. Taken together, we can reproduce time-dependent action of beta-lactams solely based on reaction kinetics. The high threshold required for activity as well as the extracellular location of the target lead to a fast onset of drug action as the antibiotic concentration rises above the MIC and a nearly immediate end of antibiotic action as the antibiotic concentration drops below the MIC. To illustrate the different dynamics of bolus injections and constant dosing, we visualize time course of ampicillin action for a bolus injection in S1 Movie (Fig 5A, left panel) and for a constant concentration in S2 Movie (Fig 5A, middle panel).
Beta-lactams are the antibiotic class for which time-dependent action is most widely accepted, and we can reproduce their time-dependent action with our model. This suggests that physicochemical characteristics may be responsible for this behavior. For most other antibiotic classes, antibacterial efficacy is better correlated with AUC or Cmax [13, 14]. We hypothesized that alteration in specific physicochemical parameters could generate AUC and Cmax-dependent action. To investigate this hypothesis, we modified the parameters for ampicillin one at a time to determine whether, through such parameter modification, we could reproduce AUC and Cmax-dependent action. Because antibiotic treatments are usually given over several days and the time between individual doses is typically in the range of hours, we first investigated parameter changes that produce an equilibration time of several hours. For example, if the drug must diffuse across a cell envelope with a diffusion rate of p = 10−4 /s, this leads to a half-life of free intracellular drug of 1h 55min. Fig 5B shows a comparison of the same dosing strategies as used for the upper panel (Fig 5A) with this additional diffusion barrier. With such a strong diffusion barrier the antibiotic concentration inside the cell also remains above MIC after a bolus injection of 50x MIC for several hours because antibiotic molecules are retained within the cell. Consequently, the activity after such a bolus administration can be extended by several hours (Fig 5B). However, the diffusion barrier also delays the onset of antibiotic action. This delay is dependent on the antibiotic concentration, the left panel of Fig 5B shows a delay of 14 minutes while the right panel shows a delay of 13h. This is because the equilibration of intra- and extracellular concentration is slower when there are smaller differences between the concentrations outside and inside the bacterial cell. If the antibiotic dose is only slightly above MIC, >10h are required to reach the threshold for inhibition fc (Fig 5B, right panel). Thus, a dosing strategy with an equivalent time >MIC as the 50x MIC bolus administration will never achieve bacterial suppression, while a dose with an equivalent area under the curve is approximately 5.4 times more effective than the bolus injection (bolus injection: antibacterial activity from 14min to 11h9min, constant concentration with same MIC: activity from 13h to 71.5h). S2 Fig shows the dynamics when keeping Cmax constant but varying the drug half-life. Again, we would expect beta-lactam action to start immediately and end immediately when the drug falls below MIC; however, for an antibiotic with a substantial diffusion barrier, we would expect delays until the onset and cessation of drug action.
This behavior is not limited to slow equilibration rates due to diffusion barriers. Earlier, we identified two parameters that affect the onset and the end of antibiotic action: target occupancy at MIC (fc) and the half-life of the drug-target complex (tbound). A long exposure to ampicillin at MIC is expected to result in a target saturation that just reaches the critical threshold fc = 95.4% [38] at equilibrium. After withdrawal of the drug, the target saturation would immediately fall below this critical threshold and the antibiotic would no longer be active. The duration of antibiotic action might be extended with higher drug concentrations since this would produce higher target saturation and lead to a longer delay until the fraction bound target falls below the critical threshold fc. However, increasing the target saturation at the beginning from 95.4% to 99.9% is expected to extend the action of ampicillin only by about 9 minutes (0.999ln(fc)−kr). On the other hand, if the critical threshold fc was 10% instead of 95.4%, achieving a target saturation of 99.9% would extend the expected time of antibiotic action by over 6h. Similarly, if the rate of de-acetylation is decreased 10-fold (kr = 10−5), the expected duration of antibiotic action after achieving a target saturation of 99.9% is extended by over 1.5h. Changing both parameters to values that result in equilibration rates in the range of hours leads to the same qualitative behavior as when the equilibration rate is in the range of hours because of a diffusion barrier (S3 Fig). Thus, our model predicts that changing a single physiochemical parameter (equilibration times to the range of hours) has major impact on pharmacodynamics: instead of TC>MIC alone, the AUC becomes another predictor of antibiotic efficacy and both are needed to predict antibiotic action. S1 Movie and S2 Movie illustrate the time course of the action of a hypothetical antibiotic that has the same binding and de-acetylation rates as ampicillin, but where the antibiotic must cross a diffusion barrier with p = 10−4 /s and a threshold of fc = 10%. As in S1 Movie and S2 Movie, we compare a bolus injection (S3 Movie) and a concentration with the same time above MIC (S4 Movie).
We now examine the consequences of slow drug equilibration rates (i.e. in the range of days) on predicted antibiotic pharmacodynamics. We slow the diffusion rate across the bacterial cell envelope to p = 10−5, which corresponds to a half-life of 19h 15min. In this case, the antibiotic concentration inside the cell remains above MIC after a bolus injection of 50x MIC for a day. Exposure to the antibiotic at a concentration only slightly above MIC (1.01 x MIC) is insufficient to achieve the required amount of bound target, even when maintained for several days (Fig 5C). Thus, in situations where antibiotics are expected to equilibrate slowly, a high peak concentration is necessary to achieve antibiotic action and the Cmax is expected to be the best predictor of antibiotic action.
We further tested this finding by investigating the reaction kinetics of a drug with a very different mechanism of action: the antitubercular pro-drug isoniazid (INH). In this case, target binding occurs after drug activation to the adduct INH-NAD which depends on NAD content and oxygen saturation. Importantly, the majority of the active drug INH-NAD remains in the mycobacterial cell and is not able to cross the cell envelope [49, 50].
Because INH-NAD remains in the cell, the expected amount of bound target does not decline even when the external concentration of INH declines. We therefore interpret treatment success here as the required time to reach fc, i.e. the expected time after which an average bacterium is killed. Again, we can reproduce experimental and clinical findings that INH treatment efficacy is significantly correlated with both Cmax and AUC in univariate regressions (Fig 6). S1 Table gives an overview of all parameters combinations used in the simulations.
In a multivariate regression, only Cmax is significantly correlated with the time to reach the required threshold to kill bacteria (S4 Fig), although the best model according to the Akaike Information Criterion includes all three pharmacokinetic indices. Due to the prodrug-activation, it takes several hours to reach the threshold required for killing, and this delay can be reduced with high peak concentrations. Since the active drug, INH-NAD, is trapped inside the cell, the intracellular drug concentration does not decrease when the external concentration decreases. Therefore, it is not necessary to keep the external pro-drug concentration above MIC for the drug to be active.
The time above the MIC is expected to be a reasonable predictor of antibiotic action in situations where antibiotic concentrations below the MIC have little effect on bacteria. For example, cells exposed to 80% MIC ampicillin show no measurable defect in either growth or elongation rates, and all cells remain intact (S1 Fig). In contrast, translation inhibitors such as chloramphenicol and tetracycline do affect bacterial growth below MIC, and it has previously been shown a nearly complete suppression of growth at 80% MIC [23]. When fitting Eq (10) to data from single cells exposed to constant sub-MIC concentrations of antibiotics [23], it has previously been estimated that a high threshold of bound ribosomes must be met to interrupt all bacterial growth (fc = 98%) and that there is a low diffusion barrier (p = 1.2x 10−2/s). Experimental values from the literature suggest a short half-life of the drug-target complex (Table 1). Based on the values of these parameters, we expect that the TC>MIC should be the best predictor of tetracycline effects. However, experimental and clinical evidence suggests that both AUC and TC>MIC determine the efficacy for tetracycline [23]. Accordingly, we used Model 3 (Eq (10) populated with parameters for tetracycline), to investigate how sub-MIC activity affects antibiotic pharmacodynamics under different dosing strategies.
Fig 7A shows simplified pharmacokinetics of a tetracycline bolus injection with initial concentrations ranging from 0.1–5 x MIC. Fig 7B shows the effects of dosages above MIC on the bacterial growth rate. Given the low diffusion barrier, bacterial growth is completely suppressed as long as the antibiotic concentration is retained above MIC. As soon as the antibiotic concentration falls below MIC, bacterial growth immediately resumes and continues to increase in rate as the antibiotic is cleared. Thus, when considering pharmacokinetic measures that correlate with the complete suppression of bacterial growth, the time above MIC is the best predictor of antibiotic action. However, the model also suggests that sub-MIC concentrations may substantially affect the total expected bacterial load over 24h (Fig 7C).
Accordingly, our model predicts that TC>MIC may be an imperfect predictor of antibiotic action (at least as measured by its effect on the total bacterial burden over 24h) since sub-MIC exposure can impact expected bacterial burden (first 5 data points in Fig 8A, highlighted in blue). In contrast, for antibiotic dosages above MIC, TC>MIC correlates well with efficacy (last 5 data points in Fig 8A, highlighted in red), however, it should be noted that the overall effect is very small. Nevertheless, in a treated patient with low remaining bacterial burden the additional killing of few bacteria can make the difference between cure (i.e. extinction of bacterial population) or relapse. Our model is deterministic and therefore cannot capture extinction, however, depending on the initial population size a very small frequency of survivors that translates to less than one bacterium effectively means extinction. Over a wider range of antibiotic dosages in bolus injection, the area under the curve correlates more strongly with antibiotic effects because it is a measure that also reflects actions that occur below MIC (Fig 8B).
It is well known that certain pharmacokinetic measures (i.e. AUC, Cmax or TC>MIC) are better predictors of the pharmacodynamics of some antibiotics than of others, but we currently have limited quantitative understanding of the mechanisms that drive this phenomenon. In this paper, we extend a model that links chemical reaction kinetics to bacterial population biology [23] and suggest a potential mechanistic explanation for this phenomenon.
Based on this model, we suggest how physicochemical and biochemical characteristics of drug-target interaction may shape antibiotic dose response curves. Differences in characteristics between antibiotics offer a compelling explanation for the observation that different measures of drug exposure correlate best with antibacterial activity. Specifically, we identified four factors that govern patterns of drug effects: i) the half-life of the antibiotic-target complex, ii) the diffusion barrier between extracellular antibiotic and its target, iii) the threshold of bound target required to suppress bacterial growth (i.e. target molecule occupancy at MIC) and iv) drug effects when the antibiotic is present only at sub-MIC levels.
The first three factors, the half-life of drug-target complex, the diffusion barrier and the threshold required for bacterial suppression, all influence the time until the antibiotic starts and stops acting (i.e. the equilibration rate of the reaction). When the onset of action of an antibiotic is rapid, we expect that achieving drug concentrations just above MIC should be sufficient to trigger the antibacterial effect. If an antibiotic stops acting quickly, antibiotic effects should cease as soon as the concentration falls below MIC. In these circumstances, we expect that the time above MIC would be a good measure for antibiotic efficacy. We demonstrated that our model, when parameterized with relevant drug-target binding data from the literature, can reproduce such time-dependent pharmacodynamics of ampicillin. Beta-lactams are somewhat unique in that their targets are located outside the cytosol [54]. Therefore, there is negligible diffusion barrier between the antibiotic molecules surrounding a bacterial cell and their targets. Our model predicts that this leads to a fast onset and end of antibiotic action. Also, almost all target molecules are occupied at MIC [38], and we demonstrate here that this also should lead to a rapid onset and cessation of antibiotic activity. Time-dependent efficacy of beta-lactams is well established both experimentally and clinically. For example, it is recommended that beta-lactams are given as continuous infusion rather than bolus injections [27].
For most other antibiotic classes, antibacterial efficacy is correlated with AUC or Cmax [10, 14]. Many antibiotics have targets that are located in the cytosol (e,g. ribosomal-targeting antibiotics such as streptomycin or gyrase-targeting antibiotics such as ciprofloxacin). Also, unlike beta-lactams, many antibiotics will have effects before the majority of target molecules are bound. We therefore investigated whether our model can also reproduce concentration-dependent patterns of antibiotic action, in which antibiotic efficacy is best described by either Cmax or AUC.
Indeed, our model predicts that the TC>MIC is not highly correlated with treatment efficacy when the time until an antibiotic starts and stops being active (i.e. the equilibration time) is in the range of hours or longer. The delay until an antibiotic is effective depends on many physiological and biochemical factors. Here, we focus on the reaction kinetics alone, which provide a lower bound for the expected time to onset of antibiotic action. We note that even these lower-bound estimates may be as long as a few hours, potentially permitting several additional generations of bacterial replication. We would therefore suggest high doses, at least initially, for antibiotics that: 1) act at low thresholds of bound target; 2) diffuse only slowly through the cell envelope; or 3) have a slow turnover rate (i.e. a long half-life of drug-target binding). A similar argument can be made for the anti-tuberculosis drug isoniazid, which is a prodrug that is activated by bacterial cells. The activation rate of the drug alone is sufficient to explain the slow onset of action of the drug [23], and this delay can likely be reduced with higher antibiotic doses. Indeed, the efficacy of isoniazid has been linked to high peak doses [55], a finding we were able to reproduce here. Additionally, when equilibration rates are slow, higher dose of antibiotics can extend the action of the antibiotic beyond the time the antibiotic concentration outside the bacteria exceeds the MIC. Thus, high doses have the additional benefit of prolonging the post-antibiotic period for antibiotic-target pairs that equilibrate slowly. In isoniazid, this extension of drug action predicted by our model is especially pronounced, because the drug is trapped in the cell such that declining external drug concentrations have little effect. In principle, these delays in onset and end of action are a similar phenomenon to the concept of a “biophase lag” [56] although the underlying mechanisms are not the same.
To examine the conditions in which each of these pharmacokinetic metrics provides the best measure of drug effect, we compared a dosing strategy with a high peak concentration that facilitates rapid onset of an antibacterial effect with a dosing strategy that has an equivalent AUC, but a lower peak concentration and a substantially longer exposure time (Fig 5). If an antibiotic equilibrates slowly, the onset of antibiotic action at low doses is so delayed that the required fraction of bound target cannot be reached before the antibiotic falls below MIC in the low dose/long exposure strategy (Fig 5). Obviously, the exact parameter ranges in which this is the case depend on the definition of “long” (in our case, days). If equilibration is too slow compared to the relevant timeframe (for example due to the accumulation of activated isoniazid in the cell), we would expect that Cmax is a better predictor of antibacterial efficacy than the AUC. Whether the peak concentration (Cmax) or the total exposure (AUC) is the best predictor of antibiotic efficacy thus depends on both the observed timeframe and the equilibration rate.
In addition to the onset and end of antibiotic action, we found that the biological activity of the antibiotic at sub-MIC concentrations also determines which pharmacokinetic measure best predicts treatment efficacy. A similar argument has been made for rifampicin therapy in tuberculosis[57]. Some antibiotics such as ampicillin (S1 Fig) have very little effect below MIC. In contrast, some antibiotics like tetracycline have some sub-MIC activity. Clearly, the time above MIC alone cannot predict treatment success when sub-MIC concentrations partially suppress bacterial growth. Indeed, our model predicts that treatment efficacy with tetracycline depends both on TC>MIC and AUC which is in concordance with clinical and experimental studies [10].
Taken together, our mechanistic model can reproduce the pharmacodynamic characteristics of both ampicillin and tetracycline. It offers an intuitive explanation for differences in optimal dosing strategies between antibiotic classes. However, the parameters needed to inform even such a simple model have not yet been measured for many antibiotic/bacterial pairs. We note that most of the kinetic measurements for antibiotic-target binding were published decades ago [24, 34, 35]. To our knowledge, beta-lactams are the only antibiotic class for which target occupancy at MIC has been experimentally determined. Furthermore, the number of target molecules per cell and especially the concentration of free antibiotic at the target site are rarely known, despite being a focus of active research in tuberculosis [58–61]. We suggest that experiments to address these knowledge gaps should be prioritized as the results of these studies could inform new approaches for the rational dosing of antibiotics.
Identifying optimal antibiotic dosing strategies is challenging and in this paper we have addressed only a subset of the considerations that must be accounted for when determining treatment recommendations. For example, antibacterial efficacy and toxicity must be balanced and the frequency of dosing may affect adherence; these are important factors that should doubtless affect treatment recommendations. In addition, our simple models do not consider host immune responses to infection, which may further modify our expectations regarding treatment success [57, 62, 63]. Nevertheless, given the urgent need to preserve the efficacy of existing antibiotics and the need to develop new agents [64], we see a promising role for mechanistic models that can suggest the most promising dosing strategies based on the physicochemical and biochemical characteristics of drug-target interactions. Such novel pharmacodynamics models can also be integrated into more complex frameworks that include host responses and more sophisticated pharmacokinetics [57, 62, 63].
Our model is general and we believe it could be usefully adapted to improve dosing strategies for treatment of other diseases. For example, we note that the effects of the physiological fluctuations of drug concentration are also poorly understood in the treatment of cancer [65], HIV [66] and malaria [67] and similar questions arise regarding the effects of exposure to harmful substances in toxicology [68].
Previously, we have shown that models that consider drug-target binding kinetics can explain complex patterns of antibiotic action such as post-antibiotic effects, inoculum effects, and persistence [23]. The central assumption of these models is that bacterial replication decreases and/or bacterial killing increases with the fraction of bound target molecules. Here, we extend this approach using three different mathematical models that incorporate additional complexity and biological realism in a stepwise fashion (Fig 1B). In all these models we follow the entire bacterial biomass rather than single cells. For our purposes here and in contrast to previous work [23], we can simplify the model by assuming that there is negligible heterogeneity between single cells. Table 2 lists all parameters and variables of these models.
To build our understanding of the drug-target reaction kinetics as antibiotic concentrations fluctuate within a host, Model 1 focuses only on the drug-target binding that occurs after exposure and withdrawal of an antibiotic. For Model 1 we make the following simplifying assumptions (which are subsequently relaxed in Models 2 and 3):
The chemical reaction of antibiotics with their targets is described by the following equation: A+T ⇌ AT. The intracellular antibiotic molecules A react with target molecules T with a rate kf and form an antibiotic-target molecule complex. If the reaction is reversible, the complex dissociates with a rate kr, leading to a dynamic equilibrium.
The dynamics of this system are governed by the concentrations [A], [T], [AT] rather than the absolute number of molecules. We assume that the total concentration of target/cell [T0] is constant. In this case, the concentration of free target can be described as [T] = [T0] − [AT]. Assuming that cells are treatment-naïve, i.e. there are no bound target molecules at the beginning, the kinetics of antibiotic-target reaction can then be described by a single differential equation, which can be simplified if we assume the intracellular antibiotic concentration [A] is constant:
d[AT]dt=kf[A]([T0]−[AT])−kr[AT]
(1)
and solved as:
[AT](t)=kf[A][T0](1−e−(kr+kf[A])t)kr+kf[A]
(2)
At a certain point, the fraction of bound target reaches a critical threshold at which the net growth of the bacterial population is zero. In this framework, the MIC is characterized as the minimal antibiotic concentration at which this critical percentage of bound target, fc, is reached. Thus, the MIC is the antibiotic concentration at which the equilibrium fraction of bound antibiotic is exactly fc: i.e.
[AT]MIC[T0]=fc. After simplifying, this yields:
MIC=KDfc1−fc
(3)
with the affinity constant KD=krkf.
Expressing all antibiotic concentrations as fold-MIC (xMIC) and thereby replacing [A] with MICKDfc1−fc, Eq (2) can then be transformed:
[AT](t)=fcT0xMIC1−fc(1−xMIC)(1−e−kr(1−fc(1−xMIC))t1−fc)
(4)
The time to the onset of antibiotic action, i.e. the delay until the fraction of bound target first exceeds fc after antibiotic administration, can be expressed as:
tonset=fc−1kr(1+fc(xMIC−1))log(−(xMIC−1)(fc−1)xMIC)
(5)
We next extend Model 1 to allow for fluctuating antibiotic concentrations after bolus dosing and to account for diffusion across the bacterial cell envelope. (These extensions effectively relax the first two assumptions for Model 1).
Model 2 includes the following compartments: Ae, the number of extracellular antibiotic molecules, Ai, the number of intracellular antibiotic molecules, T, the number of free target molecules, and AT, the number of drug-target complexes. For bolus injections, the model is described by the following set of equations:
dAedt=−ln(2)tclAe−p(AeViVe−Ai)dAidt=p(AeViVe−Ai)−kfnAViAiT+krATdTdt=−kfnAViAiT+krATdATdt=kfnAViAiT−krAT
(6)
To model an alternative drug administration approach in which the antibiotic concentration is maintained at a constant level c and after a specified time (tend) is assumed to fall instantaneously to 0 (i.e. intravenous dosing), the extracellular antibiotic concentration is given by:
Ae={cfort<tend0fort≥tend
(7)
We express the antibiotic concentration as fold-MIC (xMIC) using Eq (3). The terms describing the chemical kinetics of drug-target reaction are equivalent to Eq (1). In addition, we describe the diffusion through the cell envelope with a permeability coefficient p depending on the concentration difference inside and outside of the bacterial cells and the clearance of the extracellular antibiotic; its half-life is tcl. In our simulations, drug binding and diffusion from extra- to intracellular space changes the dynamics of external drug concentrations negligibly, even though this may change at very high bacterial loads with a high number of targets per cell [69].
Here, we use the same equations and parameters as in Figure 7 in [23], extended by diffusion across the cell envelope and a decay term that describes the elimination of the drug from the blood after a bolus injection with t1/2. In the case of the prodrug isoniazid (INH), target binding occurs after drug activation to the adduct INH•NAD (equivalent to A before) which depends on NAD content and oxygen saturation. Here, we focus on INH binding to the enoyl reductase InhA, which is then present in its inactive form InhAi. Assuming NAD and target molecule concentration as well as oxygen saturation remain constant, the number of molecules in each compartment is described by the following set of equations:
dINHedt=−ln(2)tclINHe−p(INHeViVe−INHi)dINHidt=p(INHeViVe−INHi)−kNAD,O2INHidINH∙NADdt=kNAD,O2INHi−kfnAViINH∙NADInhA+krInhAidInhAidt=kfnAViINH∙NADInhA−krInhAi
(8)
This set of equations is based on (6) and we additionally model prodrug activation.
Finally, Model 3 expands on Model 2 by allowing the reproduction of target molecules that would occur as a result of bacterial replication and also allows for unspecific binding. (This extension relaxes assumption 3 and partially relaxes assumption 4 in the list provided above.) This model describes antibiotics that only suppress bacterial growth but do not increase bacterial killing (i.e. bacteriostatic agents). For bacteriostatic translation inhibitors such as tetracycline, the bacterial replication rate depends linearly on the fraction of free ribosomes [70]. We therefore assume that the bacterial growth rate r is proportional to ffree=[T][T]+[AT] above ff = 1- fc and that there is no growth when the fraction of free ribosomes falls below this critical threshold:
r(ffree)={0forffree<ffrnodrug11−ff(ffree-ff)forffree>ff
(9)
Here, we track bacterial cells B (scaled in number of cells per liter) that can reproduce until they reach a maximal carrying capacity K, the extracellular and intracellular number of antibiotics Ae and Ai, and the intracellular concentration of drug-target complexes AT and unspecifically bound antibiotic AU. The rates kf and kr describe specific binding and dissociation, the rates ku,f and ku,r describe the rates for unspecific binding and dissociation. Data indicate that the total number of ribosomes increases linearly with cell volume; this means that the intracellular concentration within a single cell between the time of its “birth” and the split into two daughter cells remains relatively constant [31]. We can therefore write the number of free target molecules as T = BT0 − AT with T0 describing the fixed number of total target molecules per cell. The growth of bacteria exposed to sub-MIC concentrations of a translation inhibitor can then be described by the following set of differential equations (note that we are again following molecules, not molar concentrations):
dBdt=r(B,T0,[AT],fc)B(1−BK)dAedt=−ln(2)tclAe−p(AeViVe−Ai)dAidt=p(AeViVe−Ai)−kfnAViAiT+krAT−ku,fAiB+ku,rAUdATdt=−kfnAViAiT+krATdAUdt=ku,fAiB−ku,rAU
(10)
Again, this equation is based on (6), in addition, we model bacterial population biology by following the total amount of bacteria B.
|
10.1371/journal.ppat.1000361 | Caspase-7 Activation by the Nlrc4/Ipaf Inflammasome Restricts Legionella pneumophila Infection | Legionella pneumophila (L. pneumophila), the causative agent of a severe form of pneumonia called Legionnaires' disease, replicates in human monocytes and macrophages. Most inbred mouse strains are restrictive to L. pneumophila infection except for the A/J, Nlrc4−/− (Ipaf−/−), and caspase-1−/− derived macrophages. Particularly, caspase-1 activation is detected during L. pneumophila infection of murine macrophages while absent in human cells. Recent in vitro experiments demonstrate that caspase-7 is cleaved by caspase-1. However, the biological role for caspase-7 activation downstream of caspase-1 is not known. Furthermore, whether this reaction is pertinent to the apoptosis or to the inflammation pathway or whether it mediates a yet unidentified effect is unclear. Using the intracellular pathogen L. pneumophila, we show that, upon infection of murine macrophages, caspase-7 was activated downstream of the Nlrc4 inflammasome and required caspase-1 activation. Such activation of caspase-7 was mediated by flagellin and required a functional Naip5. Remarkably, mice lacking caspase-7 and its macrophages allowed substantial L. pneumophila replication. Permissiveness of caspase-7−/− macrophages to the intracellular pathogen was due to defective delivery of the organism to the lysosome and to delayed cell death during early stages of infection. These results reveal a new mechanism for caspase-7 activation downstream of the Nlrc4 inflammasome and present a novel biological role for caspase-7 in host defense against an intracellular bacterium.
| Legionella pneumophila causes a severe form of pneumonia called Legionnaires' disease. In human macrophages, L. pneumophila establishes special vacuoles that do not fuse with the lysosome and grows intracellularly. However, in mouse macrophages, the bacteria are efficiently delivered to the lysosome for degradation. Importantly, caspase-1 is activated when L. pneumophila infects mouse macrophages, but not when it infects human cells. Caspase-1 activation promotes the fusion of the L. pneumophila vacuole with the lysosome and macrophage death. However, the caspase-1 substrate mediating such effects is unknown. Experiments performed in vitro demonstrate that caspase-7 is a substrate of caspase-1. Yet, it is not known if the reaction takes place within the macrophage, and it is unclear if it has any biological effect. In this study we show that, in mouse macrophages, caspase-7 is activated by L. pneumophila downstream of caspase-1 and requires the host receptors Nlrc4 and Naip5. Remarkably, caspase-7 activation during L. pneumophila infection restricts growth by promoting early macrophage death and efficient delivery of the organism to the lysosome. Consequently, L. pneumophila grows in the macrophages and the lungs of caspase-7−/− mice. Therefore, we demonstrate a novel caspase-7 activation pathway that contributes to the restriction of L. pneumophila infection.
| Caspases are a family of cysteine proteases expressed as inactive pro-enzymes that play a central role in most cell death pathways leading to apoptosis. However, growing evidence implicates caspases in non-apoptotic functions [1]–[4]. Eleven genes were found in the human genome to encode 11 human caspases, whereas 10 genes were found in the mouse genome to encode 10 murine caspases. The human caspase-4 and -5 are functional orthologs of mouse caspase-11 and -12. The remaining caspases which share same nomenclature in human and mouse are functional orthologs of each other [1]. On the basis of their biological functions, caspases can be classified into three groups: inflammatory caspases like caspase-1, -4, -5, -11 and -12, initiator caspases like caspase-2, -8, -9, and -10, and effector caspases like caspase-3, -6, -7 and -14 [2],[4]. Caspase-1 activation mediates the maturation of the proinflammatory cytokines interleukin-1 beta (IL-1β), IL-18 and possibly IL-33 [5],[6]. Activation of caspase-1 is mediated within the inflammasome complex that is assembled when pathogen-associated molecular patterns (PAMPs) are sensed in the cytosol by special host receptors. These cytosolic receptors belong to the nucleotide binding oligomerization domain-leucine rich repeat proteins (NOD-like-receptors or CATERPILLAR family of proteins) [7]–[12]. A variety of pathogens such as Shigella, Francisella, Salmonella, Listeria, Pseudomonas, Escherichia coli and Legionella activate caspase-1, engaging different NOD-like-receptors [13]–[18].
L. pneumophila is an intracellular bacterium and the causative agent of Legionnaires' pneumonia [19]. The ability of L. pneumophila to cause pneumonia is dependent on its tendency to invade and multiply within human macrophages [20]–[23]. Once phagocytized, the bacteria reside in specialized vacuoles [20]–[26]. The L. pneumophila-containing phagosome does not fuse with the lysosome and instead acquires vesicles from the endoplasmic reticulum (ER) [20]–[26]. Within this vacuole, L. pneumophila multiply exponentially [27]. In contrast, macrophages from most mouse strains are restrictive to L. pneumophila infection. Within mice cells, L. pneumophila flagellin is sensed by Nlrc4 leading to the activation of caspase-1 [7], [28]–[30], whereas in human macrophages, caspase-1 is not activated in response to L. pneumophila. Caspase-1 activation in mouse macrophages is accompanied with L. pneumophila restriction due to the delivery of organisms to the lysosome for degradation and early macrophage death [28],[31]. Furthermore, pharmacological inhibition of caspase-1 in wild-type macrophages allows more L. pneumophila replication [28],[31],[32]. Accordingly, mouse macrophages that do not activate caspase-1 in response to L. pneumophila such as Nlrc4−/− and caspase-1−/− cells are permissive to infection [28],[31]. A/J mice and their derived macrophages are also permissive to L. pneumophila intracellular replication despite caspase-1 activation [28], [29], [33]–[36]. The downstream mechanism responsible for the permissiveness of macrophages lacking Nlrc4, caspase-1 or functional Naip5 is still unclear.
In vitro experiments revealed that caspase-1 directly processed procaspase-3 and -7 [37],[38]. Nevertheless, the biological role of this activation is unknown. Furthermore, whether this reaction is pertinent to the apoptosis or to the inflammation pathway or whether it mediates a yet unidentified effect is not clear.
Here we demonstrate that caspase-7, but not caspase-3, was activated in restrictive wild-type mouse macrophages by L. pneumophila. Caspase-7 activation by low multiplicity of L. pneumophila infection was dependent on Nlrc4, caspase-1 and functional Naip5. Such activation was accompanied by enhanced fusion of the L. pneumophila-containing phagosome with the lysosome and early death of infected cells resulting in restriction of infection in wild-type macrophages. The activation of caspases-8 and -9 which are involved in caspase-7 activation in response to apoptotic signals was not necessary for L. pneumophila-mediated caspase-7 activation. In contrast to caspase-7, caspase-3 was not activated by L. pneumophila in wild-type macrophages and its absence did not affect the activation of caspase-7 or the intracellular fate of the organism. The effect of caspase-7 activation on L. pneumophila growth was independently of IL-1β and IL-18. Remarkably, caspase-7 activation also controlled the growth of the pathogen within the murine lungs in vivo.
Therefore, our data identify a previously uncharacterized signaling pathway for caspase-7 activation through Nlrc4, caspase-1 and Naip5. We also demonstrate a new role for caspase-7 in host defense against an intracellular bacterium. These findings may be valuable in the design of compounds that could restrict the growth of not only L. pneumophila but also other organisms that tend to avoid lysosomal fusion.
L. pneumophila infection leads to caspase-1 activation in macrophages. The activation of caspase-1 is accompanied by restriction of L. pneumophila growth in macrophages and in mice [28],[31]. However, the downstream signaling pathway involved in the control of L. pneumophila growth is not known. In vitro studies suggested that caspase-1 can cleave caspase-7 and caspase-3 [37],[38]. However, it is not known if this reaction takes place in vivo during L. pneumophila infection. Therefore, we investigated whether caspase-7 and caspase-3 are cleaved within wild-type C57BL/6 macrophages upon L. pneumophila infection. Assessment of different multiplicity if infection (MOI) revealed that caspase-7 was activated at MOI ranging from 0.5 to 20 within 2 hours of infection. Caspase-1 activation was determined by the detection of the corresponding mature subunits in cell extracts in western blots using specific anti-caspase-1 antibodies (Figure 1A and 1B and Figure S1A). Infection of wild-type murine macrophages was accompanied by caspase-7 activation only in the presence of a functional Dot/Icm type IV secretion system, a bacterial apparatus that injects bacterial products into the host cytosol (Figure 1A and 1C). The bacteria induced proteolytic cleavage of pro-caspase-7 within 30 minutes even at low MOI (Figure 1B). Therefore, L. pneumophila activates caspase-7 in macrophages in a Dot/Icm-dependent manner.
Since L. pneumophila flagellin triggers caspase-1 activation via the NOD-like receptor Nlrc4 [7],[28], we tested whether a L. pneumophila mutant lacking flagellin (Fla) induced caspase-7 activation. We infected wild-type macrophages with L. pneumophila or with the Fla mutant lacking flagellin and examined the activation of caspase-7 (Figure 1C). To ensure equal internalization of wild-type and flagellin-deficient L. pneumophila, infections were followed by mild centrifugation to enhance contact between macrophages and mutant bacteria. Under these conditions, we recovered equal numbers of bacteria at 1 hour post-infection (data not shown). Unlike wild-type bacteria, the L. pneumophila Fla mutant failed to induce caspase-7 activation even at high MOI (Figure 1C, Figure S1B, and Figure S1C). Since flagellin is sensed by the host receptor Nlrc4, we tested if caspase-7 activation by L. pneumophila requires Nlrc4. Caspase-7 activation by wild-type L. pneumophila was abrogated in macrophages lacking Nlrc4 (Figure 1C). To verify the role of flagellin in caspase-7 activation, purified bacterial molecules were delivered intracellularly using streptolysin O (SLO), a protein that allows delivery of exogenous molecules into the cytosol of living cells [28],[39]. Bacterial flagellin, bacterial lipoproteins (LP), RNA, DNA, or lipopolysaccharide (LPS) activated caspase-7 when delivered to the cytosol using SLO (Figure 1C and 1D). However, flagellin required Nlrc4 for caspase-7 activation (Figure 1C and 1D). The low concentration and short time of SLO treatment used to deliver flagellin to the cytosol did not activate caspase-7 in the absence of bacterial ligands (Figure 1C). Therefore, caspase-7 activation by L. pneumophila is mediated by the bacterial flagellin through the host sensor Nlrc4.
To rule out the contribution of Toll-Like Receptor (TLR) signaling in caspase-7 activation by L. pneumophila, macrophages lacking the TLR adaptor molecules MyD88 or TRIF were infected and examined for activation of caspase-7. The lack of MyD88 or TRIF did not prevent caspase-7 activation by L. pneumophila (Figure S2A and Figure S2B). These results indicate that the sensing of flagellin through Nlrc4 mediates caspase-7 activation by L. pneumophila independently of TLRs.
Nlrc4 is indispensable for caspase-1 and caspase-7 activation by L. pneumophila. However, it is not clear if caspase-1 activation is upstream of caspase-7 as suggested by the in vitro experiments [38]. To assess the role of caspase-1 in caspase-7 activation during early stages of infection, caspase-1−/− macrophages were infected with low MOI of L. pneumophila. Particularly, caspase-7 activation by L. pneumophila was abolished in the absence of caspase-1 (Figure 2A), whereas caspase-1 activation by L. pneumophila did not require caspase-7 (Figure 2B). This result suggests that caspase-7 is downstream of the caspase-1 activation pathway. However, at high MOI caspase-7 was activated independently of caspase-1 (Figure S1B).
To verify if the enzymatic activity of caspase-1 contributes to the induction of caspase-7 activation by low MOI of L. pneumophila, we inhibited caspase-1 activity with the caspase-1 inhibitor Z-YVAD-FMK and examined caspase-7 activation upon L. pneumophila infection. Pharmacological inhibition of caspase-1 abolished caspase-7 activation by L. pneumophila (Figure S3A). Taken together, our data show that caspase-1 enzymatic activity is necessary for caspase-7 activation by L. pneumophila in macrophages.
Caspase-7 and caspase-3 are activated during apoptosis via the intrinsic and extrinsic pathways through caspase-8 and -9 [3]. First, to understand the role of caspase-3 in L. pneumophila-mediated caspase-7 activation, caspase-3−/− macrophages were infected with L. pneumophila and examined for the cleavage of caspase-1 and caspase-7. Our data demonstrate that the absence of caspase-3 had no effect on pathogen-mediated activation of caspase-1 (Figure 2B) or of caspase-7 (Figure 2C). These data support the fact that caspase-3 was not activated by L. pneumophila in wild-type C57BL/6 macrophages although activated in A/J-derived macrophages (Figure 2D). Second, to test the role of the initiator caspase-8 and -9 in caspase-7 activation by L. pneumophila infection, the pathogen was added to macrophages pretreated with the caspase-8 inhibitor (Z-IETD-FMK), or caspase-9 inhibitor (Z-LEHD-FMK), then cell lysates were examined for caspase-7 activation. In contrast to caspase-1 inhibition, the pharmacological inhibition of caspase-8 or -9 did not compromise caspase-7 activation by L. pneumophila (Figure 3A). However, inhibition of caspase-8 and -9 activities compromised caspase-7 activation induced by the apoptosis inducer doxorubicin (Figure S3B). These results indicate that caspase-7 activation by L. pneumophila is independent of caspase-8 or -9.
With the exception of the A/J mouse, most mouse strains are resistant to L. pneumophila [20],[36],[40]. The permissiveness of the A/J mouse is attributed to a polymorphism in the gene encoding the neuronal apoptosis inhibitory protein (Naip5) [34],[35],[41]. The susceptibility of A/J-derived macrophages to L. pneumophila is independent of caspase-1 activation since the levels of activation in the presence of wild-type and A/J-derived Naip5 are comparable [28],[29],[33]. Given that caspase-7 activation is mediated downstream of caspase-1 in wild-type macrophages, we tested if Naip5 plays a role in caspase-7 activation by L. pneumophila. Caspase-7 activation by L. pneumophila was compromised in macrophages derived from A/J-derived macrophages (Naip5AJ) (Figure 2E). Similar results were obtained using transgenic C57BL/6 mice harboring the mutant A/J-derived Naip5 (data not shown) [42]. These findings suggest that L. pneumophila–mediated caspase-7 activation downstream of the caspase-1 inflammasome requires Naip5.
Caspase-1 activation restricts L. pneumophila intracellular survival and replication [28],[31]. However, the downstream effectors responsible for the control of L. pneumophila growth are still unknown. Given that caspase-7 activation by L. pneumophila is dependent on caspase-1 activation, we investigated if caspase-7 controls L. pneumophila replication. Wild-type, caspase-3−/−, caspase-7−/− and caspase-1−/− macrophages were infected with L. pneumophila and intracellular bacterial replication was evaluated by quantifying the colony forming units (CFUs) throughout 72 hours of infection and by microscopy after 24 hours. Remarkably, macrophages lacking caspase-7 allowed substantial L. pneumophila replication when compared to wild-type cells (Figure 3A and 3B). The number of CFUs recovered from caspase-7−/− macrophages were comparable to the number recovered from their caspase-1−/− counterparts (Figure 3A and 3B). Peritoneal macrophages from caspase-7−/− mice were also highly permissive for L. pneumophila replication (data not shown).
To confirm the role of caspase-7 in L. pneumophila restriction, caspase-7−/− macrophages were complemented with caspase-7 plasmid and examined for the correlation between caspase-7 expression and bacterial replication. The delivery of a caspase-7 plasmid to primary caspase-7−/− macrophages was performed using transfection (Superfect) or nucleofection (Amaxa). Although caspase-7 was expressed using both techniques, nucleofection was detrimental to the cells (Figure S4A and data not shown). Expression of caspase-7 using Superfect was sufficient to restore the ability of murine macrophages to restrict L. pneumophila growth (Figure S4A and Figure S4B). The expression of the control plasmid carrying the red fluorescent protein (RFP) or green florescent protein (GFP) by the same techniques did not alter the permissiveness of the caspase-7−/− macrophages to L. pneumophila (Figure S4B and data not shown).
Since caspase-1 activity was necessary for L. pneumophila-mediated caspase-7 activation (Figure S3A), we tested if inhibition of caspase-1 activity permitted more bacterial replication. Distinctly, the inhibition of caspase-1 activity by Z-YVAD-FMK allowed more bacterial growth in wild-type B6 macrophages (Figure 3C). Unlike caspase-1 inhibition, the inhibition of caspase-8, -9 or -3 did not allow L. pneumophila replication in B6 macrophages (Figure 3C). Therefore, caspase-1 enzymatic activity is required for L. pneumophila growth restriction in macrophages. Unlike caspase-7, and despite the suggested similarity in small substrate specificity, the absence of caspase-3 had no effect on L. pneumophila growth in macrophages (Figure 3A).
Since Legionnaires' disease is caused by the replication of L. pneumophila in alveolar macrophages, we investigated if caspase-7 regulates bacterial growth within the lungs of mice in vivo. Wild-type and caspase-7−/− mice were infected intratracheally with 1×106 CFU of L. pneumophila and the number of bacteria in the lungs was determined at 96 hours post-infection (Figure 3D). The lungs of caspase-7−/− mice supported substantially more L. pneumophila CFUs than their wild-type counterparts (Figure 3D), although initial bacterial counts in the lungs were identical (Figure 3E). Therefore, these results indicate that caspase-7 restricts L. pneumophila replication both in vitro and in vivo.
In wild-type macrophages, a fraction of phagocytized L. pneumophila are contained inside phagosomes that rapidly fuse with lysosomes [22],[43],[44]. To determine the mechanism by which caspase-7 controls L. pneumophila growth in macrophages, we observed the trafficking of the organism intracellularly. First, we examined the rate of acquisition of the lysosomal-associated membrane protein-1 (LAMP-1) by phagosomes harboring L. pneumophila. Whereas, in wild-type macrophages, greater than 60% of the phagosomes containing the pathogen co-localized with LAMP-1 within 30 min post-infection, less than 30% of the L. pneumophila containing vacuoles in caspase-7−/−, caspase-1−/− and A/J-derived macrophages acquired LAMP-1 (Figure 4A and 4B).
Next, to test if caspase-7−/− macrophages contained a general defective phagosome-lysosome fusion function, we examined the ability of the non-pathogenic type IV secretion mutant (Dot) to traffic to the lysosome. In contrast to pathogenic wild-type L. pneumophila, caspase-7−/− macrophages efficiently delivered the Dot mutants to LAMP-1 labelled compartments within 2 hours of internalization (Figure S5A). Then, in order to assess the role of caspase-7 in mediating L. pneumophila degradation [28],[45], macrophages were infected with L. pneumophila, and the intact rod-shaped and degraded distorted-shaped bacteria were differentiated by labelling with anti-L. pneumophila antibody as previously described (Figure S5B). In wild-type macrophages, by 2 hours post-infection, about 40% of internalized bacteria lost their rod-shaped contour and were degraded into multiple small rounded particles compared to only 10 % in caspase-7−/− macrophages (Figure S5B).
Given that the L. pneumophila-containing phagosome in permissive cells acquires ER-derived vesicles [20],[22],[26], we examined the recruitment of calreticulin, an ER protein, to phagosomes harboring L. pneumophila. By 4 hours post-infection, around 35% of the bacteria co-localized with calreticulin in caspase-7−/− macrophages, whereas less than 5% bacteria associated with the ER marker in B6 macrophages (Figure 4C and 4D). The trafficking of L. pneumophila in caspase-7−/− macrophages was comparable to that in the caspase-1−/− and A/J-derived counterparts (Figure 4A and 4C). Therefore, caspase-7 activation promotes the fusion of the L. pneumophila-containing phagosome with the lysosome mediating the degradation of the pathogen.
Caspase-7 activation by L. pneumophila is regulated by caspase-1 which also regulates the maturation of the pro-inflammatory cytokines IL-1β and IL-18 [2],[5]. Our data show that caspase-7 activation is mediated by Nlrc4 and caspase-1 (Figure 1C and Figure 2A and 2E). To examine the possibility that caspase-7 is downstream of IL-1β and IL-18, we infected wild-type and IL-1β/IL-18 double knockout-derived macrophages with L. pneumophila and examined the activation of caspase-7. The absence of IL-1β and IL-18 did not prevent caspase-7 activation by L. pneumophila infection (Figure S6A). Next, to test if IL-1β and IL-18 control L. pneumophila growth, macrophages lacking both cytokines were infected and bacterial growth was quantified throughout 72 hours. Cells lacking IL-1β and IL-18 did not allow L. pneumophila replication intracellularly when compared with their caspase-7−/− and A/J-derived counterparts (Figure S6B). To further understand the role of IL-1β and IL-18 in infection, recombinant mouse IL-1β (rIL-1β) or IL-18 (rIL-18) was added to B6 and caspase-7−/− macrophages upon L. pneumophila infection. Exogenous cytokines had no effect on the growth of the pathogen whether caspase-7 was present or not (Figure S6C). Next, to examine the role of IL-1β and IL-18 receptors during L. pneumophila infection, corresponding antibodies to IL-1 receptor and IL-18 receptor were added during infection. The intracellular growth of L. pneumophila in macrophages in the presence of the antibodies was identical to that of untreated cells (Figure S6C). Our data demonstrate that despite being downstream of caspase-1, IL-1β and IL-18 do not play a role in caspase-7 activation by L. pneumophila or restriction of macrophage infection.
To determine if caspase-7 is upstream of IL-1β and IL-18 and hence controls their activation, we examined IL-1β and IL-18 release in culture supernatants of wild-type and caspase-7−/− macrophages in response to different organisms. IL-1β was released by L. pneumophila in the presence or absence of caspase-7 after 24 hours of infection. However, IL-18 was barely detected in both macrophage cell types in response to L. pneumophila (data not shown). Similarly, within 24 hrs of infection, IL-1β was released by wild-type and caspase-7−/− macrophages in response to Salmonella typhimurium (Salmonella) (Figure S7A and Figure S7B). These results demonstrate that caspase-7 does not regulate the caspase-1–dependent inflammatory substrates IL-1β or IL-18.
Our data show that at physiological stages of infection, L. pneumophila leads to caspase-7 activation in a caspase-1–dependent manner (Figure S1). Nevertheless, when high numbers of organisms invade the macrophage, caspase-7 is induced independent of caspase-1.
It has been shown that macrophage cell death contributes to the restriction of L. pneumophila infection [44],[46]. Therefore, we examined the role of caspase-7 in induction of cell death during L. pneumophila infection at different MOIs and durations of infection. First we measured LDH release in the overall population of macrophages infected with low MOI for 24 hrs. Minimal cell death was detected in response to wild-type organism in infected macrophages whether they expressed caspase-7 or not (Figure 5A).
To better understand the role of caspase-7 in L. pneumophila-mediated cell death, we measured macrophage apoptosis and necrosis in the overall population of macrophages by determining the cytoplasmic (apoptosis) and extracellular (necrosis) histone-associated-DNA-fragments during low and high MOI. Low MOI of L. pneumophila infection lead to minimal apoptosis after 24 hrs of infection in WT macrophages and in those lacking caspase-7, -1, -3, or functional Naip5 (Figure 5B). However, at high MOI, L. pneumophila induced significant apoptosis within 2 hrs of infection in all macrophages except those lacking caspase-1 (Figure 5C). Indeed, after 24 hrs of high MOI, all macrophages were apoptotic irrespective of their type (data not shown).
At low MOI, around 20% of macrophages were necrotic after 24 hrs of infection (Figure 5D). The high bacterial MOI did not dramatically increase the number of necrotic cells within 2 hrs of infection (Figure 5E).
To follow the role of apoptosis in L. pneumophila infection, apoptosis of individual cells after labeling of DNA strand breaks (TUNEL) was quantified by fluorescence microscopy. We scored 100 infected cells and quantified how many of those were TUNEL positive. Within 2 hrs of infection at low MOI, no more than 2% of infected macrophages lacking caspase-1, -7 or functional Naip5 were TUNEL positive whereas 20% infected wild-type macrophages were apoptotic (Figure 5F). After 24 hrs of infection, the majority of infected wild-type, caspase-7−/− and A/J-derived macrophages were apoptotic while infected caspase-1−/− macrophages did not show signs of apoptosis even when harboring several bacteria (Figure 5F). Remarkably, after the 24 hrs infection, many wild-type B6 macrophages that did not seem to harbor bacteria also underwent apoptosis (data not shown). Therefore, infected wild-type macrophages respond to L. pneumophila infection by early apoptosis which render them unsuitable for optimal bacterial replication.
Then, to investigate the role of apoptosis in vivo, we infected wild-type, caspase-1−/−, caspase-7−/− and A/J-derived macrophages intratracheally for three days. Then, harvested infected lung sections were stained with TUNEL to detect apoptotic cells. The degree of apoptosis observed in lung tissues was comparable among different infected lungs (Figure S8). These results suggest that apoptosis may not play a major role in L. pneumophila infection in vivo especially at latter stages of infection.
Human monocytes are permissive to L. pneumophila [20],[47]. The failure of human cells to restrict L. pneumophila infection is still under extensive studies. Remarkably, Caspase-1 activation is not detected during L. pneumophila infection of human cells [48]. To investigate the role of caspase-7 in human cells, we examined caspase-1 and capsase-7 activation in fresh human monocytes infected with L. pneumophila or with Salmonella. Caspase-1 was strongly activated in response to Salmonella but not in response to L. pneumophila. Similarly, caspase-7 was activated during Salmonella infection but not during L. pneumophila infection (Figure S9A and Figure S9B). Therefore, caspase-1 and caspase-7 are not activated in human monocytes upon L. pneumophila infection.
Taken together, our data show that the lack of caspase-7 activation is associated with permissiveness to L. pneumophila infection in mice and in humans.
The ability of L. pneumophila to survive in human cells presents a challenge to host defense. One common strategy for the host to deal with intracellular infection is to eliminate infected cells by caspase-mediated apoptosis. However, emerging reports demonstrate new functionally distinct roles for executioner caspases independent of cell death [1],[49]. Another strategy for eliminating intracellular bacteria is to deliver them to the lysosome for degradation. However, several pathogens have developed ways to deter such fate [43], [50]–[52].
Here we reveal a novel role for caspase-7 in host defense against L. pneumophila. We have identified a new activation pathway for caspase-7 that requires Nlrc4 and caspase-1, independent of the classical apoptosis pathway employing caspase-8 and -9. We also show that Naip5 contributes to caspase-7 activation downstream of caspase-1 during physiological levels of infection.
Caspase-7 activation restricts L. pneumophila infection in vitro and in vivo. This restriction is lost in caspase-7−/− macrophages, but restored after expression of caspase-7 via transfection (Figure S4). Taken together, the role of caspase-7 in restriction of L. pneumophila infection is confirmed, and the possibility that other molecules that may cause permissiveness to L. pneumophila such as Nramp-1 or Naip5 are responsible is ruled out [34],[40].
To recognize the role of caspase-7 during different stages of infection and different MOI, we pursued macrophage infections with low MOI (0.5) or high MOI (20). Under low MOI, caspase-7 activation by L. pneumophila required Nlrc4 and caspase-1 (Figures 1 and 2). It is possible that caspase-1 is required for the proper assembly of the Nlrc4-inflammasome while the cleavage of caspase-7 is mediated by another unidentified molecule. This possibility was deemed unlikely since the pharmacological inhibition of caspase-1 activity abolished caspase-7 activation in response to L. pneumophila (Figure S3) and because caspase-1 directly cleaves pro-caspase-7 in vitro [38].
In addition to flagellin, bacterial ligands not sensed by Nlrc4 but known to activate caspase-1 through other NOD-like receptors also led to caspase-7 activation, suggesting that this may apply to other inflammasome complexes (Figure 1D). The role of caspase-7 activation in response to organisms that engage different inflammasomes remains to be elucidated.
L. pneumophila-triggered caspase-7 activation is distinct from the activation seen during apoptosis, as inhibition of caspases-8 or -9 did not prevent activation following L. pneumophila infection (Figure S3), nor did it permit additional bacterial growth (Figure 3C). Our results show that, along with taking part in apoptosis, caspase-7 activation performs an additional unrecognized role.
The route of intracellular trafficking of L. pneumophila affects at least in part, the outcome of infection [20],[23],[25],[26],[47],[53],[54], but the factors that dictate this route are not very well understood. Here we show that in caspase-7−/− macrophages, only 20% of the internalized L. pneumophila were delivered to the lysosome as early as 30 min after infection. Within 4 hrs, less than 10% of the bacteria were still in the lysosome while 40% resided in endoplasmic reticulum-labelled vacuoles (Figure 4). Further, there were more rod-shaped (undegraded) organisms in the absence of caspase-7 (Figure S5). Therefore, the restriction of infection in macrophages is achieved at least in part, by the delivery of more organisms to the lysosome when caspase-7 is activated.
The presumptive role of caspase-7 as an executioner caspase, involved in the cleavage of apoptotic substrates, is based primarily on its close relationship with caspase-3. However, recent reports demonstrate that caspase-7 and caspase-3 are functionally distinct [46],[55]. As reported by others, we found that unlike caspase7, caspase-3 was activated by L. pneumophila in A/J-derived macrophages and not in wild-type macrophages (Figure 2D). Therefore, our data demonstrate that caspase-3 and caspase-7 are not simultaneously activated during L. pneumophila infection as they are during apoptosis.
Furthermore, the absence of caspase-7 but not caspase-3 was accompanied with permissiveness to L. pneumophila infection in vitro and in vivo (Figure 3). Caspase-3 is also activated independently of the classical apoptosis pathways [55],[56], and reports have suggested that caspase-3 activation promotes cell survival [57]. Several studies suggest that caspase-3 activation is essential for establishment of L. pneumophila infection by mediating the cleavage of Rabaptin5 [55],[56]. These data strongly suggest different roles of caspase-7 versus caspase-3, as they restrict and permit L. pneumophila infection, respectively (Table S1).
Interestingly, TUNEL analysis of infected macrophage population revealed that during the first 2 hrs of infection, 20% of infected wild-type macrophages were apoptotic which may hinder the optimal replication of the organism, whereas infected macrophages lacking caspase-7, -1 or a functional Naip5 did not undergo early apoptosis (Figure 5F). These data agree with previous reports that suggest that L. pneumophila delay apoptosis in permissive macrophages to allow for the establishment of the replicative vacuole [46],[58]. Since early apoptosis is delayed in macrophages lacking caspase-1, caspase-7, or functional Naip5, and since these cell types are defective in caspase-7 activation in response to L. pneumophila, we conclude that caspase-7 activation contributes to early apoptosis observed in wild-type macrophages. However, our conclusion does not rule out the contribution of other mechanisms [55],[59],[60].
High MOI led to significant apoptosis in wild-type, caspase-7−/− and A/J-derived macrophages but not in caspase-1−/− cells (Figure 5C). Hence, caspase-1 appears as a more general regulator of L. pneumophila-triggered apoptosis. Caspase-7, however, regulated apoptosis only in the early stage of infection at low MOI. Similarly, Korfali et al showed that caspase-7 is involved earlier than other effector caspases in the apoptotic execution process in DT40 B lymphocytes [61]. Despite this, L. pneumophila replicated to a similar extent in macrophages lacking caspase-1 or caspase-7 (Figure 3), suggesting that for L. pneumophila to establish infection, it is particularly important to delay apoptosis during the early stages of infection.
Lungs from infected mice did not show significant differences in apoptosis after 3 days whether they lacked caspase-1, caspase-7, or a functional Naip5 (Figure S8). This does not exclude the role of cell death in vivo but strongly suggests that there must be at least one other complementary mechanism employed through caspase-7 to restrict infection.
Macrophages harbouring A/J Naip5 allele are capable of caspase-1 activation in response to L. pneumophila as reported by our group and by Lightfield et al [28],[29],[33], nevertheless, they are defective in caspase-7 activation (Figure 2E). Therefore, it is possible that the lack of caspase-7 activation is responsible at least in part, for the permissiveness of the A/J cells to L. pneumophila infection. Increasing evidence suggest that the Naip family of proteins may have yet uncharacterized functions [33], [62]–[64]. We propose that wild-type Naip5 mediates the activation of caspase-7 by caspase-1 during infections at low MOI. Our suggestion is supported by the fact that Naip5 interacts with Nlrc4 (which binds caspase-1) and with caspase-7 [31],[48],[65].
However, recent reports showed that the complete absence of Naip5 prevents caspase-1 activation [33]. The authors suggested that Naip5 is upstream of caspase-1, but it is possible that that Naip5 is required for proper assembly of the Nlrc4 inflammasome and hence caspase-1 activation but that A/J-derived Naip5 is partially functional as suggested by the authors [33]. This partial functionality may permit inflammasome assembly and caspase-1 activation but fail to mediate downstream caspase-7 activation, as suggested by our results. The isolation of the Nlrc4 inflammasome and identification of its components could clarify these possibilities.
Since IL-1β and IL-18 are downstream of caspase-1 and IL-1β has been implicated in controlling the maturation of the phagosome containing the Mycobacteria zmp-1 mutant, we examined the role of IL-1β and IL-18 in L. pneumophila infection and in caspase-7 activation [66],[67]. IL-1β/IL-18 double knockout macrophages restricted L. pneumophila infection as efficiently as wild-type macrophages (Figure S6B). In addition, caspase-7 activity in the double-knockout cells was equal to that in wild-type cells (Figure S6A). Finally, exogenously-added IL-1β or antibodies against the corresponding receptors did not alter the number of L. pneumophila as measured by colony-forming-unit assays (Figure S6C). Therefore, our model suggests that caspase-7 activation and L. pneumophila restriction are mediated downstream of caspase-1 but independently of IL-1β and IL-18 (Table S1).
In summary, caspase-7 activation restricts L. pneumophila infection by mediating macrophage apoptosis during early stages of infection and by affecting the trafficking of the organism within the cell. How caspase-7 mediates these distinct functions remains unclear. It is possible that caspase-7 modulates host or bacterial factors involved in controlling apoptosis and vesicular trafficking in the cell. The identification of caspase-7 substrates during L. pneumophila infection is the focus of ongoing work.
Therefore, our findings establish a previously uncharacterized caspase-7 activation pathway downstream of the Nlrc4 inflammasome during L. pneumophila infection. Moreover, these results demonstrate a novel biological role for caspase-7 in the control of L. pneumophila intracellular infection in macrophages and in mice. This new pathway can be a target for compounds that could have therapeutic application in the context of intracellular infection.
Legionella pneumophila (L. pneumophila) strain Lp02, is a thymine auxotrophic derivative of Philadelphia-1 [19]. The dotA mutant strain Lp03 is defective in the Dot/Icm Type IV secretion system [68]. The flaA mutant L. pneumophila was previously described [69]. Bacterial strains were supplemented with a plasmid that complements thymine auxotrophy and expresses green fluorescent protein (GFP) at the post-exponential phase (PE) [22],[45]. L. pneumophila was cultured as described previously [22],[45] in ACES-yeast extract broth supplemented with ferric nitrate and L-cysteine. All experiments were performed in the absence of ferric nitrate and L-cysteine from the macrophage culture medium, to allow L. pneumophila multiplication only intracellularly. All in vitro infections were performed at an MOI of 0.5 for 30 minutes followed by rinsing of the infected macrophages which allowed the infection of 20–25% of macrophages with usually 1 organism, unless stated otherwise [28]. The quantification of the colony-forming units (CFU) in vitro and in vivo was performed as described [28].
Wild-type C57BL/6 (B6), caspase-7−/−, caspase-3−/−, and A/J mice were previously described and purchased from the Jackson laboratory [41],[70],[71]. Caspase-1−/− mice were from Dr. Amy Hise at Case Western University. MyD88−/−, TRIF−/−, and Nlrc4−/− mice were previously described [72],[73]. All knockout mice were in C57BL/6 background. Caspase-1−/− and caspase-7−/− mice were backcrossed to C57BL/6 background for 10 generations and previously described [37],[70]. IL-1β/IL-18 double knockout mice [74] in C57BL/6 background were obtained from Dr. A. Zynchlinsky, Max-Plank-Institute, Berlin (authorized by Dr. S. Akira, Japan).
Bone marrow macrophages were prepared from femurs of five to eight-week-old mice as previously described [28],[75].
Mouse caspase-7 plasmid pCAGGS vector (LMBP 3818) was obtained from Gent University in Belgium. The plasmid was deposited by Dr. P. Vandenabeele [76]. Control plasmids used for transfection were obtained from Amaxa (pMaxGFP) or cloned in Dr. Wewer's Laboratory (pLenti-dsRed). To deliver control and caspase-7 plasmids into mouse macrophages, two approaches were used: transfection with SuperFect (Qiagen) and nucleofection with Amaxa (Lonza). Briefly, in transfection, mouse macrophages were seeded in 24-well plate at a density 0.5×106 cells per well 24 hours prior the transfection. Next day, the media was changed leaving 400 µl of full media per well. To make a transfection mix, control or caspase-7 plasmids (1 µg/well) were mixed with SuperFect reagent (2.5 µl/well) in a total volume of 100 µl of serum-free media for 20 minutes. Then, transfection mix was added to the mouse macrophages bringing volume to 0.5 ml per well. Three hours later, transfection mix was replaced with 1 ml of full media and cells were left to recover for additional 24 hr before adding bacteria. In nucleofection approach, 6×106 cells were resuspended in 100 µl of nucleofection solution (mouse macrophage nucleofector kit) supplemented with 10 µg plasmid. Plasmids were delivered into macrophages with Y-01 program. After nucleofection, macrophages were resuspended in 6 ml of full media and 0.5×106 cells were plated per well in 24-well plate. Cell death was monitored throughout the assays. Bacteria were added 24 hr later.
Immunofluorescence experiments were performed as previously described [28]. L. pneumophila was detected with mouse anti-Legionella (Abcam) and secondary Texas Red conjugated antibody. Localization of markers on L. pneumophila phagosomes was performed as previously described [22]. Antibodies used were rat anti-lysosomal-associated membrane protein 1 (LAMP1; 1D4B; Developmental Hybridoma Bank) [77], rabbit anti-calreticulin (Stressgen Bioreagents) followed by fluorescent secondary antibodies (Molecular Probes). Nuclei were stained with nucleic acid dye 4′,6′-diamino-2-phenylindole (DAPI; Molecular Probes). In each experiment one hundred bacteria were scored. Experiments were performed at least three times. Samples were analyzed with The Zeiss 510 META Laser Scanning Confocal microscope and Zeiss Axioplan 2 upright microscope at The Ohio State University Microscopy Core Facility.
Macrophages were plated at 5×105 cells per well and infected as described above [28]. At designated time points, macrophages were lysed and plated on AYE plates for colony forming units (CFUs). When indicated, macrophages were treated with recombinant IL-1β or IL-18 (Calbiochem) or with anti-mIL-1-RI or mIL-18 R/IL-1 R5 antibodies (R & D systems) at time of infection. Caspase-8 inhibitor (Z-IETD-FMK), caspase-9 inhibitor (Z-LEHD-FMK), or caspase-1 inhibitor (Z-YVAD-FMK) (Calbiochem) were used at 50 µM concentration when indicated. Inhibitors were maintained during the course of infection.
Cell extracts were prepared and immunoblotted with an anti-body that recognizes the large subunit of mouse caspase-1, -3 or -7 (Cell Signalling), followed by appropriate secondary anti-rabbit antibody as described [28]. When indicated, macrophages were permeabilized with 10 ng/ml streptolysin O for 5 minutes in the absence or presence of ligands as previously described [28],[39], then rinsed and incubated for 2 hours [28]. Purified bacterial ligands were purchased from Invivogen.
IL-1β measurements were performed as previously described [72],[73]. Experiments were performed in triplicates.
The percentage of macrophage death was determined by measuring the release of host cell cytoplasmic LDH using the CytoTox 96 non-radioactive cytotoxicity assay (Promega) as previously described [28],[45]. The apoptosis inducer doxorubicin was added at 100 ng/ml, when indicated (Calbiochem). In vitro quantification of cytoplasmic (apoptosis) and extracellular (necrosis) histone-associated DNA fragments was performed using The Cell Death Detection ELISAplus photometric enzyme immunoassay kit from Roche to the specifications of the manufacturer. Apotosis of macrophages in vitro was assessed with fluorescent TUNEL (terminal deoxynucleotidyl transferase-mediated dUTP nick end-labelling) assay according to the manufacturer's specifications using In Situ Cell Death Detection Kit, TMR red from Roche. Experiments were performed in triplicates. Sections of infected mice lungs were stained for apoptosis using Apop Tag In Situ Apoptosis Detection Kit (Chemicon). TUNEL-positive stained cells (brown) were evaluated by capturing digital images using a 20× objective lens and covering the entire lung (at least 32 images per lung). All samples were handled in a blinded manner. The percent of brown pixels per high powered field (HPF) were quantified using Adobe Photoshop CS2 software histogram analysis.
Wild-type C57BL/6 and caspase-7−/− mice were infected intra-tracheally with 1×106 wild-type bacteria, and the number of bacteria in the lungs was determined at 6 hours and at 96 hours post-infection [28],[75]. All animal experiments performed were done according to animal protocols approved by the Animal Care Use Committee of The Ohio State University College of Medicine.
All experiments were done at least three independent times and yielded similar results. The data points represent the average ±S.D. Data were analyzed by Student's t-test. *, P value≤0.05 and was considered significant.
|
10.1371/journal.pntd.0005773 | Antibody trapping: A novel mechanism of parasite immune evasion by the trematode Echinostoma caproni | Helminth infections are among the most prevalent neglected tropical diseases, causing an enormous impact in global health and the socioeconomic growth of developing countries. In this context, the study of helminth biology, with emphasis on host-parasite interactions, appears as a promising approach for developing new tools to prevent and control these infections.
The role that antibody responses have on helminth infections is still not well understood. To go in depth into this issue, work on the intestinal helminth Echinostoma caproni (Trematoda: Echinostomatidae) has been undertaken. Adult parasites were recovered from infected mice and cultured in vitro. Double indirect immunofluorescence at increasing culture times was done to show that in vivo-bound surface antibodies become trapped within a layer of excretory/secretory products that covers the parasite. Entrapped antibodies are then degraded by parasite-derived proteases, since protease inhibitors prevent for antibody loss in culture. Electron microscopy and immunogold-labelling of secreted proteins provide evidence that this mechanism is consistent with tegument dynamics and ultrastructure, hence it is feasible to occur in vivo. Secretory vesicles discharge their content to the outside and released products are deposited over the parasite surface enabling antibody trapping.
At the site of infection, both parasite secretion and antibody binding occur simultaneously and constantly. The continuous entrapment of bound antibodies with newly secreted products may serve to minimize the deleterious effects of the antibody-mediated attack. This mechanism of immune evasion may aid to understand the limited effect that antibody responses have in helminth infections, and may contribute to the basis for vaccine development against these highly prevalent diseases.
| Helminthiases are highly prevalent neglected tropical diseases, affecting millions of people worldwide, mainly in the poorest regions. The lack of vaccines against these infections is one of the major constraints in the current parasitology and massive efforts are being done in that direction. Herein, we present a potential mechanism for parasite immune evasion consisting in trapping of surface-bound antibodies within the excretory/secretory products that are deposited over the parasite. This mechanism is aided by parasite-derived proteases, well documented virulence factors that degrade the entrapped antibodies. Altogether, this parasite strategy may serve to minimize the antibody-mediated response and promote the development of chronic infections. The present study has been done using the model trematode Echinostoma caproni, though is expected to work in other helminths, even in other groups of extracellular pathogens. This opens new expectative to better understanding of host-parasite interactions and susceptibility to helminth infections. Therefore, the results presented in this manuscript may contribute to the basis of anti-helminth vaccine development.
| Parasites are able to actively evade or manipulate the host immune system for their own benefit, either increasing their transmission or reducing clearance. That is crucial in the evolution of host-parasite interactions, pathology and virulence. Evasion strategies, such as antigenic variation, antigen masking, molecular mimicry or protease secretion are common among both protozoa and helminth parasites [1]. Ultimately, these mechanisms let the parasites disrupt or manipulate the host immune responses, both innate and adaptive, and/or prevent the formation of a memory response [1]. It is well known that antibodies can affect the development of helminth parasites by hindering processes such as attachment, feeding or motility, among others [2]. Several studies show that antibodies are able to target parasites, such as Echinostoma caproni or Nippostrongylus brasiliensis, for other immune effector mechanisms such as granulocyte and macrophage binding, or complement system activation [3–5]. Nevertheless, though antibody responses are commonly needed for controlling helminth infections, generally these are not sufficient to prevent nor overcome the infection [2].
The tegument of trematodes is a highly active structure with a key role in host-parasite interactions and the maintenance of tegument integrity is crucial for worm survival [6]. For these reasons, a number of tegumental proteins have been proposed as promising vaccine candidates against these helminthiases. However, though different levels of protection have been observed, antibody responses per se normally have limited effect and complete protection against infection has not been reached so far [7–9]. This suggests the existence of intrinsic mechanisms that limit the susceptibility of the tegument to the immune attack. Herein, we describe a potential novel mechanism for parasite immune evasion, which consists in the entrapment of surface-bound antibodies to limit the effects mediated by the humoral response.
E. caproni is an intestinal trematode, broadly employed as an experimental model for the study of the biology of this group of parasitic helminths, with emphasis on the host-parasite interactions. One of the key features that makes this trematode a suitable model for studying host-parasite relationships is its different compatibility among laboratory rodents [10]. Low-compatible hosts, i.e. rats or jirds, are able to rapidly expel the parasites. Conversely, hosts of high compatibility, such as mice or hamsters, develop chronic infections lasting more than 25 weeks [10–12]. In highly compatible hosts, such as mice, strong, Th1-type inflammatory responses are developed at the site of infection, together with elevated levels of oxidative stress and mucosal antibodies [13,14]. This response, however, is not effective in the clearance of the infection and does not affect worm establishment nor development [10–14]. Worm recovery rates in mice are high, and adults are larger and more fecund than those recovered from hosts of low compatibility are [12]. Altogether, these facts suggest that the parasite is well adapted to this environment and it is capable of avoiding, or minimizing somehow, the deleterious effects mediated by the immune response, including antibodies, developed in mice. Thereupon, the experimental model E. caproni-mouse have been used herein to further study the mechanism through which parasites are able to withstand the immune response and ensure their survival inside the host.
The strain of E. caproni employed and the infection procedures have been described previously [15]. Briefly, encysted metacercariae were removed from kidneys and periacardial cavities of experimentally infected Biomphalaria glabrata snails and used for infection. CD1 mice (male, 30–35 g) were infected by gastric gavage with 75 metacercariae of E. caproni. At 4 weeks post-infection mice were necropsied and the small intestine was longitudinally opened to collect the adult parasites.
The animals were maintained under conventional conditions with food and water ad libitum. This study has been approved by the Ethical Committee of Animal Welfare and Experimentation of the University of Valencia (Ref#A18348501775). Protocols adhered to Spanish (Real Decreto 53/2013) and European (2010/63/UE) regulations.
E. caproni adults were fixed by immersion in 4% paraformaldehyde, either immediately after isolation (0 min) or after incubation in RPMI 1640 culture medium (Life Technologies), at 37°C, during increasing time intervals (15, 30, 60 and 120 min). The immunostaining was performed as follows. Briefly, adults were blocked for unspecific unions in 5% BSA (Sigma-Aldrich) in PBS for 1 h, and then incubated for 1h 30 min with a mixture of two primary antibodies, which consisted of: 1) rabbit sera against either E. caproni-actin [16, 17] or E. caproni-enolase [16, 18] and 2) goat anti-mouse IgA or goat anti-mouse IgG, both conjugated with HRP (Nordic). Antibody solutions were prepared in PBS by mixing one of the antibodies in 1 (rabbit against E. caproni protein) and another one from 2 (goat against mouse immunoglobulin), both diluted 1/50 in the final mixture. Different combinations of these antibodies were used to confirm that staining patterns do not depend on specific parasite antigens nor immunoglobulin isotypes, i.e. that different parasite-secreted proteins and/or different antibody isotypes share a common pattern regarding the trapping process.
After carefully washing in PBS (3 times of 10 min each), adults were incubated simultaneously with 2 secondary antibodies: 1) goat anti-rabbit IgG conjugated with Alexa Fluor 647, which tagged rabbit antibodies specifically bound to parasite antigens in the previous step, and 2) goat anti-HRP conjugated with FITC, tagging the HRP-conjugated goat antibodies bound to mouse immunoglobulins. This incubation was performed for one hour in the dark and parasite specimens were washed again in PBS before their examination by confocal microscopy. Secondary antibodies (both from Jackson ImmunoResearch) were diluted to a final concentration of 1/250 each. All incubations were performed at room temperature, under gentle agitation. Antibody solutions were prepared in PBS and no detergents were employed to permeate the samples. Negative controls, employed to set acquisition parameters for confocal microscopy, were performed likewise, excepting the incubation with primary antibodies.
Specific anti-actin and anti-enolase antibodies were prepared in our laboratory through immunization of New Zealand white rabbits with recombinant proteins as described in [18]. Antibody specificity is proved herein by western blot (see below).
Fluorescent staining was visualized by laser scanning confocal microscopy on 10 specimens at each time point. Adult worms were obtained from 3 experimentally infected mice and randomly allocated in the different experimental groups (i.e. times of in vitro incubation), so that each group comprised adults from the different hosts. Images were analysed using FV10-ASW 4.2 and Imaris software.
The loss of in vivo bound antibodies on worm surface along time was quantitated using ImageJ software to calculate the percentage of image area covered by the fluorescent tag (FITC). Confocal micrographs (x400) were stacked to create Z projections that were converted into binary (black and white) images. Raw integrated density (RawIntDen), which is the sum of the values of all the pixels in the image, was measured and used to calculate the percentage of area covered by the fluorescent tag (% AC) according to the following formula, in which 255 is the density value of a positive (tagged) pixel in the binary image and areas are expressed in pixels:
% AC= RawIntDen/255Total area⋅100
Statistical significance in relation to non-incubated worms (0 min) was assessed by unpaired Student’s t test (p<0.05). Prior to analysis data normality was confirmed by Shapiro-Wilk test.
To address the role of parasite-derived proteases, adult specimens were incubated for 2 h in RPMI in the presence or absence of a cocktail of protease inhibitors (cOmplete, Mini, EDTA-free, Roche) added at proper concentrations from a 7x stock solution, following the manufacturer’s instructions. The worms were processed as described above and the immunostaining compared between the two groups. To see if E. caproni adults are susceptible of being recognized by specific antibodies after in vitro incubation, the parasites were incubated for 4 h in RPMI to totally eliminate in vivo-bound antibodies. After blocking, the worms were incubated for 1 h with the serum of E. caproni-infected mice, before to proceed with the general protocol for double indirect immunofluorescence using anti-mouse IgG as primary antibody. Pre-immune mouse serum was employed for negative controls.
For SEM, E. caproni adults were fixed in Karnovsky’s fixative (0.5 M glutaraldehyde, 2.5 M formaldehyde), washed in buffer solution and post-fixed in 2% osmium tetroxide in 0.1 M sodium phosphate buffer, pH 7.2, for 2 h before dehydration by critical point. Mounted specimens were sputter-coated with gold-palladium and examined in a Hitachi S4100 scanning electron microscope at 5 kV.
Inclusion in LR-white resin for TEM was performed by fixing the adult parasites in glutaraldehyde 2.5% overnight, washing them in phosphate buffer 0.1 M pH 7.2, and then post-fixing in 2% osmium tetroxide in phosphate buffer for 2 h. After several washes in water, parasites were sequentially dehydrated in 30%, 50%, 70% and 96% ethanol, 5 min each. Finally, the worms were sequentially incubated for 2 h in 33% LR-white resin in 96% ethanol, 66% LR-white resin in 96% ethanol, 66% LR-white resin in 100% ethanol and 100% LR-white resin in 100% ethanol. Samples were filtered in resin and polymerized at 60 uC for 48 h. Ultra-thin slices (60 nm) were stained with 2% uranyl acetate prior to visualization by TEM at 70 kV in a microscope Jeol JEM1010. Images were acquired using a digital camera MegaView III with Olympus Image Analysis software.
For the immunogold assay, E. caproni adults were fixed in Karnovsky’s fixative and included in LR-resin as described above. Grids were washed 5 times, 1 min each, in 20 mM Tris-HCl buffer, pH 8.2, (TB) containing 0.1% BSA and 0.05% Tween-20 and blocked for unspecific unions using goat serum, diluted 1:20 in TB, for 30 min. After washing, free aldehyde groups were blocked for 5 min in TB containing 0.02 M glycine, and washed again. Rabbit sera against E. caproni actin or E. caproni excretory/secretory products (ESPs) [17], diluted 1/10 in TB-0.1% BSA, was applied as primary antibody at 4°C overnight. Grids were washed as previously described and incubated for 1 h with gold-labelled secondary antibody, donkey anti-rabbit IgG coupled to 12 nm gold particles (Jackson ImmunoResearch), diluted 1/20 in TB. For double immunogold, samples were processed likewise, using rabbit sera against E. caproni ESPs as a primary antibody and a mixture of 2 secondary antibodies, which consisted of the same donkey anti-rabbit IgG described above plus a donkey anti-mouse IgG conjugated with 18 nm colloidal gold (Jackson ImmunoResearch), both diluted 1/20 in TB. Negative controls were performed using grids incubated with pre-immune rabbit sera as primary antibody.
The specificity of the primary antibodies employed herein was confirmed by protein electrophoresis in SDS-PAGE and western blot (S1 Fig). A total of 30 μg of ESPs, obtained as previously described [16], were loaded onto 4% stacking and 12% resolving polyacrylamide gels and electrophoresed in Tris-glycine SDS buffer. Proteins were elctrotransferred onto nitrocellulose membranes (0.45 μm) in 20 mM Tris, 192 mM glycine and 20% methanol buffer, pH 8.3, for 90 min at 200 mA. After 1 h blocking in 5% skimmed milk in PBS containing 0.05% of Tween-20 (PBS-T) at room temperature, blots were incubated overnight at 4°C in PBS-T containing each antiserum: rabbit polyclonal anti-E. caproni actin (1/2,000) [16, 17]; rabbit polyclonal anti-E. caproni enolase (1/2,000) [16, 18]; and rabbit polyclonal anti-E. caproni ESPs (1/4,000). Membranes were washed and probed with peroxidase-conjugated secondary antibody, goat anti-rabbit IgG in PBS-T (diluted 1/10,000 for actin and enolase, and 1/20,000 for complete ESPs) for 2 h at room temperature. Negative control was performed likewise, by incubating ESPs against serum of pre-immune rabbit (1/2,000) and secondary antibody (1/10,000). Blots were developed with Amersham ECL Advance Western Blotting Detection Kit (GE Healthcare) following the manufacturer’s instructions and images were taken with ChemiDoc Imaging System (Bio-Rad). The results of this trial are shown in S1 Fig.
High levels of mucosal antibodies are characteristic of E. caproni infections in mice [13]. With the aim to confirm the in vivo binding of luminal antibodies over the parasite surface, double indirect immunofluorescence was performed on E. caproni adults. Specific mouse IgA and IgG were detected on worms at 0 min, indicating that they are susceptible of being affected by antibody-mediated responses (Fig 1, S1 Movie, S2 and S3 Figs). At this time point, an intense staining with the different tags used (i.e. anti-E. caproni-actin, anti-E. caproni-enolase and anti-mouse IgA/G) was observed (Fig 1, S1 Movie, S2 and S3 Figs), indicating that the host immune response effectively targets E. caproni adults for antibody-mediated attack. Actin and enolase are immunogenic proteins, commonly found in ESPs of E. caproni and other trematodes, so that antibodies against both molecules were used as general markers to tag the parasite surface and the deposit of ESPs. Considering that distal tegument consists of a syncytial cytoplasm, externally limited by a plasma membrane, and that no detergents were employed for immunofluorescence, the ESP molecules detected on the parasite surface are outside the tegument itself. These results are in agreement with those from Simonsen and co-authors [3, 19], indicating that mouse antibodies bind to secreted antigens and form an external layer of immune complexes that covers the parasite. Moreover, Sotillo et al. [18] found that both molecules (actin and enolase) were among the most antigenic proteins in the ESPs of E. caproni.
The decrease in the fluorescent signal for in vivo bound antibodies on worm surface during in vitro incubation is shown in S4 Fig. Loss of surface-bound antibodies during in vitro culture has been previously described in E. caproni and other trematodes [1,3]. It was suggested that the shedding of surface antigens and, consequently, the antigen-bound antibodies, might be an adaptation of this group of parasites to withstand the host immune attack [20]. However, the approach we have followed herein, based on monitoring in vitro the dynamics of bound antibodies by double immunofluorescence, reveals a different mechanism of immune evasion. This new mechanism consists in entrapping the surface-bound antibodies within newly secreted products. Fig 1, S1 Movie and S3 Fig show how, as the time culture increases, an external layer of ESPs, stained in red (anti-E. caproni actin), appears over the in vivo-bound antibodies (seen in green/yellow). This new layer is almost continuous after 30 min of incubation in RPMI. Andresen et al. [20] found that antigen-antibody complexes were released from the parasite surface into the culture medium during no longer than 20–25 min, suggesting that surface turnover was completed by this time. Our results, however, yield a novel interpretation of this finding. It seems that the loss of surface-bound antibodies in vitro is not due to the turnover of surface antigens, but instead to the trapping of the antibodies underneath a layer of excreted/secreted molecules. Results were displayed for anti-IgA (Fig 1, S1 Movie, S2 and S3 Figs).
A 3D reconstruction of antibody trapping and degradation process was created at two time points of in vitro culture using Imaris software (Fig 2A). After 1 h incubation, it can be appreciated how the layer of in vivo-bound antibodies is beneath a continuous and relatively thick layer of secreted material, tagged either with anti-actin or anti-enolase antibody. This indicates that antibodies are not lost from the surface of the parasite, as previously suggested [3, 20] but, in contrast, they become hidden beneath a layer of ESPs. After 2 h in RPMI, in vivo-bound antibodies are scarcely detected on the parasite surface, suggesting that trapped antigen-antibody complexes were removed or degraded somehow. Indeed, antibody degradation by parasite-secreted proteases is well recognized as a mechanism to evade the host immune response [21].
The in vitro study has let us elucidate the dynamics of bound-antibodies through the incubation of worms in an antibody-free medium. However, antibody trapping is expected to function also within the host. In that case, both antibody binding and trapping must occur simultaneously, and the continuous entrapment of surface-bound antibodies may serve to disable their harmful impact over the parasite. To confirm our hypothesis that this is a dynamic process and to examine the role that secreted proteases may have in the context of this mechanism of immune evasion, two different experiments were carried out.
Firstly, adult worms were incubated for 2 h in culture media in the presence and absence of a cocktail of protease inhibitors. As it was expected, only a slight reduction in the anti-IgA staining was observed after 2 h when protease inhibitors were added to the culture media (Fig 2B and S5 Fig). This result indicates that antibody trapping not only hampers the accessibility of other immune molecules and cells to the bound antibodies, but also facilitates antibody degradation by parasite-derived proteases. Furthermore, this proves that the loss of green staining (in vivo-bound mouse antibodies) in vitro is not an artifactual result. The fact that in vivo-bound antibodies are still detected over the surface after culture in the presence of protease inhibitors demonstrates that the reduction in anti-mouse antibody staining (green) during incubation in not-supplemented medium is not consequence of a passive release of antibodies due to low-affinity bindings (Fig 2B and S5 Fig). A variety of proteases have been previously detected in the ESPs of E. caproni [16, 22] and, using protease inhibitors, herein we have also shown that ESPs of E. caproni have protease activity. Future studies will show which proteases are involved in the degradation of entrapped antibodies on the parasite surface.
Secondarily, to verify that the ESPs deposited over the surface of the parasite can be recognized by new antibodies, adult specimens were incubated for 1 h with the serum of E. caproni-infected mice and a HRP-conjugated anti-mouse IgG as a secondary antibody. Previously, these worms had been kept for more than 2 h in RPMI to ensure the elimination of in vivo-bound antibodies. Immune mouse sera effectively tagged the surface of worms, whereas unspecific antibody binding was not observed over those incubated with pre-immune serum, indicating that ESPs accumulated on the surface can be the target for new antibodies (Fig 3). Furthermore, this confirms that antibody binding to parasite antigens is specific and that nonspecific unions, e.g. through Fc, do not occur. Altogether, these findings suggest a constant and reciprocal interplay between parasite- and host-released molecules at the site of infection. Antibody responses have been proven to have little effect on worm survival and development in E. caproni primary infections, with the highest rates of establishment and longevity being associated with high levels of mucosal antibodies [12, 13]. The evasion mechanism described herein may serve to explain the lack of effectiveness of these responses.
To examine if the proposed mechanism is compatible with the tegument dynamics and ultrastructure, the surface of E. caproni was studied by transmission and scanning electron microscopy (TEM and SEM, respectively). Moreover, immunogold labelling using anti-actin and anti-ESPs polyclonal antibodies was performed. Secretory vesicles of different morphology are highly abundant in the tegumental syncytium, indicating a very active secreting surface. Elongate and circular vesicles [20] accumulate manly at the apex, where they fuse with the external plasma membrane and empty their content to the outside (Fig 4 and S6 Fig). Immunogold labelling with anti-E. caproni ESPs showed a widespread staining of both the tegumental syncytium and the parasite surroundings, i.e. external surface, secretions and extracellular vesicles (Fig 4A). Gold particles inside membrane-bound vesicles, either elongated or circular, are seen profusely thorough the syncytium (Fig 4A). Similar results were observed with anti-E. caproni actin, though specific staining was much less extensive as could be expected when detecting a discrete molecule (Fig 4B and 4C). Unstained negative control is shown in S6 Fig. This demonstrates that secreted proteins are incorporated in tegumentary vesicles that fuse with the plasma membrane in the apex and discharge their content to the extracellular milieu. In Fig 4B, specific anti-actin staining is seen inside an elongated vesicle opened in the apex and in the vicinity of an opened vesicle. Packing of actin molecules within apical circular vesicles is shown in Fig 4C. According to these results, antibody trapping by newly secreted products is mechanistically feasible and it may occur in vivo. This was further confirmed by double immunogold for ESPs and mouse antibodies, showing that host antibodies are trapped by ESPs both on the parasite surface (Fig 5A) and within the extracellular secretions in the tegument vicinity (Fig 5B and 5C). High-resolution SEM reveals that a layer of extracellular material is deposited on the parasite surface, both on the ventral and dorsal sides (Fig 6). Highly likely, this layer consists of a mixture of parasite secreted proteins and host-derived molecules (i.e. trapped antibodies) and corresponds to what is detected by double indirect immunofluorescence.
Although we cannot discard that antibody shedding effectively operates to evade the host immune response, our results indicate that it is less relevant than the mechanism proposed herein. Antibody shedding was suggested as a mechanism of immune evasion based on the facts that adult worms lost surface-bound antibodies during in vitro culture [3] and they rapidly released antigen-antibody complexes into the culture medium [20]. Hence, it was hypothesized that bound antibodies were removed from the surface due to antigen turnover. Our results of confocal microscopy show that, in culture, most of antibodies are not shed, as suggested by Andresen et al. [20], but they remain bound to the surface beneath a layer of ESPs. Loss of antibodies from the parasite surface is mainly related to an initial trapping by ESPs, which masks surface-bound antibodies, and the subsequent degradation of trapped antibodies by parasite-derived proteases. The fact that surface-bound antibodies became entrapped within a layer of ESPs may explain the interpretation of Andresen and co-workers [20], since this layer prevents antigen-antibody shedding, thus antibodies were non-detectable by the modified ELISA method used by those authors. The question that raises, however, is how newly secreted antigens get to cover the layer of surface-bound antibodies. In view of our results, it is tempting to hypothesize that discrete antigens are able to diffuse through the layer of antigen-antibody complexes, generating a gradient of ESPs from the parasite to the outside that gets to cover the antibody layer. As secreted antigens are not anchored, the most external molecules are progressively released into the medium, thus the covering of antibodies is maintained due to the continuous secretion of ESPs. In contrast to discrete molecules, antigen-antibody complexes are not expected to diffuse due to their greater size, becoming entrapped within this mesh of antigens, which further facilitates their retention.
In vivo, antibody binding and trapping by ESPs is a continuous process in which antibodies and ESPs are overlapping. In this context, parasite-derived proteases may play a critical role by degrading the layers of molecules that are continuously formed on the parasite surface. Despite the complexity of this process, our experimental design has allowed to determine how it occurs and its potential consequences on parasite survival within a hostile environment.
In conclusion, the results presented in this paper lead to a new interpretation of classic studies on tegument dynamics in parasitic trematodes. As with other common evasion mechanisms (i.e. antigen shedding or protease cleavage), antibody trapping and degradation is expected to function in other helminth parasites in addition to E. caproni, alleviating the deleterious effects of antibodies and promoting parasite survival. Our current results suggest that antibody trapping may occur through the covering of surface-bound antibodies with secreted antigens. Nevertheless, future studies with E. caproni and other helminths are needed to elucidate how the trapping process happens in vivo and which proteases are specifically involved in antibody degradation. Helminth infections affect millions of people worldwide, mainly in the poorest regions, and have a tremendous economic impact in the livestock sector. In the context of host-parasite relationships, the immune evasion mechanism described herein may help to understand the limited effectiveness that the antibody responses, per se, have against this group of parasites.
|
10.1371/journal.ppat.1000690 | Functional Memory B Cells and Long-Lived Plasma Cells Are Generated after a Single Plasmodium chabaudi Infection in Mice | Antibodies have long been shown to play a critical role in naturally acquired immunity to malaria, but it has been suggested that Plasmodium-specific antibodies in humans may not be long lived. The cellular mechanisms underlying B cell and antibody responses are difficult to study in human infections; therefore, we have investigated the kinetics, duration and characteristics of the Plasmodium-specific memory B cell response in an infection of P. chabaudi in mice. Memory B cells and plasma cells specific for the C-terminal region of Merozoite Surface Protein 1 were detectable for more than eight months following primary infection. Furthermore, a classical memory response comprised predominantly of the T-cell dependent isotypes IgG2c, IgG2b and IgG1 was elicited upon rechallenge with the homologous parasite, confirming the generation of functional memory B cells. Using cyclophosphamide treatment to discriminate between long-lived and short-lived plasma cells, we demonstrated long-lived cells secreting Plasmodium-specific IgG in both bone marrow and in spleens of infected mice. The presence of these long-lived cells was independent of the presence of chronic infection, as removal of parasites with anti-malarial drugs had no impact on their numbers. Thus, in this model of malaria, both functional Plasmodium-specific memory B cells and long-lived plasma cells can be generated, suggesting that defects in generating these cell populations may not be the reason for generating short-lived antibody responses.
| Malaria causes considerable human suffering resulting from associated high mortality, morbidity and reduced economic productivity in endemic areas. Current control methods are thwarted by a multiplicity of problems including rapidly developing resistance for anti-malarial drugs and insecticide-treated nets, and huge costs and hence poor coverage with bed nets in poor countries. Understanding the basis of the inefficiency of immunity to malaria in childhood will greatly aid the search for effective vaccines, which together with drugs and vector control, will be essential in the drive to eliminate malaria. Because of the strong evidence associating anti-malarial antibodies with anti-parasitic and anti-disease effects, vaccines inducing protective long-lasting antibody responses are attractive. However, it has been suggested that antibody responses to some Plasmodium antigens may be not long-lived. It would be important to determine whether long-lived plasma cells and memory B cells are generated after a malaria infection; however, these studies are difficult to perform in humans. Therefore we investigated the kinetics, duration and characteristics of the two cell types responsible for long-term antibody production in a mouse model of malaria. We show here that malaria-specific memory B cells and plasma cells are still detectable more than eight months after infection, and that both long-lived malaria-specific antibody-secreting cells and functional malaria-specific memory B cells can be made after a single infection.
| There is longstanding evidence that naturally acquired immunity to the erythrocytic stages of malaria is strongly dependent on antibodies (Abs) [1]–[6]. However, acquisition of immunity to P. falciparum malaria in humans is a relatively inefficient process; slow to develop, never sterile and wanes quickly in the absence of continued exposure to infection [7]–[9]. This would suggest that intermittent exposure to parasite antigens is required, at least for several years, for maintenance of both the memory and effector arms of the immune response to P. falciparum. The biological explanation for this apparent dependence of naturally acquired immunity to continued antigen exposure in residents of malaria endemic areas is still a subject of debate.
During both experimental and human malaria, there is evidence for loss of memory or activated CD4+ T cells, B cells and plasma cells and short-lived malaria specific Abs after a primary acute infection [10]–[13], suggesting that some of the components contributing to the humoral response may be short-lived. Moreover, some studies have suggested that maintenance of malaria-specific Abs is dependent on the presence of chronic parasitemia [14]. However, there are conflicting reports on the longevity of Ab reponses to Plasmodium; in some longitudinal studies [14]–[16], short-lived Ab responses with reduced half lives [17] have been reported, whereas other studies report that Ab responses persist [18],[19] and are protective [5], and it has yet to be settled whether there are any deficiencies in the generation and maintenance of Plasmodium-specific memory B cells and Abs.
Long term production of Abs is maintained by a combination of short-lived and long-lived plasma cells (PC), usually defined functionally as Ab secreting cells (ASC). Although short-lived ASC die within 3–5 days, Ab levels can be maintained by continuous proliferation and differentiation of memory B cells (MBC) into short-lived ASC upon continuous re-activation by either persistent antigen (chronic infection) [20],[21] or polyclonal stimulation [22]–[24]. Alternatively, long-term production of Ab is maintained by long-lived ASC, which migrate to survival niches within the bone marrow [25]–[29] and spleen [30] and can exist for the life-time of the mouse [28], [30]–[32], and this is probably also the case in humans [33]. Long-lived PC are thought to be independent of MBC [33],[34], suggesting that MBC do not have a direct role in the maintenance of pre-existing serum Ab. However, antigen specific MBCs provide rapid ASC responses upon re-encountering specific antigen, resulting in high titres of specific Ab. One explanation for the potentially short-lived nature of anti-Plasmodium Abs could be that they are predominantly produced by short-lived ASC, which are not replenished due to a defect in the MBC compartment, or that there is a defect in the long-lived ASC compartment.
There are very few studies investigating the cellular basis of the Ab responses to P. falciparum antigens in humans. In one report, MBC have been detected in blood as long as 8 years after a P. falciparum infection [35], whereas more recently we have reported that stable populations of circulating P. falciparum-specific memory B cells are not maintained in exposed adults in an endemic area of malaria transmission [36]. The discrepencies may be due to the difficulties of doing such studies in humans, where there is only access to peripheral blood as a source of lymphocytes and ASC. Although MBC can traffic in peripheral blood [23], [37]–[41], ASC are normally only seen in peripheral blood mononuclear cells (PBMC) either en route to the bone marrow [42] after differentiation in secondary lymphoid organs, or after dislodgement from their survival niches in the bone marrow [42],[43]. Therefore blood cannot give an accurate readout of MBC or long-lived ASC.
Experimental models of malaria where lymphoid organs including bone marrow can be accessed may provide valuable information on the contribution of long- and short- lived MBC and ASC to the protective Ab response. Plasmodium chabaudi infections in mice give rise to a primary infection with high parasitemia followed by a 2 to 3 month low grade chronic infection [44], and therefore can inform us about the impact of both acute and persistent low-level parasitemia on the subsequent generation and maintainence of Plasmodium-specific MBC and ASC. Here, we have investigated the kinetics and duration of malaria-specific MBC and ASC responsible for serum Ab in this infection, using a fragment of the P. chabaudi protein, Merozoite Surface Protein 1 (MSP1) to track specific cells. We used the region of P. chabaudi-MSP1 analogous to the c-terminal 19kDa part of P. falciparum MSP1 (MSP119), a well described candidate for a potential malaria vaccine. We show that malaria- (P. chabaudi MSP119) specific ASC and MBC are long-lived, and are detectable for more than eight months following a primary infection. Memory B cells are functional, giving rise to elevated levels of MSP119-specific ASC in a second infection secreting the classical T-cell dependent isotypes IgG1, IgG2c (the IgG2a equivalent of C57BL/6 mice) and IgG2b Abs characteristic of a memory B cell response. Using cyclophosphamide treatment and drug-induced parasite clearance, we demonstrate that maintenance of ASC is not dependent on chronic infection and that long-lived ASC resident in both spleen and bone marrow are generated. Our data support the idea that despite the drop in Ab titres following acute malaria infection and regardless of chronic infection, long-lived memory B cells and plasma cells secreting anti-MSP119 Abs can be generated.
We examined whether MSP119-specific Ab secreting cells (ASC) and MBC (MBCs) could be detected several months after a P. chabaudi infection in C57BL/6 mice. An acute blood stage infection following an inoculum of 105 parasite-infected red blood cells (iRBC) characteristically shows a maximium parasitemia at day 8 with approximately 30% of red blood cells infected, before dropping rapidly to very low parasitemias by day 20 after which low-level sub-patent chronic infection can ensue for up to 75–90 days (Figure 1A, [44]). Such sub-patent chronic infections have been demonstrated by passive transfer of chronically infected blood into immunocompromised mice [44]. Additionally, splenomegaly [45]–[48] and a transient depletion of bone marrow cells [48] at the peak of infection always accompany these infections.
The numbers of MSP119-specific IgG ASC and MBCs in the spleens and bone marrow of infected (and uninfected-control) mice were determined by ELISpot at various time-points up to 250 days post infection. There were low background MSP119-specific IgG ASC numbers averaging 276.42±400.79 (standard deviation, n = 18) and 238.43±176.70 (standard deviation, n = 12) per spleen and bone marrow of uninfected mice, respectively. However, MSP119-specific IgG ASC (above naïve-background) were detected in the spleens of infected mice as early as 10 days after infection (Figure 1B). Consistent with the rise and fall of the MSP119-specific IgG Ab response described previously [44], MSP119-specific ASC increased rapidly with peak cell numbers at day 30 but dropped by 95% by day 45 (Figure 1B). Thereafter, the numbers of splenic MSP119-specific IgG ASC were maintained at relatively low numbers, and by day 250 post infection there were still approximately 2000 (median) MSP119-specific ASC per spleen. In contrast, the kinetics of appearance of MSP119-specific ASC in bone marrow was different. MSP119-specific ASC were not detectable until day 20 of infection, whence they increased rapidly to a peak at day 30, followed by a 2-fold drop, and maintained at these levels thereafter (Figure 1C). Although there was a trend for more anti-MSP119 ASC in bone marrow than in spleen by day 250, the medians were not different (p≥0.05, Mann Whitney). In addition, the numbers of MSP119-specific IgG ASC in the bone marrow and spleen were strongly correlated with the concentrations of MSP119-specific IgG in plasma (Supplementary Figure S1), suggesting that concentrations of plasma Ab can be a good surrogate for ASC (Plasma cells) in the bone-marrow.
The numbers of MSP119-specific IgG MBC, determined by limiting dilution analysis as described [49] were maximal in the spleen at day 20, but then dropped rapidly to approximately 1000 MBC per spleen at day 30 and further until day 75, after which time they were present in relatively small numbers for the rest of the observation period (Figure 1D). On average and in agreement with other reports of memory B cell maintenance [49], there was a 95% reduction of MSP119-specific IgG MBCs between the peak at day 20 and the relatively reduced numbers at day 75. Unlike MSP119-specific ASC, MBC were not detected in the bone marrow at any time during the observation period (data not shown).
Together, these data show that C57BL/6 mice infected with P. chabaudi generate both MSP119-specific IgG ASC and MBC. ASC are present in both spleens and bone marrow in the chronic infection and long after, suggesting that both organs are sites for long-term Ab production in this infection. MSP119-specific MBC, although not present in bone marrow, are similarly sustained long-term in low numbers in the spleen.
The ASC that were detectable several weeks and months after the primary infection may have been present because of persistent stimulation by the subpatent chronic infection, which can last up to 90 days in C57BL/6 mice [44]. To investigate whether this was the case, infected mice were treated with curative doses of either chloroquine (CQ, Figure 2) or mefloquine (Supplementary Figure S2) to eliminate the chronic P. chabaudi infection. CQ and MQ were given in 3 (days 30, 32 and 34 of infection) and 4 (days 30, 31, 32 and 33 of infection) doses, respectively. The numbers of MSP119-specific IgG ASC in spleens and bone marrows of infected mice and IgG Ab levels were analysed 15 and 30 days after the inception of treatment (days 45 and 75 of infection). Treatment of mice with either of the two drugs did not affect the total numbers of splenic and bone marrow cells at any of the various time points that were analysed (data not shown).
Elimination of chronic parasitemia by CQ and MQ did not affect the numbers of splenic and bone marrow MSP119-specific IgG ASC nor the specific Ab levels with no significant differences in the numbers between the drug-treated and untreated mice at either time point measured (Figure 2, and Supplementary Figure S2). Thus, the MSP119-specific IgG ASCs responsible for the maintenance of MSP119-specific IgG production are retained independently of chronic infection for at least 6 weeks. In addition, there were no differences in the isotype distribution of plasma Abs between the drug-treated and untreated (chronically-infected) mice (Supplementary Figure S3), suggesting that presence of low grade chronic infection may not influence Ab function.
CQ can inhibit MHC class II antigen-presentation in vitro [50], and thus could itself affect a helper T cell/Ab response irrespective of P. chabaudi infection. This did not appear to be the case, as uninfected mice immunised with MSP119, and given the same chloroquine regimen as infected mice after 30 days of immunisation had similar levels of MSP119-specific Abs and CD4 T cell responses (Supplementary Figure S4) and similar numbers of specific ASC in spleen and bone marrow compared with those of untreated immunised mice (data not shown), suggesting that this dose of CQ in vivo does not affect the magnitude of a B cell response. This is consistent with previous observations showing that chloroquine treatment did not reduce T cell responses [51] or anti-Plasmodium Abs [44] in vivo.
Although ASC could be detected for up to 250 days following primary infection (Figure 1), this type of analysis could not distinguish between intrinsically long-lived ASC, which survive and secrete specific Ab for the life of the mouse [28],[30], and continuous proliferation and differentiation of MBC into short-lived ASC. Therefore, we investigated whether any long-lived ASC were generated during the primary infection.
To determine whether long-lived ASC were generated after a primary infection of P. chabaudi, we first determined the longevity of the total ASC regardless of specificity. In this experiment, mice received BrdU in drinking water for 2 to 4 week-periods at different times of the infection (i.e., 0–2, 2–4, 4–6, 6–8 and 8–12 weeks), and the resultant CD138+ PC (gated as shown in Supplementary Figure S5) at 12 weeks of infection were analysed to determine whether PC in the spleen and bone marrow retained BrdU from any of the labelling periods (Figure 3A–F). BrdU treatment did not affect the course of the P. chabaudi infection (data not shown).
The majority of the PC labelled during the early acute infection (mice were given BrdU only during weeks 0–2), were not present at 12 weeks after infection, suggesting that the majority of CD138+ PC generated within the first 2 weeks of the acute infection were indeed short-lived in either bone marrow or spleen (Figure 3B and C). Similarly, when BrdU was given in 2 or 4 week-periods (0–2, 2–4, 4–6, 6–8 or 8–12, Figure 3D) after a primary infection, the largest populations of BrdU+ PC found at 12 weeks in spleen and bone marrow were generated in the 6 weeks immediately preceding sampling (ie, between 6 to 12 weeks of infection, Figure 3E and F). However, a small proportion of PC in spleen and bone marrow that had incorporated BrdU in the earlier periods of the infection were still present after 12 weeks post infection in spleen, suggesting that some longer lived PC were residing in both spleen and bone marrow. In total, approximately 50% of PC present at 12 weeks of infection were formed over the course of the infection. The PC remaining unlabelled after 12 weeks of infection most probably pre-date the infection, and were likely to be be specific for non-malarial antigens and therefore not important for this analysis.
The specificity of the PC for P. chabaudi antigen(s) could not be determined by this flow cytometric analysis. Therefore to demonstrate the presence of long-lived antigen-specific PC, MSP-1 specific ELISpot assays were performed, in which infected mice were treated for 4 days with the immunosupressive drug, cyclophosphamide (CY), at different times in the infection (i.e., days 8, 30 and 45, Figure 4), and the number of splenic and bone marrow MSP119-specific ASC determined seven days after the the initiation of the CY treament (Figure 4A). This regimen has been shown previously to delete short-lived plasmablasts entirely, whilst not significantly affecting numbers of already established long-lived ASC [52],[53]. In addition, CY-treatment did not affect the course of P. chabaudi infection in the experiments described here (Supplementary Figure S6).
Splenic and bone marrow MSP119-specific ASC in CY-treated mice were compared with those of similarly infected, but untreated age-matched infected mice. CY-treatment at days 8 and 30 of infection resulted in significant reductions of MSP119-specific IgG ASC in spleen (Figure 4B). In contrast, CY-treatment at day 45 did not affect the size of the MSP119-specific IgG ASC pool in the spleen at day 52 suggesting that by this time, new splenic MSP119-specific ASC were not being generated (Figure 4B) and the ASC detectable in both groups are terminally differentiated long-lived cells (resistant to CY). The differences between CY-treated and untreated mice observed early in the infection could not be ascribed to obvious diffences in parasitemia (antigen dose) as the 7 day treatment period did not appear to affect the course of infection (Supplementary Figure S6).
In the bone marrow, the pool of MSP119-specific IgG ASC was not affected by CY treatment at day 30 of primary infection, consistent with the idea that the ASC that migrate to the bone marrow are long-lived (Figure 4B). Surprisingly, although the pool of MSP119-specific IgM ASC was completely depleted by CY-treatment at day 8, it was not affected by treatment at day 45 in either spleen or bone marrow suggesting that even antigen specific IgM secreting ASC could be long-lived (Figure 4B).
Thus infection of C57BL/6 mice with P. chabaudi induces the generation of long-lived ASC that maintain anti-malaria Ab levels independently of chronicity of infection. However, this does not tell us whether persistent low-grade infections affect the longevity of ASC themselves. To test this possibility, infected mice were either treated with CQ or left untreated (sex and age matched controls) after 30 days of infection. Mice were then treated with CY on days 45, 46, 47 and 48 and sacrificed for the determination of splenic and bone marrow MSP119-specific ASC (IgG and IgM) at day 52 of infection (Figure 5A). There were no significant differences in the numbers of long-lived ASC between the chronically infected and CQ-cured mice (Figure 5B) suggesting that persisting low grade infections do not affect longevity of MSP119-specific ASC.
Together, these data suggest that whilst the initial acute MSP119-specific ASC response following a primary infection of mice with P. chabaudi is comprised of predominantly short-lived ASC, a significant proportion of long-lived ASC of both IgG and IgM isotypes are also generated. Long-lived MSP119-specific IgG ASC are observed in the bone marrow from day 30, and in the spleen from 45 days of infection. Long-lived MSP119-specific IgM ASC can be detected in both spleen and bone marrow after 45 days of infection. Generation of long-lived ASC is not affected by persistent low grade infections that characterise P. chabaudi infections of mice.
MSP119-specific MBC were detectable in spleens of P. chabaudi-infected mice up to 250 days following primary infection (Figure 1). However, it is possible that these MBC were not fully functional memory cells as a result of the prolonged chronic infection that characterises this infection. Therefore we asked whether MBC present in the spleens of previously infected mice could generate a classical secondary response as evidenced by rapid increase in the number of specific ASC secreting the full range of IgG isotypes typical of a recall response, and replenishment or increased size of the specific memory B cell pool. C57BL/6 mice which had recovered from a primary P. chabaudi infection initiated 100 days previously were given a second challenge with the same dose of iRBC (schematically shown in Figure 6A). As reported previously [44] this resulted in low transient parasitemia (approximately 0.01% parasitemia at day 10, data not shown). Consistent with a memory response, there was an increase in the numbers of splenic MSP119-specific IgG ASC already by day 20 of the second infection in spleen and in bone marrow (Figure 6B, i–iv). At this time of a primary infection there were only very few MSP119 IgG ASC suggesting that the secondary B cell response was more rapid.
In addition to a faster MSP119-specific IgG ASC response, the secondary response was composed predominantly of IgG ASC, (97%, 93% and 83% of all anti-MSP119 ASC at days 10, 20 and 30 of secondary infection respectively), in contrast to the primary response which contained a large IgM component (Figure 6B, i–iv). IgM MSP119-specific ASC were the first to appear in the spleens of primary infection and were predominant at 91% and 61% of the total MSP119-specific ASC in the spleen, at days 10 and 20, respectively (Figure 6B, i).
Breakdown of the IgG MSP119-specific ASC in to the different IgG isotypes in primary and secondary responses revealed classical primary and secondary B cell responses, respectively (Figures 6B, v–viii). IgG3 MSP119-specific ASC comprised the majority of IgG ASC in the spleen in a primary infection, whereas they became the minority in the secondary ASC response, when IgG2c ASC predominated. In both primary and secondary responses in the spleen the greater part of the ASC response was short-lived, and ASC were reduced to low levels by day 45 and 30 respectively.
The isotype composition of ASC in the bone marrow was different from that in the spleen (Figure 6B). Firstly, as described in Figures 6B-iii and 6B-vii, there was no evidence of a large number of shortlived ASC in the primary infection in the bone marrow, rather, relatively steady maintenance of numbers after day 30 was observed. In the secondary response there was a large transient IgG ASC response in the spleen (and not the bone marrow) peaking at day 20 and lower at day 30. The maintenance of ASC of IgM isotype in the bone marrow of primary infection (Figure 6B, iii and vi) was unexpected, but is consistent with the idea that IgM PC can be long-lived (Figure 4B, [30],[34]). Unlike in the spleen, there were relatively few MSP119-specific ASC secreting IgG3 (Figure 6B, viii). All other IgG isotypes, with IgG2c being predominant were maintained in the bone marrow for longer than in spleen (compare Figures 6B, vi and viii).
Together, these results show that there is an MSP119-specific memory B cell response that is faster and composed predominantly of the IgG isotypes more typically associated with a secondary or memory response, particularly the opsonising IgG2c isotype which is thought to play an important role in the protective Ab response to blood stage malaria infections [54]–[56].
Although a classical secondary B cell response was observed after rechallenge with P. chabaudi, it is possible that the presence of prolonged low-level infection could have impaired the secondary response. We therefore investigated whether the removal of the chronic primary P. chabaudi infection by drug-cure would result in a quantitatively improved memory ASC response upon a secondary infection (Figure 7A).
There was a trend towards faster and higher MSP119-specific IgG ASC and MBC responses in the spleen and bone marrow of previously drug-cured mice (Figures 7B, and Supplementary S7), consistent with our earlier observation of higher titres of malaria-specific Ab in these mice upon rechallenge [44]. However, the differences were not significant and therefore these data suggest that the low level chronic parasitisation of C57BL/6 mice has only minimal impact on the generation and development of functional MSP119-specific IgG MBCs and ASC. In addition, there were no differences in the levels of IgM and IgG subclass Abs between the drug-cured and chronically infected mice at day 30 of re-infection suggesting that low-grade chronic infections do not affect the isotype distribution of Abs in memory responses (Supplementary Figure S8).
In humans infected with the malaria parasite, both short-lived [14]–[16] and long-lived [18],[19] anti-Plasmodium Ab responses have been reported in longitudinal and cross-sectional surveys. However, the cellular and molecular determinants of the longevity of specific Ab responses have not been investigated in detail. Previously we have shown in an experimental infection of C57BL/6 mice with P. chabaudi that the MSP119-specific IgG Ab response is stably maintained, but at low levels, for several months following the decay of the acute peak Ab response [44], suggesting that a single infection can result in long-lasting production of some Plasmodium-specific Abs. Here we have shown that both MSP119-specific IgG MBC and ASC are generated, and are maintained above naïve background for over 8 months, and importantly that long-lived ASC are maintained independently of the presence of a chronic infection. Whether the naïve background observed is the result of B cell responses to cryptic self-epitopes that are cross-reactive with MSP119, and whether it has a genetic basis will become clearer in future studies. The increase and decrease in the acute anti-MSP119 IgG response reported previously mirror that of ASC in the spleen, while the later lower IgG Ab levels correlate well with maintenance of ASC numbers in both spleen and bone marrow. Although the kinetics of appearance of MSP119-specific IgG ASC in the spleen and bone marrow appear to be different, they were continuously found in both locations suggesting that both organs are sites of long-term anti-Plasmodium Ab production. Our findings are consistent with observations in other infectious disease models such as LCMV infection of mice [30] where long-lived specific ASC have been found in both spleen and bone marrow.
It has been thought that persistent secretion of serum Abs is the result of the continuous activation of antigen-specific memory B cells and their differentiation into short-lived ASC [57]–[62]. However, several investigators have shown that a substantial fraction of antigen-specific PC can survive for years in the bone marrow of immunized mice from where they continue to secrete Abs for extended periods of time in the absence of detectable MBC and antigen stimulation [30],[31],[63],[64]. In the P. chabaudi infection of mice described here, parasitemia typically peaks at day 8 of infection and then declines rapidly and a chronic phase ensues that can be maintained for up to three months [44], thus allowing us to determine the impact of low grade chronic infection on generation of ASC. Using two anti-malarial treatment regimens, CQ and MQ, to remove the chronic infection, we have observed that by day 45 of infection MSP119-specific ASCs are maintained independently of the low-grade chronic infection.
Since PC are terminally differentiated non-dividing cells, they are neither sensitive to irradiation nor to cell division inhibitors. Thus the persistence of auto-, as well as anti-microbial, Abs in humans suffering from autoimmune diseases despite high doses of treatments with immunosuppressive drugs is an indirect evidence for the presence of long-lived PC [65]. The contribution of long-lived PC to the maintenance of serum Ab levels for prolonged periods is further strengthened by the observation that depletion of peripheral B cells with anti-CD20 monoclonal Abs in both humans [33] and mice [34],[66] does not affect concentrations of anti-microbial Abs. In agreement with these reports, our data suggest that in this P. chabaudi infection CY-resistant long-lived PC that secrete anti-Plasmodium Abs are generated and maintained in the later stages of malaria infections. This finding also agrees with the results of our other experimental approach; namely that a proportion of BrdU-labelled PC were detected in the spleens and bone marrow of infected mice for up to 12 weeks of observation. However, PC labelled with BrdU during the acute phase of infection contributed little to the long-lived PC pool suggesting that the initial anti-Plasmodium response is predominantly composed of short-lived PC. Consistent with this result, CY treatment of infected mice at days 7 and 30 of infection completely abrogated and significantly reduced the respective anti-MSP119 PC responses. This predominance of short-lived PC in the acute B cell response is to be expected and can be explained by a rapid expansion of polyclonally activated B cells, their differentiation into ASC and subsequent depletion via clonal selection and affinity maturation [67]. Thus it is unlikely that short-lived and long-lived PC are derived from different precursors.
There are no comparable studies on specific ASC in humans with either low-grade chronic malaria infections, or acute malaria infections. These would be very difficult to carry out, as ASC are located primarily in bone marrow and lymphoid organs. Cross-sectional studies measuring Plasmodium specific Ab levels (as a measure of ASC activity) suggest that like this P. chabaudi infection, acute P. falciparum infection is accompanied by higher specific Ab levels [14],[17],[68], but that some Ab responses are maintained in the absence of obvious clinical malaria or parasitemia [18],[19]. Whilst not addressing ASC/PC lifespan, well-designed longitudinal studies in regions of differing malaria endemicity would provide the closest approximation of a study of longevity of the humoral response. It is possible that the decay in Abs responses previously observed in some field settings [15],[17] is a reflection of recent acute infection and/or exposure to new antigenic variants resulting in short-lived and primary responses rather than defective MBC and long-lived PC formation. Our studies here would suggest that investigation of serum Ab responses following immunizations or infections should not be restricted to the acute response as this might give misleading results. Rather, the natural history of the serum Ab response should be followed for longer periods during which measurements are taken at different time points. Our data demonstrate a strong correlation between numbers of MSP119-specific IgG ASC in the bone marrow and spleen with concentrations of Ab in sera, suggesting that measurements of serum-Ab in such studies would be a good surrogate for the numbers of PCs.
An important question is whether the low-grade chronic P. chabaudi infection affects the longevity of MSP119-specific PC themselves. One possible underlying mechanism that could explain the reported short-lived Ab responses to some antigens in malaria, is a continuous mobilisation (dislodging) of old PC from their survival niches by continuously generated new ones in the face of a chronic infection. Alternatively, circulation of low affinity Abs and their immune complexes can induce the killing of PC from their survival niches via a recently described mechanism that involve cross-linking of the inhibitory FcγIIR on the surface of PC [69]. We found no differences in the numbers of MSP119-specific ASC between the mice treated with anti-malarial drugs and chronically infected mice suggesting that low-grade persistent infections may not affect the longevity of malaria-specific ASC. Ab isotypes and their respective subclasses are determined by differences in the structures of their Fc-portions, which in turn determine Ab function. Here, there were no differences in the concentrations of MSP119-specific Abs of IgM and the four IgG subclasses suggesting that Ab function is similar between the drug-cured and chronically infected mice. However, new tools and functional assays are required to determine whether low-grade chronic infections affect the fine specificity of anti-MSP119 Abs.
Unlike PC, which are terminally differentiated ASC, MBC are capable of rapid proliferation and differentiation into ASC upon re-exposure to antigen resulting in the amnestic Ab responses that characterize humoral memory. It is generally accepted that MBC are long-lived. In addition to antigen-specific stimulation of MBC, polyclonal activation by TLR ligands or bystander T cell help and their subsequent differentiation into ASC can also contribute to the maintenance of circulating antigen specific Ab [23],[24]. The relative roles of MBC and ASC in sustaining enduring levels of protective Ab after clearance of the inducing antigen are unclear, and remain a subject of investigation. On the basis of reports of some short-lived Ab responses in P. falciparum infected people, it has been suggested that chronic Plasmodium infections may prevent generation of long-lived and/or functional MBC [12],[13],[70]. In this regard, P. yoelii infections in BALB/C mice were found to delete MSP119-specific MBC [11]. One consequence of an antigen-specific immune response is that pre-existing memory cells will differentiate into other phenotypes while others may apoptose upon re-infection/immunisation. A similar phenomenon has been reported in Trypanosoma brucei infections of C57/BL6 and BALB/C mice, where parasite-induced B cell apoptosis resulted in abolishment of pre-established protective anti-parasite and vaccine induced MBC responses. Clearly, more studies are needed to confirm whether the reported abrogation of previously established memory B cell responses by parasitic infections fits a general rule and whether there are differences in various mouse strains [71]. However, here we show that functional MBC are generated after a single P. chabaudi infection and give rise to faster secondary ASC/Ab responses upon re-infection. Furthermore, the memory MSP119 response is composed mainly of IgG isotypes, while the primary response is initially predominated by IgM ASC. Although IgM may augment antimalarial immunity [72],[73], Abs of the IgG isotype are considered to be more superior and better suited for humoral memory than those of IgM, mainly due their longer half-lives, and more specialized effector functions. Consequently, IgG Abs allow for the maintenance of plasma Ab titres by fewer ASC than IgM Abs. In addition, whilst the primary MSP119-specific IgG response was mainly comprised of IgG3, the memory response consisted of the CD4 T-cell dependent IgG2c isotype. Although, general rules about the importance of the respective IgG subclasses in immune protection cannot be made as yet, IgG2c (IgG2a in other mouse strains) has been shown to be more superior in the effector functions requiring complement activation and binding to FcγRs compared to the other IgG subclasses [74]–[76], and human opsonising IgG Abs are thought to play an important role in control of blood-stage parasites [77],[78]. Consistent with these observations, purified hyperimmune IgG2a(c) Abs have been shown to be more efficient at inhibiting invasion of red cells, in vivo, in a P. chabaudi infection of mice than IgG1 [54]. In addition, passively transferred IgG2a(c) Abs from hyper-immune mice were better at transferring immune protection in to naïve mice than Abs of other isotypes [54]–[56]. Collectively, these findings together with the current study, suggest the generation of functional MBC that result into a classical memory response upon re-infection of mice. The ability of these MBC to proliferate and differentiate into MSP119-specific ASC, as demonstrated by the comparable numbers of PC and levels of plasma-Ab between chronically infected and drug-cured mice ex vivo, was not significantly affected by the presence of a chronic infection.
Our study demonstrates that long-lived ASC and MBC can be induced by malaria infection and these cells mount improved humoral responses to a secondary challenge. Furthermore, their maintenance and functionality are not altered by chronic infection. The importance of these findings is highlighted by the fact that B cells and Abs (perhaps in collaboration with ‘parasiticidal’ mediators from macrophages and/or other similar cells of the innate immune system probably activated by T cells) are critical for the elimination of malaria parasites in mouse models [79],[80]. In addition, passive transfer of purified IgG from immune adults into children suffering from malaria had both therapeutic and strong anti-parasitic effects [2],[6] and malaria-antigen specific Abs have been variously associated with reduced incidences of clinical malaria (and/or parasitisation) in longitudinal studies of humans in endemic populations [4],[5],[81],[82], respectively. Further studies in humans are required to determine the persistence of ASC specific for various malarial antigens (as opposed to the single antigen used here) in malaria infected individuals and the conditions under which short- and long-lived Ab responses are generated. Although cellular studies are more complicated to perform in humans, they will have the advantage (over mouse models) of providing data on the generation and maintenance of ASC and MBC specific for P. falciparum parasites in their natural hosts. In addition, the availability of new methods to generate human monoclonal Abs (e.g. cloning of antigen-specific MBC [83]) from multiple malaria-specific clones of MBC in immune individuals coupled with the appropriate functional assays will help distinguish the protective MBC-specificities from those that are not. Such studies will provide insight into which cellular phenotypes and specificities should be targeted for the induction of long-lived anti-malarial Ab based therapeutics.
Female C57BL/6 mice bred in the specific pathogen-free unit at the National Institute for Medical Research (London, U.K.) were used at 6–12 wk of age. They were conventionally housed on sterile bedding, food, and water.
P. chabaudi chabaudi (AS) was routinely injected from frozen stocks. Further infections were initiated by i.p. injection of 105 iRBCs obtained from infected mice before the peak of parasitemia, and the infection monitored by Giemsa-stained thin blood films as previously described [46].
Drug-mediated elimination of chronic P. chabaudi infection was accomplished with chloroquine (CQ) (Sigma, UK), or mefloqine hydrochloride (MQ) (Sigma, UK). A curative regimen of CQ consisted of 25 mg per Kg of mouse body weight in 0.9% saline solution given in 3 doses (at days 30, 32 and 34 of infection) by intraperitoneal injection. This regimen has been shown previously to be effective in removing residual parasites [44]. For MQ, curative treatment consisted of 4 consecutive daily doses (starting at day 30 of infection) at 20 mg/kg of mouse body weight. To investigate the effect of cyclophosphamide (Sigma, UK) on the numbers of MSP119-specific ASC, mice were injected i.p. with 35 mg/kg of mouse body weight of cyclophosphamide daily for 4 days at various times of P. chabaudi infection (depicted in Figure 4A and 5A), as described elsewhere [52],[53]. Single-cell suspensions of spleen and bone marrow were harvested 7 days after initiation of treatment and analysed for MSP119-specific ASC by ELISpot.
All animal work has been conducted according to the relevant British Home Office and international guidelines.
Infected and non-infected mice were given BrdU in drinking water (0.8mg/ml) for 2 or 4 weeks, as previously described [28]. Spleens and bone marrow were harvested and single-cell suspensions created. Erythrocytes were lysed with red blood cell lysis buffer (Sigma) and lymphocytes were enumerated using a Coulter counter. 1×106 cells per well were plated out into 96-well V-bottom plates and incubated with anti-Fc receptor Ab (Fc block, BD), anti-CD138 PE (BD) (or an IgG2a isotype control) and anti-CD19 biotin (BD) and Streptavidin Tricolour (Caltag). Cells were fixed in 2% paraformaldehyde, permeabilised with NP-40 and incubated with anti-BRDU-FITC Ab with DNase (BD). Cells were acquired on a FACSCalibur and analysed using FlowJo (Treestar Inc, OR, USA).
Nucleotide sequences corresponding to the C terminal amino acids 4960 to 5301 of P. chabaudi (AS) MSP1 were re-synthesized for expression in Pichia pastoris, and removal of potential glycoslyation sites, respectively. They were inserted into the pIC9K vector (Invitrogen, San Diego, CA), modified to code for a hexa-His-Tag after the α-factor cleavage site at the N terminus), and protein expression was induced in Pichia pastoris SMD1168, as described previously [84]. The MSP119 protein was purified by binding to a Ni-NTA agarose column (Qiagen, Hilden, Germany) and eluted with 250 mM imidazole, as described [84]. The recombinant protein fragment including the HIS tag has a molecular weight of approximately 14 kDa, and corresponds to the C-terminal MSP-119 fragment of P. falciparum. For clarity and for reference purposes, the P. chabaudi fragment will be referred to as MSP-119 in this paper.
In order to determine whether the anti-malarial drug chloroquine had any direct effects in vivo on the ability of mice to make CD4 T cell and Ab responses, mice were immunised with MSP119 on days 0, 21 and 42 with 50µg, 25µg and 25µg respectively of recombinant MSP119 in Sigma adjuvant (Sigma, UK) according to the manufacturer's protocol. Chloroquine (25mg/kg) was administered in three doses 30, 32 and 34 days after the first immunisation to be as close as possible to the timing of chloroquine treatment in P. chabaudi infections (as described above). Plasma samples and spleens were taken for analysis at day 54.
MSP119- and malaria-specific IgM and IgG Abs were measured as described previously [84],[85] using MSP119 as coating antigen. IgG was revealed with AP-labeled goat anti-mouse IgG, IgG1, IgG2a, IgG2b, IgG3 Abs (Southern Biotechnology Associates) and p-nitrophenyl phosphate. Normal plasma was used as a negative control. Hyper-immune plasma was used as a standard for the IgG and IgM specific ELISAs, and the results were expressed as relative units, as described previously [86]. For the MSP119 Abs measured after immunisation with MSP119 in Supplementary Figure S4, the results are expressed as µg/ml using an anti-mouse Ig ELISA to quantify the amounts of the different isotypes. Anti-mouse Ig (Southern Biotechnology) was used as the coating antibody, purified mouse immunoglobulins of different isotypes as standards (Sigma, UK), and the same AP-labelled antibodies as described above.
Spleen and bone marrow single-cell suspensions were cleared of erythrocytes by a single round of 0.83% NH4Cl treatment and resuspended in Iscove's modified Dulbecco's medium (Sigma) containing 10% fetal calf serum (Sigma), 100 units/ml of penicillin (Sigma), 100 µg/ml of streptomycin (Sigma), 1 mM of L-glutamine (Sigma), 12 mM of Hepes (Sigma) and 5×10−5 M of 2-mercaptoethanol (Invitrogen). MSP119-specific plasma cells were quantitated by a modification of the ELISpot method as described previously [29]. Briefly, nitrocellulose-bottom 96-well Multiscreen HA filtration plates (Millipore Corporation, San Francisco, CA, USA) were coated at 50 µl per well with phosphate-buffered saline (PBS) containing 10 µg/ml of recombinant MSP119 per ml and incubated overnight at 4°C. Additionally, some wells on each plate were coated with a purified goat anti-mouse isotype-specific Ig (CALTAG, San Francisco, CA, USA) as a positive control, and for the determination of total isotype specific ASCs. Plates were washed twice with PBS, and then blocked with 200 µl of Iscove's medium containing 10% fetal calf serum for 1 h at room temperature. Thereafter, blocking media was replaced with 100ml of complete media containing four threefold dilutions of cells and incubated for 5 h at 37°C in a humidified incubator with 6% CO2. Plates were emptied by flicking and washed three times with PBS and then three times with PBS containing 0.1% Tween (PBS-T). For the detection of IgG ASC, a 100µl volume of biotinylated, affinity-purified goat anti-mouse immunoglobulin G (IgG) (CALTAG) diluted 1/1,000 in PBS-T containing 1% fetal calf serum was added to each well and incubated overnight at 4°C. Otherwise, for detection of Ig isotype specific ASC, anti-mouse IgG1, IgG2a, IgG2b, IgG3, and IgM Abs (Caltag Laboratories, Burlingame, CA) were used for the primary detection reagents. The anti-IgG2a Ab used here recognizes the IgG2c isotype expressed in C57Bl/6 mice [87],[88]. The plates were washed four times with PBS-T, 100 µl of alkaline phosphatase-conjugated avidin D (Vector Laboratories) at a concentration of 5 ug/ml in PBS-T–1% fetal calf serum was added, and the mixture was incubated at room temperature for 1h. The plates were washed three times with PBS-T and three times with PBS, and detection carried out by adding 100 µl of substrate. Granular blue spots appeared in 30 min to 1h, and the reaction was terminated by thorough rinsing with tap water. Spots were enumerated with a Immunospot analyser (CTL, Germany).
Using 59Fe in distribution studies, it was demonstrated that 12.6% of the whole bone marrow in located in the two femurs [89], as used in this study. Therefore, we have multiplied the numbers of ASC from the 2 femurs/mouse by a factor of 7.9 to get the total bone marrow ASC response, as described [29].
MSP119-specific memory B cells were measured by a modification of a described limiting dilution method [49]. Splenocytes were cultured for 6 d in flat-bottomed 96-well plates in complete Iscove's medium in a total volume of 200µl in the presence of 1×106 irradiated (1,200 rad) feeder splenocytes, 0.4µg of R595 lipopolysaccharide (Alexis Biochemicals) and 20µl of a culture supernatant from concanavalin A-stimulated C57Bl/6 spleen cells as a source of T and B cell cytokines prepared as described previously [90]. Four-fold dilutions of splenocytes were tested in replicates of 22 wells each. After 6 d of polyclonal activation, cells were washed and transferred to MSP119-antigen-coated 96-well Multiscreen-HA filter plates (Millipore) and ASC ELISpots performed as described above.
CD4 cells were purified from spleens of mice immunised with MSP119 or from unimmunised mice by separation on MACS columns using the manufacturer's protocols (Miltenyi, Germany). For antigen-presenting cells, spleen cells were depleted of T cells using Abs to Thy1.2 and CD4 with rabbit complement (Zymed, UK) as described previously [79]. Responder CD4+ T cells (6×104 per well) were co-cultured with 2×105 antigen presenting cells in 200µl of complete Iscove's medium with 5µg/ml of recombinant MSP119 for 4 days at 37°C, 7% CO2. The proliferative response of the CD4+ T cells was measured by incorporation of 3H-Thymidine as described previously [91].
The frequencies of MSP119-specific memory B cells were determined from the zero-order term of the Poisson distribution using the least squares method of curve fit, and the goodness-of-fit was analysed by linear regression. R2 values of greater than 0.8 were accepted. Differences between groups were tested with a nonparametric test (Mann-Whitney) for significance at 95% confidence intervals. Probabilities of less than 0.05 were considered significant.
|
10.1371/journal.ppat.1002788 | Control of Virulence by Small RNAs in Streptococcus pneumoniae | Small noncoding RNAs (sRNAs) play important roles in gene regulation in both prokaryotes and eukaryotes. Thus far, no sRNA has been assigned a definitive role in virulence in the major human pathogen Streptococcus pneumoniae. Based on the potential coding capacity of intergenic regions, we hypothesized that the pneumococcus produces many sRNAs and that they would play an important role in pathogenesis. We describe the application of whole-genome transcriptional sequencing to systematically identify the sRNAs of Streptococcus pneumoniae. Using this approach, we have identified 89 putative sRNAs, 56 of which are newly identified. Furthermore, using targeted genetic approaches and Tn-seq transposon screening, we demonstrate that many of the identified sRNAs have important global and niche-specific roles in virulence. These data constitute the most comprehensive analysis of pneumococcal sRNAs and provide the first evidence of the extensive roles of sRNAs in pneumococcal pathogenesis.
| Pneumonia is a leading cause of childhood mortality worldwide, resulting in more deaths in young children than any other infectious disease. One of the leading causes of pneumonia is the human pathogen, Streptococcus pneumoniae, the causative agent of over six million infections each year in the United States. Understanding how bacterial pathogens rapidly respond to dynamic host environments is a central aspect of microbial pathogenesis. Accumulating evidence has implicated sRNAs as vital regulators in a number of important cellular processes though few have been implicated in virulence. In our investigations we have applied next-generation sequencing to define the sRNA repertoire of S. pneumoniae. In addition, we utilized both targeted genetic knockouts and transposon mutagenesis to show that a significant portion of these sRNAs play important roles at various stages of pneumococcal pathogenesis. These data represent the first example of sRNAs being involved in pneumococcal pathogenesis and greatly expand the number of sRNAs that play important roles in bacterial pathogenesis.
| Gene regulation and intercellular communication are fundamental aspects of bacterial adaptation to dynamic environments. As such, bacteria have evolved numerous strategies to facilitate tight control of genetic networks in response to diverse extracellular stimuli. Roles have been described for DNA, RNA and protein in gene regulation. Only recently have we begun to appreciate the global roles of sRNAs, particularly in regards to bacterial pathogenesis, as the traditional genetic screens for virulence factors have typically not focused on these small, rarely annotated sRNAs. In recent years there has been a constantly expanding repertoire of sRNAs being identified in a number of bacterial pathogens using both tiling arrays as well as high-throughput sequencing of RNA (RNA-seq). Bioinformatic approaches have also predicted numerous sRNAs in many bacterial pathogens indicating a high prevalence of sRNAs encoded by diverse bacterial species [1], [2]. The increasingly important role of sRNAs in controlling gene expression in bacteria suggests a subset of these molecules may have roles in bacterial virulence [3], [4].
One of the more compelling cases for the role of sRNAs in bacterial pathogenesis arose from studies of Hfq, a chaperone providing stability to sRNA, which substantially advanced our knowledge of the diversity and functional roles of sRNAs in bacteria [5]. Homologs of Hfq are found in diverse species of Gram-negative and Gram-positive bacterial pathogens [6]. Deleting Hfq, which has pleiotropic effects on the stability of several sRNAs, predictably results in numerous phenotypes, mainly consisting of resistance to various environmental stresses, suggesting potential roles in host pathogenesis [6], [7], [8]. There are also numerous examples of sRNAs that function independently of Hfq, even in bacterial species that encode the chaperone. While deletion of Hfq in Listeria has a discernable effect on virulence, its absence does not affect the level of expression of sRNAs [7], [9]. Additionally, deletion of Hfq in S. aureus was found to have no detectable effect on the microbial stress response nor the function of sRNAs [10]. Despite the apparent absence of Hfq, pathogenic streptococci nonetheless encode and express an abundance of sRNAs [11], [12], [13]. In S. pyogenes, the regulatory RNAs RivX and FasX have been implicated in virulence gene regulation and interactions with host cells, respectively [14], [15], [16], [17]. Additionally, a specific sRNA, tracrRNA, serves a central function in the CRISPR system that mediates the silencing of foreign nucleic acid sequences [18]. Regulatory RNAs targeting virulence gene expression in streptococci function both at the transcriptional and translational levels [19]. The interactions of sRNAs are complex, with examples of the same sRNA functioning to both activate and repress target genes by a number of mechanisms [20]. Despite the increase in our knowledge of sRNAs, their contribution to virulence has been much less well established though examples have been demonstrated [3], [21], [22]. In S. pyogenes, deletion of the 4.5S RNA component of the signal recognition particle pathway results in significant attenuation of tissue disease [23]. S. aureus encodes numerous sRNAs, of which the best characterized example is RNAIII, which coordinates the expression of virulence genes [24], [25], [26], [27], [28]. Examination of the transcriptome of L. monocytogenes indicated the presence of several sRNAs implicated in pathogenesis that were not found in closely related non-pathogenic species [9], [29]. Recent reports have also shown sRNAs being involved in pathogenesis in Salmonella and Yersinia [30], [31]. Despite these examples, the contribution of the vast majority of sRNAs to bacterial pathogenesis, particularly in Streptococcus pneumoniae, remains uncharacterized.
S. pneumoniae is a leading cause of childhood mortality worldwide and is a major health concern despite widespread vaccination. The pneumococcus is remarkably adept at colonizing and infecting diverse niches in the human body, readily establishing itself as a commensal in the nasopharynx in over 40% of healthy individuals as well as being a major causative agent of pneumonia, otitis media, sepsis, and meningitis [32], [33]. A number of well characterized virulence genes have tissue-restricted virulence phenotypes, underscoring the diverse pneumococcal arsenal for targeting dissimilar host tissues [34], [35]. One major facet of gene regulation is the set of 13 two-component systems (TCSs) encoded in the pneumococcal genome that control a multitude of gene networks and are implicated in pathogenesis [36]. Included in these networks are sRNAs, some of which are controlled by the CiaR response regulator in the pneumococcus [37]. This phenomenon is not restricted to pneumococci, as other streptococcal species harboring CiaR also are predicted to encode numerous sRNAs, indicating that downstream sRNAs may be an important facet of regulation by this TCS [38]. Of the sRNAs identified thus far in the pneumococcus, none have been found to play a definitive role in the regulation of virulence genes or networks.
A substantial number of sRNAs have been predicted in the sequenced pneumococcal reference strains D39 and TIGR4 using bioinformatics, tiling arrays, and sequencing [11], [12], [39]. However, none have been assigned a role in host pathogenesis. To address this possibility, we undertook a sequencing based approach to identify sRNAs in pneumococcus coupled with both targeted and random gene deletions to ascertain the impact of sRNAs on pneumococcal disease. We present data identifying sRNAs in the pneumococcus by RNA sequencing (RNA-seq). Furthermore, using both transposon mutagenesis (Tn-seq) and targeted deletions, we describe data indicating that many sRNAs play vital roles in progression of infection with unique sRNAs being required for specific tissue tropism. These data provide the first comprehensive analysis of the contribution of sRNAs to pneumococcal pathogenesis and greatly expand the repertoire of sRNAs that play definitive roles in bacterial virulence.
To initially identify sRNAs, we isolated, enriched, and fully sequenced small (<200 nt) transcripts of the TIGR4 strain of pneumococcus. To broaden sRNA capture, we also analyzed mutants in genes encoding the response regulator of three two-component systems (TCS): GRR (TCS03), CbpR (TCS06), and VncR (TCS10) - all of which influence the expression of many transcripts in pneumococcus [40], [41]. TCSs monitor environmental cues to precisely control networks of gene expression; elimination of TCS control could potentially broaden total transcript abundance and thereby capture sRNAs that would otherwise be overlooked. In addition, TCSs have been shown to control the expression of sRNAs both in Gram-negative and Gram-positive bacteria, both as positive and negative regulators [37], [42]. The TCS mutants and TIGR4 were sequenced individually and the data were pooled to generate the composite of sRNAs. For each strain analyzed, coverage exceeded 99.9% with a read depth ranging from 100–400 providing high confidence in sequence quality. The data were next processed to eliminate all sequences within known ORFs to focus on intergenic regions or those running antisense to known ORFs as well as further constraints as detailed in the methods. The position of the identified sRNAs both from our analysis and previous reports were mapped to the TIGR4 genome. The sRNAs were found to be more abundant on the positive strand, though numerous sequences were identified on the negative strand (Figure 1).
We identified 89 putative sRNAs (Table 1). Of these, 56 were novel and the rest have been recently identified by various studies (Table 1, column 11). By BLAST analysis, 85 sRNAs were highly conserved (>90%) amongst pneumococci, 11 were conserved amongst streptococci, and 17 were conserved amongst other Gram-positive bacteria, typically other lactic acid bacteria. Figure 2 outlines the order of analyses applied to the identified sRNAs. Of the 89 sRNAs identified by sequencing, 41 were confirmed for expression and size via Northern blot analysis (Figure S1 in Text S1), an additional 4 were confirmed by qRT-PCR analysis (Table 1), and 10 sRNAs were confirmed by previous studies. Seventeen of the novel sRNAs contained a highly conserved BOX element, making specific detection by Northern blotting or qRT-PCR difficult as the BOX element encompassed a majority of the predicted sRNA sequence in many instances. RNA-seq of the TCS knockouts allowed for the identification of additional sRNAs that were not expressed in the parental TIGR4. An example is shown in Figure S2 in Text S1; the F13 sRNA had high expression in the TCS knockout while being undetectable in the parental TIGR4. In total, there were 24 sRNA candidates that failed to meet the cutoff criteria in all three TIGR4 RNA-seq assemblies but were present in at least one of the TCS knockouts. These data indicate the pneumococcus expresses numerous, highly conserved sRNAs.
We next sought to determine if any of the sRNAs detected by RNA-seq shared any conserved motifs that could facilitate the identification of additional sRNA candidates. Five sequence motifs were conserved across several sets of sRNAs (Figure S3 in Text S1). Each of these motifs was found at additional locations in intergenic regions in the TIGR4 genome, raising the possibility that these motifs could be used to identify additional sRNAs (Table S3 in Text S1). Part of Motif 1 shares homology with a boxA BOX element. The areas around 17 of these motifs had increased signal based on the Illumina reads compared to the nearby flanking region, indicating the possibility of sRNAs being encoded in these domains. Northern Blots using probes against flanking regions immediately outside the conserved motif for these 17 putative sRNAs identified detectable bands between 250–350 bps for each of these new putative sRNAs (Figure S1 in Text S1), indicating that the conserved motifs can be used to predict additional sRNAs. All identified sequences were also analyzed by using Rfam to identify potential RNA families. The R6 and F17 were predicted to be members of the T-box family; F26 and R15 were predicted to be members of Pyr; F27 and F32 were predicted to be members of the TPP and tmRNA families, respectively. Members of these families were found upstream of the class of genes typically regulated by cis-acting riboswitches, namely tRNA synthases and amino acid biosynthesis genes in the case of the T-box, and genes involved in pyrimidine biosynthesis for the Pyr families, indicating these regulatory RNAs may function in a similar manner. The remaining identified sequences had no significant homology to described RNA families.
As indicated in Figure 2, the sRNAs were next analyzed for a role in virulence. Fifteen sRNAs were chosen for further study on the basis of favorable predicted free energy for folding into secondary structures and high levels of expression by Northern blot. These included ΔF6, ΔF7, ΔF20, ΔF22, ΔF24, ΔF25, ΔF32/tmRNA, ΔF41, ΔF42, ΔF43, ΔF44, ΔF48, ΔF55, ΔR6, and ΔR12. These sRNAs were deleted with most having no polar effects on flanking genes (Figure S4 in Text S1; note SP0625 is a pseudogene and partially overlapping with ΔF22). One mutant, ΔF48 resulted in approximately 20-fold upregulation of the upstream gene sp1872. The mutants were assessed for their ability to establish invasive disease in a murine model of infection in which intranasal challenge progresses to pneumonia, sepsis, and meningitis. All mutants caused equivalent levels of bacteremia 24 hours post challenge (data not shown) but further progression of sepsis was attenuated in 8 of the sRNA knockouts tested (p<0.05, Mantel-Cox log rank test): ΔF20, ΔF32/tmRNA, ΔF41, ΔF44, ΔF48, ΔF22, ΔF7, and ΔF25 (Figure 3). These data represent the first report of sRNAs playing a definitive role in pneumococcal pathogenesis whereby deletion of the sRNA results in a significant attenuation of invasive disease.
In order to obtain organ-specific information on the relative contribution of the identified sRNAs to pneumococcal pathogenesis, we next utilized Tn-Seq, an approach that measures the relative fitness of bacterial mutants in different environments (Figure 2, right arm of flowchart). We also included the sequences for the sRNAs identified in TIGR4 by previous studies to obtain the most comprehensive analysis of the contribution of sRNAs to pathogenesis. We analyzed three sites of the host that are vital for the progression of pneumococcal disease- the nasopharynx, lungs, and bloodstream. A comprehensive, large pool of pneumococcal mutants generated by random transposon insertions was administered to these respective host sites and bacteria were harvested subsequent to disease progression. By sequencing the respective mutants in the input and output pools, the relative fitness level of the sRNA mutants was quantified (Table 2, unfiltered data Table S4 in Text S1). A fitness level below 1 means the mutant had decreased fitness whereas a fitness level of 0 indicates that the mutant was attenuated to a degree that no mutants were recovered from the output pools. A number of sRNAs were found to have reduced fitness during colonization of the nasopharynx including F14, F20, F38, F41, F63, and F66. A further 12 sRNAs identified by other groups were also found to have significantly reduced fitness during nasopharyngeal colonization. During lung infection, sRNAs F7 and F32/tmRNA were among the 5 genes identified in our study to be significantly impaired during infection. When the comprehensive list of sRNAs was included, a total of 28 sRNA mutants were predicted to have defects during lung infection. In the sepsis model of infection, a total of 18 sRNA mutants were found to have highly significant reductions in fitness in the bloodstream, including the F25 and F41 that were amongst the knockouts originally tested. These data were in agreement with and further supportive of our data from the targeted genetic knockouts (5 of the 8 attenuated knockouts predicted from RNA-seq were also identified by Tn-seq).
In order to confirm the Tn-seq analysis, individual sRNA knockouts were tested in a competitive index model of infection in which the sRNA mutant was inoculated together with the TIGR4 wild type into the nasopharynx, lung, or blood and differential bacterial density was determined at 24 hours post infection. The capacity of a subset of sRNA mutants predicted by Tn-seq to colonize the nasopharynx, infect the lungs, and replicate in the bloodstream were analyzed in respect to TIGR4 (Figure 4A–C). The ΔF24 strain which was avirulent in sepsis showed a slight decrease in colonization of the nasophaynx (Figure 4B). In addition, ΔR12, which was not significantly attenuated in our initial model of infection, showed dramatic differences in both nasopharyngeal colonization and in the intraperitoneal bacteremia model (Figure 4). In addition, two new sRNA mutants were generated from the Tn-Seq predictions, ΔF5 and ΔF62, both of which displayed defects in their respective host niches of the bloodstream and lung.
RNA-seq coupled with Tn-seq and validated with targeted knockout mutants proved to be a robust method for determining the contribution of sRNAs to pathogenesis. A total of 28 sRNAs in the lung, 26 in the nasopharynx, and 18 in the blood were predicted to have significantly altered fitness in these respective host niches. While a majority of the Tn-seq sRNA mutants attenuated the bacteria, it should be noted that a small number of mutations actually resulted in a fitness benefit in certain host sites (Table 2). In addition, most of the attenuated sRNAs were predicted to be defective in only one host organ, underscoring the contribution of these sRNAs to these distinct environments. These data indicate that sRNAs contribute to pneumococcal pathogenesis both for systemic infections as well as for tissue specific tropisms.
To identify the step in host-bacterial interactions affected by the attenuated sRNA knockouts, the ability of the mutants to adhere to and invade endothelial and nasopharyngeal cell lines was determined. The sRNA mutant F20 had a significant defect in adhesion and invasion of Detroit nasopharyngeal cells (Figure S5 in Text S1), a finding in agreement with the decreased nasopharyngeal fitness (Table 2). A striking defect in adherence to activated endothelial cells was observed in six of the sRNA mutants, while invasion of endothelial cells was only impaired in F20 and F32/tmRNA. These data indicate that many of the attenuated sRNAs have specific defects in interactions with host cells, an underlying cause for attenuation of disease.
We then hypothesized that sRNAs could target either gene networks or individual genes. To investigate global gene expression, we compared the transcriptome of TIGR4 to that of each of the attenuated sRNA mutants via microarray analysis. Several pathways were significantly different upon deletion of the sRNAs (Table S4 in Text S1). The ΔF25, ΔF41, and ΔF44 mutants upregulated a putative ABC transporter encoded by SP1688–1690 that is predicted to be involved in carbohydrate transport. The SP1721–1725 genes, predicted to play roles in sucrose metabolism, were also highly differentially regulated in several of the sRNA mutants. The ΔF32 mutant substantially downregulated several metabolic networks encompassing the lactose transport system and multiple PTS systems. This highlights the potentially pleiotropic effects that the deletion of the sRNAs could have on pneumococcal biology and pathogenesis in the host.
Many sRNAs function at the post-transcriptional level [43], suggesting that there may be important changes in bacterial physiology that potentially could have been missed by global transcriptional analysis. We next sought to determine the effect of the deletion of sRNAs on the global proteome of the pneumococcus. Replicate two-dimensional gels were analyzed for each attenuated sRNA mutant and compared to the parental TIGR4. Every individual protein spot on the gels was then quantified from these duplicate gels to obtain a comprehensive analysis of changes in protein abundance resulting from the deletion of the respective sRNA. The quantitation of the respective spots for each bacterial strain, along with both the predicted pI and molecular weight of the protein, are listed in Table S6 in Text S1. The image of a TIGR4 gel with the individual spot identifications is provided in Figure S6 in Text S1. A number of proteins spots found in increased or decreased abundance are summarized in Figure 5. Deletions in F20 and F32/tmRNA resulted in dramatic alterations in abundance, of 88 and 100 proteins respectively. Of note is that both the ΔF20 and ΔF32 mutants were the only attenuated sRNA mutants to have significant defects in the invasion of endothelial cells, indicating that a subset of these misregulated proteins are important for cell-cell interactions. Analysis by mass spectrometry (Figure 5) indicated that the ΔF20 mutant had decreased abundance of two proteins involved in purine biosynthesis, PurM and PurC, potentially explaining the defect in virulence. The overexpression of the NrdI flavoprotein, essential for the conversion of nucleotides to deoxynucleotides, suggests defects in DNA synthesis and repair [44]. These data indicate that the deletion of sRNAs can have multiple effects on bacterial pathogenesis by influencing numerous putative targets.
Advances in sequencing technologies have driven an explosion in our knowledge of the non-coding genetic repertoire of bacterial species. This study illustrates the first example of a global approach to both sRNA identification and pathogenesis profiling, an amalgamation of RNA-seq and Tn-seq. The RNA-seq tactic identified 89 putative pneumococcal sRNAs, capturing both sRNAs previously detected by sequencing and tiling arrays and many additional previously unknown sRNAs [11], [12], [37], [39]. Use of RNA-seq has certain advantages for the identification of sRNAs. The mean level of sequence coverage was over 100-fold on both the forward and reverse strands, with each sRNA corresponding to a minimum of 10x coverage allowing for high confidence in the data. It should be noted that low abundance sRNAs identified in other studies from a single read will likely be missed by our analysis [39]. Unlike tiling arrays, RNA-seq identifies the origin of transcription. This permits the precise mapping of sRNAs that contain highly repetitive regions, such as the over 100 BOX elements found in intergenic regions of the pneumococcal genome. BOX elements are short AT-rich repeats that are highly transcribed and were also detected in sRNA searches using tiling arrays, though precise locations could not be mapped [11]. Eighteen BOX element containing sRNAs were mapped, a finding particularly important as the Tn-seq analysis implicated a subset of four BOX-element sRNAs in pathogenesis. Although BOX elements have traditionally been thought to be parasitic sequences mobilized by transposases [45], recent evidence supporting their placement in sRNAs indicates that they can form RNA structures with riboswitches [46]. In addition, BOX elements can stimulate expression of downstream genes by increasing the half-lives of the mRNA [47].
Another important aspect of this study was the identification of five novel shared sRNA sequence motifs that were conserved at multiple locations in the pneumococcal genome. Upon closer examination of the sequence read depth in the areas surrounding these motifs, we identified 17 with increased signal compared to the surrounding region. All 17 of these predicted sRNAs were subsequently validated by expression analysis underscoring the robustness of the predictions. While members of the T-box, Pyr, TPP, and tmRNA sRNA families described in other bacteria were also found in pneumococcus, a majority of the predicted pneumococcal sRNAs could not be assigned to a functional family. These data indicate that the pneumococcus is a rich source of new motifs that can expand sRNA prediction algorithms in Gram-positive bacteria.
Although numerous sRNAs have been identified in the pneumococcus, there have been no sRNAs implicated in pathogenesis and more broadly, there have been no attempts to apply transposon-mediated mutagenesis to determine the role of sRNAs in bacterial virulence in specific host tissues. This study represents the first use of transposon-mediated mutagenesis to address the global role of sRNAs in discrete host tissues during disease. Using a comprehensive list of sRNAs identified in this study together with those found by others, we identified a number of sRNAs that played distinct roles in pathogenesis in the nasophaynx, the lung, or the bloodstream. The lungs provided the most comprehensive analysis of the contribution of sRNAs to virulence, since bottleneck constraints in the nasophaynx and the blood imposed by a limitation of bacterial binding sites and clearance by the spleen, respectively, may have impaired detection in these sites. A number of sRNAs had no inserts in the Tn-seq deletion library (n.i. in Table S4 in Text S1) and it is tempting to speculate that there is a selective pressure against the loss of these sRNAs; however this observation could be random due to their small size. All three body sites had a distinct list of sRNA candidates that were involved in pathogenesis. The Tn-seq analysis proved to be robust, as mutants predicted to be attenuated in their respective host niches were confirmed in in vivo competition experiments pitting each sRNA mutant individually against wild type (Figure 4). Thus the multi-organ Tn-Seq approach captured this diversity as exemplified by R12 that did not have a significant virulence defect in overall survival in our initial studies but was attenuated both during colonization of the nasophaynx and in the blood following intraperitoneal infection. The Tn-seq analysis also provides insight into the organ-specific defects of the sRNAs found to have reduced virulence in Figure 3. Both the ΔF41 and ΔF25 strain had greatly reduced fitness in the blood, in agreement with their inability to progress to sepsis. The ΔF7 and ΔF32/tmRNA strains were both defective in the lung infection, indicating that this might be the most crucial site for clearance of these mutants. This comprehensive analysis of the contribution of all the identified sRNAs to pneumococcal pathogenesis in discrete host sites can provide a framework for future investigations elucidating the precise functions of these sRNAs. These data add to the growing understanding of the contribution of sRNA in the virulence of bacterial pathogens [3].
The sRNA mutants displaying defects in virulence exhibited a number of characteristics that could potentially explain an inability to cause disease. Several of the attenuated sRNA mutants had defects in adhesion and invasion of nasopharyngeal or endothelial cells, capabilities important to the progression of invasive disease. ΔF20 and ΔF32/tmRNA showed decreased adhesion/invasion of nasopharyngeal or endothelial cells, respectively, in concert with Tn-seq and competitive index data indicating lack of fitness in the nasopharynx and lung. F32 encodes a tmRNA and these have been implicated in the pathogenesis of other bacteria [48], [49]. The central role of tmRNA in the rescue of ribosomes on stalled mRNA as well as targeting defective mRNA for degradation, is consistent with the strong defect in pathogenesis observed in the ΔF32 strain [50]. In the case of the ΔF20 mutant, proteomic analysis indicated proteins responsible for purine metabolism were strongly down regulated whereas DNA synthesis and repair pathways were greatly increased. Thus deletion of F20 had pleiotropic effects on DNA metabolism that could explain attenuation of the mutant. Taken together, these data provide compelling evidence that sRNAs play important roles in virulence, that their affects can arise at several levels of control of virulence gene/protein expression, and that these roles can be restricted to specific host tissues.
Our study expanded the search for sRNAs and their role in gene regulation to three mutants in TCSs. Control over gene networks by TCSs is typically mediated by a direct interaction of the response regulator with a target sequence shared by many genes dispersed over a genome. However, TCSs have also been found to control the expression of sRNAs in pneumococcus and other bacteria [37], [51]. For example, control of porin expression in E. coli involves multiple sRNAs that exert posttranscriptional control over the targets of TCSs [42]. The prospect of sRNA functioning as an intermediary, finely tuning the control of and expanding the regulatory scope by a TCS, would allow for another layer of control for more precise regulation. Our observation that the abundance of sRNAs was altered when each of the three TCSs were disrupted is consistent with TCSs acting through sRNAs to broadly control gene expression. This is further supported by the observed alterations of the global transcriptome as well as the abundance of multiple protein targets upon deletion of an individual sRNA (Tables S5 and S6 in Text S1, Figure 5). These data suggest that the impact of sRNAs on multiple aspects of pneumococcal biology and pathogenesis could potentially be exerted by an additional layer of posttranscriptional control over the gene networks controlled by TCSs.
The widespread utilization of RNA-mediated regulation of diverse processes has a number of potential advantages for bacteria [52]. Protein regulators incur greater metabolic costs to the cell, being encoded by larger segments of the genome and requiring translation. In contrast, sRNAs do not require translation and occupy a very limited amount of the genome. The additional layer of regulation conferred by sRNAs may also allow for more precise control of gene expression, as evidenced by the fact that sRNAs can have multiple targets as well as the fact that multiple sRNAs can regulate a single target under different conditions [4], [53]. Additionally, sRNAs can have dramatically different half-lives in the cell, ranging from under 2 minutes to greater than 30 minutes [54]. Such differences in stability could potentially mediate the duration of control mediated by sRNAs. The challenging task that remains following the identification and characterization of sRNAs in pathogenesis is assigning discrete functional roles to these molecules. We have shown the feasibility of applying Tn-seq to identify changes in bacterial fitness in response to deletion of the corresponding sRNA in various host tissues. The feasibility of this approach to investigate the gene networks and functional roles of sRNAs suggest the combination of RNA-seq and Tn-seq will be a unique and powerful tool for future investigations of the precise functional roles of these sRNAs in the pneumococcus.
The S. pneumoniae strains used are listed in Table S1 in Text S1. All experiments were conducted in the sequenced TIGR4 strain [55]. Cultures were grown overnight on tryptic soy agar plates supplemented with 3% sheep blood and were transferred to a defined semisynthetic casein liquid medium supplemented with 0.5% yeast extract (i.e., C+Y) [56].
To initially identify sRNAs in Streptococcus pneumoniae, we designed a method to isolate, enrich, and fully sequence small (<200 nt) transcripts of the TIGR4 strain of pneumococcus. Cultures were grown in triplicate in C+Y (200 mL) until an OD620 of 0.5 was reached, corresponding to mid log phase growth. Bacteria were diluted (1∶2) in RNAProtect stabilization buffer (Qiagen) and centrifuged; the resulting bacterial pellets were then frozen at −80°C. The pellets were thawed and resuspended in Lysis Buffer Mirvana miRNA Isolation Kit (Applied Biosystems). To each sample, 200 µL of 0.1 mm glass beads (Sigma) were added before they were lysed using a mini-beadbeater. Samples were incubated for 10 minutes at 70°C and subsequently processed through a Qiashredder column (Qiagen). sRNA was purified using organic extraction and sRNA enrichment procedures as described in the Mirvana protocol. Purified sRNA was DNAse-treated by using Turbo DNAse (Applied Biosystems) according to the manufacturer's instructions. Purified sRNA was prepared for sequencing by using the Small RNA Sample Prep kit (Illumina). Details about the cluster generation, sequencing, and Northern Blot confirmation are provided in the Supplementary Materials section.
Detection of biologically meaningful sRNA regions was based on the assumption that sequence reads are enriched in such regions. The sequence reads were first mapped to the T4 genome using the program GMAP recursively by quality based trimming. Then the coverage information for both strands was calculated based on high quality matches. When a read mapped to multiple positions on the genome, the highest quality match was selected. For each intergenic region and anti-sense coding region of size greater than 150 bases, a simple method was used to identify a potential read enriched region (peak). Due to the degradation of the sample mRNA, these reads were mapped all over the genome and it was necessary to remove those background signals. Signal noise was not uniformly distributed along the genome, so a baseline detection algorithm (linear interpretation of minimum value) was used. After baseline correction, a cut off value of 20 was utilized to identify potential peaks such that any consecutive region with minimum coverage of 20 is considered as a potential peak. The peak detection methods were applied on both strands separately.
These detected peaks were subjected to further biological constraints. First, a promoter region would be expected on the upstream sequence. We used the Prokaryotic promoter prediction program (http://bioinformatics.biol.rug.nl/websoftware/ppp/ppp_start.php) to search for promoters. Second, a rho-independent terminator would be expected downstream of the sequence. We used the TransTermHP (http://transterm.cbcb.umd.edu/index.php) predicted terminator for the T4 genome. For each potential peak, the promoter must appear between −75 and 20 bases around peak starting position and the terminator must present between −20 and 75 bases around the peak ending position. Those two criteria remove 83% to 98% of potential peaks. These criteria are similar to those used previously to identify candidate sRNAs using RNA-seq data [57], [58].
Mutants were made by using PCR-based overlap extension [59]. Briefly, regions upstream and downstream of the target region were PCR-amplified and spliced into an antibiotic resistance cassette. The final PCR product was transformed into the pneumococcus by conventional methods, replacing the targeted region with the antibiotic resistance cassette. To confirm transformation, primers outside of the transformed region were used for PCR and subsequent region sequencing. The lists of mutants made and oligonucleotides used are included in Tables S1 and S2 in Text S1, respectively.
Bacterial RNA was harvested from mid–log phase cultures (OD600 = 0.4) grown in C+Y by using the Qiagen RNAeasy minikit. Microarray experiments were performed as described previously [60]. Briefly, whole-genome S. pneumoniae version 8.0 cDNA microarrays were obtained from the Pathogen Functional Genomics Resource Center (PFGRC). Microarray experiments were performed by the Functional Genomics laboratory, Hartwell Center for Bioinformatics and Biotechnology, St. Jude Children's Research Hospital using standard protocols provided by PFGRC (http://pfgrc.tigr.org/protocols.shtml) as previously described [61].
Secondary structures were predicted using mfold to obtain ΔG values [62]. The MEME program was used to perform the MOTIF search. The meme web server was used with default options although negative training sequences were used to delineate true motifs from the background sequence patterns of S. pneumoniae.
Proteomic profiling was performed by Kendrick Laboratories Inc (Madison, WI). Two-dimensional electrophoresis was performed using the carrier ampholine method of isoelectric focusing. Isoelectric focusing was carried out in glass tubes of inner diameter 2.3 mm using 2% pH 4–8 mix Servalytes (Serva, Heidelberg Germany) for 9,600 volt-hrs. Fifty ng of an IEF internal standard, tropomyosin, was added to each sample prior to loading. After equilibrium in SDS sample buffer (10% glycerol, 50 mM dithiothreitol, 2.3% SDS and 0.0625 M tris, pH 6.8), each tube gel was sealed to the top of a stacking gel that overlays a 10% acrylamide slab gel (0.75 mm thick). SDS slab gel electrophoresis was carried out for about 4 hrs at 15 mA/gel. The following proteins (Sigma Chemical Co, St. Louis, MO) were added as molecular weight markers: myosin (220,000), phosphorylase A (94,000), catalase (60,000), actin (43,000), carbonic anhydrase (29,000), and lysozyme (14,000). The gels were dried between cellophane sheets with the acid end to the left. Duplicate gels were obtained from each sample and were scanned with a laser densitometer (Model PDSI, Molecular Dynamics Inc, Sunnyvale, CA). The scanner was checked for linearity prior to scanning with a calibrated Neutral Density Filter Set (Melles Griot, Irvine, CA). The images were analyzed using Progenesis Same Spots software (version 4.5, 2011, Nonlinear Dynamics, Durham, NC) and Progenesis PG240 software (version 2006, Nonlinear Dynamics, Durham, NC). The general method of computerized analysis for these pairs included image warping followed by spot finding, background subtraction (average on boundary), matching, and quantification in conjunction with detailed manual checking.
Spot % is equal to spot integrated density above background (volume) expressed as a percentage of total density above background of all spots measured. Difference is defined as fold-change of spot percentages. For example, if corresponding protein spots from different samples (e.g. mutant versus wild type) have the same spot %, the difference field will show 1.0; if the spot % from the mutant is twice as large as wild type, the difference field will display 2.0 indicating 2-fold up regulation. If the spot % from the mutant has a value half as large, the difference field will display – 2.0 indicating a 2-fold down regulation.
A subset of proteins were chosen for further analysis. Protein spots were excised from duplate Coomassie stained gels. The protein sample was digested with trypsin and mass spectrometric analysis was performed using an Orbitrap Velos Mass Spectrometer from Thermo Electron (San Jose, CA). This instrument employs electrospray ionization (ESI), in conjunction with an Orbitrap mass analyzer. The digest was introduced into the instrument via on line chromatography using reverse phase (C18) ultra-high pressure liquid chromatography on the nanoAcquity (Waters, MA). The column used was a New Objective C18 with an I.D. of 75 um and bed length of 10 cm. The particle size was 2.7 um. Peptides were then gradient eluted into the linear ion trap through a non-coated spray needle with voltage applied to the liquid by increasing the concentration of acetonitrile. Data acquisition involved acquiring the peptide mass (MS) spectra followed by fragmentation of the peptide to produce MS/MS spectra that provides information about the peptide sequence. Database searches were performed using raw files in combination with the Mascot search engine. Protein/peptide assignments are made on the basis of MS/MS spectra.
Detroit nasopharyngeal cells and rBCEC6 rat brain capillary endothelial cells were grown in 24-well plates at 37°C in 5% CO2 to 80% confluency and activated with TNF-α (10 ng/mL) for 2 hours [40]. Pneumococcal cultures were grown until the OD620 was 0.5, washed with PBS, and then added to eukaryotic cells (1×107 cfu/well). Three wells were used for each mutant or TIGR4R and the assays were repeated a minimum of 3 times. For adherence assays, cells were incubated 30 minutes with bacteria, a time chosen to minimize internalization of adherent bacteria. After washing 3x in dPBS, the cells were released from the plate with trypsin but not lysed before plating on blood agar plates. Colonies grown overnight were counted as bacteria adherent to cells. For invasion assays, cells were incubated with the bacteria for 2 hours, washed 3 times in dPBS, and subjected to 1 hour of treatment with penicillin (10 µg/mL) and gentamycin (200 µg/mL). The cells were washed, trypsinized, and lysed with 0.025% Triton X-100. The lysates were then incubated overnight on blood agar plates and the resulting colonies were counted.
All mice were maintained in BSL2 facilities, and all experiments were done while the mice were under inhaled isoflurane (2.5%) anesthesia. For survival studies, bacteria were introduced by intranasal administration of 107 CFU of bacteria in PBS (25 µL), a model which effectively recapitulates the progression of disease from nasopharyngeal colonization, to pneumonia, and finally to the development of sepsis and meningitis [63]. A minimum of 10 mice per group was used in the studies from at least two independent experiments. Mice were monitored daily for signs of infection, and differences in time-to-death among the mice were compared via Mantel-Cox log rank test. For the competitive index studies, equivalent CFUs of the parental TIGR4 and the respective mutants were administered to the mice. For nasopharyngeal colonization, bacteria were administered at a dose of 107 CFU in 25 µL PBS [40]. Bacteria were administered intratracheally at a dose of 105 CFU in 100 µL PBS to model lung infection [40]. For the sepsis model, 2×103 CFU in 100 µL PBS was administered by intraperitoneal injection [63]. Tissues and blood were collected from all animals 24 hours following infection. For lung collection, mice were perfused with saline prior to organ collection to remove contaminating blood from the lung which was then homogenized. The parental TIGR4 and sRNA mutants were enumerated by serial dilution and counting on TSA blood agar plates with and without erythromycin. The CFU counts were then utilized to calculate competitive indexes [64] (1 = equivalent numbers of mutant recovered to TIGR4).
Tn-seq, both the experimental procedure as well as data analysis, was performed essentially as described previously [65], [66]. For two time points (t1 and t2) the number of reads at each genome location was determined by massively parallel sequencing on an Illumina Genome Analyzer II. Mice were challenged with transposon mutant libraries administered directly to the nasopharynx, lungs, or to the bloodstream. On average, 250 reads were mapped per insertion/time point. Since insertions with a very low number of reads that slightly fluctuate over time can influence the data disproportionately, only insertions with fifteen or more reads at t1 are included in the analyses. For each insertion, fitness Wi, is calculated by comparing the fold expansion of the mutant relative to the rest of the population with the following equation [67]:In which Ni(t1) and Ni(t2) are the frequency of the mutant in the population at the start and at the end of the experiment, respectively, and d (expansion factor) represents the growth of the bacterial population during library selection. Details regarding the data analysis and methodology are included in the Supplementary material.
All experiments involving animals were performed with prior approval of and in accordance with guidelines of the St. Jude Institutional Animal Care and Use Committee (Protocol #250). The St Jude laboratory animal facilities have been fully accredited by the American Association for Accreditation of Laboratory Animal Care. Laboratory animals are maintained in accordance with the applicable portions of the Animal Welfare Act and the guidelines prescribed in the DHHS publication, Guide for the Care and Use of Laboratory Animals.
|
10.1371/journal.pgen.1000324 | ATR and Chk1 Suppress a Caspase-3–Dependent Apoptotic Response Following DNA Replication Stress | The related PIK-like kinases Ataxia-Telangiectasia Mutated (ATM) and ATM- and Rad3-related (ATR) play major roles in the regulation of cellular responses to DNA damage or replication stress. The pro-apoptotic role of ATM and p53 in response to ionizing radiation (IR) has been widely investigated. Much less is known about the control of apoptosis following DNA replication stress. Recent work indicates that Chk1, the downstream phosphorylation target of ATR, protects cells from apoptosis induced by DNA replication inhibitors as well as IR. The aim of the work reported here was to determine the roles of ATM- and ATR-protein kinase cascades in the control of apoptosis following replication stress and the relationship between Chk1-suppressed apoptotic pathways responding to replication stress or IR. ATM and ATR/Chk1 signalling pathways were manipulated using siRNA-mediated depletions or specific inhibitors in two tumour cell lines or fibroblasts derived from patients with inherited mutations. We show that depletion of ATM or its downstream phosphorylation targets, NBS1 and BID, has relatively little effect on apoptosis induced by DNA replication inhibitors, while ATR or Chk1 depletion strongly enhances cell death induced by such agents in all cells tested. Furthermore, early events occurring after the disruption of DNA replication (accumulation of RPA foci and RPA34 hyperphosphorylation) in ATR- or Chk1-depleted cells committed to apoptosis are not detected in ATM-depleted cells. Unlike the Chk1-suppressed pathway responding to IR, the replication stress-triggered apoptotic pathway did not require ATM and is characterized by activation of caspase 3 in both p53-proficient and -deficient cells. Taken together, our results show that the ATR-Chk1 signalling pathway plays a major role in the regulation of death in response to DNA replication stress and that the Chk1-suppressed pathway protecting cells from replication stress is clearly distinguishable from that protecting cells from IR.
| The integrity of the genetic information in cells is protected by elaborate mechanisms that ensure that an accurate DNA copy is passed from generation to generation. These mechanisms repair errors in DNA sequence or stop growth if DNA structure is compromised. However, if the level of DNA damage is too severe, cells may also respond by inducing death rather than attempt repair. Relatively little is known about how cells decide whether to repair damage or commit to death. The purpose of our work was to identify genes that control this decision-making process while cells are duplicating DNA. We show that two genes play a major role in this process; however, our work also suggests considerable complexity in this death response as different death pathways are triggered in response to different forms of DNA damage. Since DNA replication inhibitors are used widely in the treatment of cancer, our work may enable us to more effectively kill cancer cells in treatment protocols employing these agents.
| Cells respond to DNA damage by triggering cell cycle arrest, DNA repair, or death. The related PIK-like kinases ATM (Ataxia-Telangiectasia Mutated) and ATR (ATM- and Rad3-related) are major coordinators of this damage response [1]. ATM is central to the DNA double-strand break (DSB) response. It delays DNA synthesis and the onset of mitosis following DSB induction by agents such as ionizing radiation (IR) through a complex signalling cascade that includes p53, Chk2 and NBS1 as phosphorylation targets [2]–[4]. This signalling cascade also plays a major role in the onset of apoptosis following IR through the p53-mediated transcriptional activation of pro-apoptotic proteins such as BAX and PUMA [5]–[7]. However cells deficient in ATM are only partially defective in the induction of apoptosis by IR while p53 deficient cells show a more complete resistance [8],[9]. These observations indicate that both ATM-dependent and independent pathways regulate the induction of apoptosis by IR. Chk2 may be particularly important for the ATM-independent pathway as mouse cells with knockouts of both Chk2 and ATM show levels of apoptosis similar to those found in p53−/− cells [9].
ATR and its downstream phosphorylation target, Chk1, are generally activated in response to UV and agents that stall DNA replication forks [10],[11]. Activated Chk1 coordinates many of the cellular responses to replication fork stress. More specifically, it prevents the inappropriate firing of late replication origins, the abandonment of replication forks, and premature chromosome condensation following disruption of replication [12]–[15]. In contrast to the proapoptotic role of the ATM-mediated protein kinase cascade in the response to IR, Chk1 has an anti-apoptotic effect in the cellular response to replication inhibitors [13], [16]–[18] as well as IR [19]. SiRNA mediated ablation of Chk1 (but not Chk2) causes cells arrested in S-phase by a range of replication inhibitors to undergo apoptosis. This death response is p53 independent, but cells that lack both Chk1 and p21 show a more robust death response and reduced cell survival [17]. Thus the Chk1 pathway plays a key role in protecting S-phase cells from apoptosis during replication stress and p21 mediates this role, presumably by preventing entry into S-phase. Intriguingly depletion of the replication helicase cofactor Cdc45 that plays an essential role in DNA replication origin firing and fork elongation protects cells lacking Chk1 from undergoing apoptosis, suggesting that the role of Chk1 in controlling origin firing and maintaining fork integrity is key to its anti-apoptotic effect [20].
A role for Chk1 in the suppression of apoptosis in response to IR was revealed in a zebrafish embryo-based screen [19]. The novel death pathway triggered in p53 mutant embryos in the absence of Chk1 required ATM, ATR and caspase 2 but not other caspases. It was further shown that this response was not limited to Zebrafish as IR triggered a caspase-2 dependent apoptotic response in cultured p53-deficient human tumour cells treated with a Chk1 inhibitor. Taken together these reports establish a major role for Chk1 in the protection of cells from apoptosis in response to a wide range of DNA damage.
While the role of the ATM signalling cascade in the induction of apoptosis following IR is well established, relatively little is known concerning the contribution of this signalling pathway in response to replication fork stress. DNA replication inhibitors trigger a rapid ATM autophosphorylation and ATM-dependent phosphorylations of Chk2, NBS1 and p53 [21]. More recent work has implicated BID as a downstream ATM phosphorylation target and a suppressor of apoptosis in response to DNA replication inhibitors [22]. In the work reported here we compared the effects of ATM, NBS1 or BID deficiency on apoptosis with those obtained following siRNA-mediated depletion of ATR or Chk1. We show that depletion or loss of ATM, NBS1 or BID has little or no significant effect on the induction of apoptosis in two human tumour cell lines or immortalized human fibroblasts treated with DNA replication inhibitors, even though depletion of ATR or Chk1 in these cells led to high levels of cell death. Furthermore, unlike the Chk1 suppressed pathway responding to IR identified in Zebrafish embryos, the pathway regulated by Chk1 in response to replication inhibitors was triggered in both p53 proficient and deficient cell lines, did not require ATM or ATR, and was primarily characterized by caspase-3 activation.
To determine the role of ATM or ATR relative to Chk1 in the regulation of apoptosis in response to DNA replication stress, HCT116 cells were treated with control, ATM, ATR, or Chk1 siRNAs (Figure 1A) for 24 hours before treatment with thymidine or HU. After 24 or 48 hour treatment with replication inhibitors, cells were fixed, stained with PI, and analysed for DNA content by flow cytometry or assayed for Annexin V binding. Cells treated for 24 hours with thymidine or HU accumulated in S-phase, however there was no significant increase in the level of cells with a subG1 DNA content in cultures treated with any of the siRNAs.
After a 48 hour exposure to thymidine or HU, cells treated with the control siRNA continued to accumulate in S as well as G2 but there was only a small increase in cells with a subG1 DNA content relative to controls (Figure 1B & C). HCT116 cells depleted of Chk1 or ATM showed a cell cycle distribution similar to that found for cultures treated with the control siRNA while there was a small increase in the fraction of cells showing a subG1 DNA content in cultures depleted of ATR. When Chk1 or ATR depleted cells were treated with thymidine or HU, fewer S or G2 DNA phase cells were detected relative to the cultures treated with the control siRNA while a markedly higher fraction of cells (40 to 50%) with a subG1 DNA content was evident (Figure 1B & C). ATM depleted HCT116 cells showed a somewhat different response following treatment with thymidine or HU. The fraction of S phase cells increased in these cultures like cultures treated with control siRNA (Figure 1B). However after thymidine treatment, ATM-depleted cells arrested earlier in S-phase while HU treated cells showed a higher frequency of cells with a late S-phase DNA content. There was an increase in the level of cells with a subG1 DNA content relative to control cultures after a 48 hour exposure to the inhibitors. This reached significance for HU treated cells but not those treated with thymidine. Notably the fraction of subG1 cells was consistently lower in ATM depleted cells relative to Chk1 or ATR depleted cultures (Figure 1B & C). Similarly, there was little effect on the fraction of cells with a subG1 DNA content in HCT116 cells treated with the ATM inhibitor KU-55933 [23] following a 48 hour exposure to thymidine (Figure S1). p53 defective SW480 cells depleted of Chk1 or ATR also had a significantly higher level of subG1 cells than those depleted of ATM following a 48 hour treatment with HU (Figure S2).
Analysis of Annexin V+ cells gave similar results (Figure 1D). HCT116 cultures depleted of either Chk1 or ATR showed a significant increase in the level of Annexin V+ cells relative to those treated with the control siRNA following either thymidine or HU treatment. The fraction of Annexin V+ cells in ATM depleted cultures exposed to thymidine or HU increased relative to control siRNA treated cells but this did not reach significance for either replication inhibitor.
To determine whether the depletions of these checkpoint proteins affected apoptosis through related pathways, HCT116 cells were treated with combinations of siRNAs for the checkpoint proteins (Figure 2A). Apoptotic responses after treatment with thymidine were measured by the AnnexinV assay. In cultures depleted of both Chk1 and ATR, the increased level of AnnexinV+ cells was not significantly different from that produced by depletion of either protein alone (Figure 2B). When HCT116 cells were depleted of both Chk1 and ATM there was no significant difference in the fraction of AnnexinV+ cells relative to cultures depleted of Chk1 alone, although the level of apoptotic cells was significantly higher than that found in cells depleted of ATM alone (Figure 2C).
We further examined the induction of apoptosis following treatment with replication inhibitors in an immortalized human fibroblast line derived from an AT patient (pEBS) and a derivative of this line corrected for the ATM defect (YZ5) [24]. In these cultures, the level of pEBS fibroblasts with a subG1 DNA content was not significantly different from that found in the ATM corrected cells following treatment with thymidine (Figure 2D). However, cultures of both pEBS and YZ5 depleted of Chk1 (Figure 2E) showed a significantly higher level of cells with a subG1 DNA content relative to controls following thymidine treatment (Figure 2D). Furthermore there was no significant difference in fraction of cells with a subG1 DNA content in the two lines. Thus our results suggest that Chk1 and ATR regulate apoptosis in response to replication stress through a common pathway while ATM does not play a significant role in this response.
RPA foci appear as an early event in Chk1 depleted cells in response to replication inhibitors [20]. We next determined whether cells depleted of ATR or ATM showed induction of such foci. HCT116 cells were treated with control, Chk1, ATR, or ATM siRNAs for 24 hours before treatment with thymidine (Figure 3A). After 24 hours cells were fixed and stained for RPA. Following depletion of Chk1 or ATR, RPA foci accumulated (>10 foci/cell) in up to 50 to 60% of cells following treatment with thymidine (Figure 3B & C). In contrast cultures depleted of ATM or treated with the ATM inhibitor KU-55933 showed a significantly lower percentage of cells accumulating these foci. In addition hyperphosphorylation of RPA 34 was evident in Chk1 or ATR depleted HCT116 cells after thymidine treatment but not cells depleted of ATM (Figure 3D). Since ATM has been reported to contribute to the phosphorylation of RPA34 following DNA damage [25], this decrease in the level of phosphorylation could simply be due to a decrease in ATM kinase activity. However, RPA34 hyperphosphorylation was also evident in HCT116 cells depleted of both Chk1 and ATM (Figure 3E). Furthermore AT5 fibroblasts (derived from AT patients) depleted of Chk1 showed hyperphosphorylation of RPA34 as early as six hours after thymidine treatment (Figure 3F) demonstrating that cells are still capable of RPA34 phosphorylation when ATM function is compromised. Thus the early events detected in Chk1- or ATR-depleted cells treated with replication inhibitors are not evident in ATM depleted cells.
We previously reported that caspase 3 was activated in HCT116 cells depleted of Chk1 following treatment with replication inhibitors [17]. To determine whether caspase 3 was activated in cells depleted of ATM, HCT116 cells treated with control, Chk1, ATR or ATM siRNAs were exposed to thymidine for 24 or 48 hours. Activated caspase 3 was assayed by Western blotting using cell free lysates prepared from these cultures. This analysis revealed a strong increase in the level of the activated (cleaved) form of caspase 3 in Chk1 or ATR depleted HCT116 cells exposed to thymidine (Figure 4A). A lower level of the activated caspase 3 was detected in HCT116 cells treated with control or ATM siRNAs following exposure to thymidine consistent with the lower level of apoptosis found in such cells.
Recently it was reported that p53 deficient cells treated with a Chk1 inhibitor or siRNA showed cleavage of caspase 2 but not caspase 3 following exposure to IR while p53+/+ HCT116 cells predominantly showed cleavage of caspase 3 [19]. To determine the effect of p53 status on caspase 2 and 3 cleavage in Chk1 depleted cells treated with thymidine, the cleaved forms of these caspases were analysed in HCT116 p53−/− cells treated with control or Chk1 siRNAs. In agreement with the previous report, cleaved caspase 2 was detected in Chk1 depleted HCT116 p53−/− cells exposed to 10 Gy IR while cleaved caspase 3 was not evident (Figure 4A & B). Following a 48 hour exposure of Chk1 depleted HCT116 p53−/− cells to thymidine, there was little change in the level of the cleaved caspase 2 relative to cells treated with the control siRNA while more robust levels of cleaved caspase 3 were evident (Figure 4B). Similarly Chk1 depleted SW480 cells (that are defective in p53 function) showed a weak increase in the level of cleaved caspase 2 while caspase 3 cleavage was clearly induced (Figure 4C). Cleaved caspase 2 was not detected in Chk1 depleted p53+/+ HCT116 cells following thymidine treatment (Figure 4A).
To determine whether the induction of apoptosis was dependent upon caspase 3 activation following thymidine exposure, Chk1 depleted HCT116 or SW480 cells exposed to thymidine were treated with the caspase 3 inhibitor II (Z-DEVD-FMK). In such cultures the accumulation of cells with a subG1 content (Figure 4D) or Annexin V+ cells (Figure 4E) was markedly reduced. Taken together these results indicate that a caspase 3 dependent pathway is activated in Chk1 depleted cells exposed to thymidine. In contrast to the response of such cells to IR, caspase 3 activation is not dependent upon p53 status although a low level of caspase 2 cleavage can be detected in the p53 deficient cells.
Recent work has shown that mice carrying a carboxyl terminus deletion of NBS1 are defective in apoptosis in many tissues and in response to IR [9]. To investigate the contribution of NBS1 to apoptosis in response to DNA replication stress, we determined the apoptotic response of HCT116 cultures depleted of NBS1 or immortalized human fibroblasts obtained from Nijmegen breakage syndrome patients to DNA replication inhibitors. HCT116 cultures depleted of NBS1 (Figure 5A) and treated with thymidine showed a slightly reduced level of cells in S- and G2-phases relative to cultures treated with the control siRNA (Figure S3). The fraction of apoptotic cells in thymidine treated cultures reach significance (p = 0.049) when measured by the Annexin V assay, but not in the assay of subG1 cells (Figure 5B). The response of NBS1 depleted cells in either assay was not as robust as that seen with Chk1 depleted cells. Co-depletion of NBS1 and Chk1 did not produce any significant changes in the level of apoptotic cells relative to cells depleted of Chk1 alone. When HCT116 cells depleted of NBS1 were treated with HU, the level of S and G2-phase cells was not greatly affected (Figure S3) and the increase in Annexin V+ or subG1 cells did not reach significance (Figure 5B). The level of Annexin V+ cells in HCT116 cultures depleted of both NBS1 and Chk1 was similar to that of cultures depleted of Chk1 alone.
NBS1−/− fibroblasts (NBS1-LB1) obtained from Nijmegen breakage syndrome patients and fibroblasts corrected for the defect (p95wt, [4] were next examined for their response to replication inhibitors. Thymidine had no significant effect on the level of cells with a subG1 DNA content in either mutant or corrected fibroblast lines treated with the control siRNA while HU produced a small increase in both cell types (Figure 5C & D). Chk1 depletion of the corrected fibroblasts resulted in a ∼two-fold increase in cells with a subG1 DNA content. Intriguingly there was a ∼five-fold increase in the fraction of subG1 cells in Chk1 depleted NBS1−/− fibroblasts (Figure 5C & D). The level of cells with a subG1 DNA content was significantly increased in mutant and corrected fibroblasts treated with Chk1 siRNA relative to cells treated with the control after exposure to either thymidine or HU. However there were no significant differences in the response of the NBS1−/− fibroblasts relative to the corrected cells in these conditions (Figure 5C & D). Interestingly the replication inhibitors did not further increase the level of apoptosis in the NBS1 LB1 (−/−) cells depleted of Chk1.
Given the controversy arising over recent reports of an anti-apoptotic role for BID in response to some forms of DNA damage [26],[27], we determined the effect of BID depletion on the induction of apoptosis following treatment with thymdine or HU. Depletion of BID in HCT116 (Figure 6A) had only minor effects on cell cycle distribution after treatment with thymidine or HU (Figure 6B), although, like ATM depleted cells, an accumulation of cells in mid S-phase was detected. Similarly there were only small changes in the level of cells with a subG1 DNA content or AnnexinV+ after treatment of BID depleted cells relative to cells treated with the control siRNA (Figure 6B–D). The strong induction of subG1 or AnnexinV+ cells in cultures depleted of Chk1 after thymidine or HU treatment was not altered in cells depleted of both Chk1 and BID. In addition co-depletion of BID and ATR or BID and ATM had no further effect on the level of apoptotic cells relative to cells treated with ATR or ATM siRNAs alone (Figure S4). Thus BID does not appear to play a major role in the commitment to apoptosis during replication stress in the tumour cells tested here.
In the work reported here we analysed the roles of key proteins controlling the cellular response to DNA damage in the control of apoptosis following DNA replication stress. The role of ATM in promoting apoptosis in response to ionizing radiation is well established [8],[28]. Our data suggest that it plays little or no role in tumour cell lines in response to DNA replication fork stress relative to the ATR-Chk1 pathway. SiRNA mediated depletion of ATM and downstream ATM phosphorylation targets NBS1 and BID had little or no significant effect on the level of apoptotic cells in response to the replication inhibitors in the tumour cell lines tested. We previously reported that Chk2 depletion did not affect the level of apoptotic cells induced by replication inhibitors [17]. In addition early events occurring after the disruption of DNA replication (accumulation of RPA foci and RPA34 hyperphosphorylation) in ATR- or Chk1-depleted cells committed to apoptosis are not detected in ATM-depleted cells. In immortalized fibroblast lines derived from patients with inherited defects in ATM or NBS1, there was no difference in the apoptotic response of mutant cells to replication inhibitors relative to the corrected lines while depletion of Chk1 or ATR gave robust apoptotic responses in all cell types. The only exception to this pattern was NBS1−/− cells that showed an elevated level of apoptosis following Chk1 depletion in the absence of the replication inhibitor. Since thymidine or HU treatment of these cells did not further increase the level of apoptosis, we speculate that the synthetic lethality observed in these conditions may be a consequence of some disruption of replication in these cells.
Depletion of ATR or Chk1 leads to a consistent robust apoptotic response to replication inhibitors in both tumour and immortalized fibroblast cell lines. The response of the ATR-Chk1 pathway is largely directed at the stabilization of DNA replication following stress [29] and it does not appear to be required to activate downstream proapoptotic proteins. The precise event that initiates the death response in the absence of this signalling pathway is not yet clear. Previous work showing that the replication helicase cofactor Cdc45 is required for the Chk1 suppressed apoptotic response suggests that the role of this signalling pathway in maintaining replication fork integrity and preventing firing of new origins following the disruption of DNA replication is critical [20]. Co-depletion of proteins involved in ATR and ATM signalling pathways does not enhance or inhibit the apoptotic response to replication inhibitors indicating that the ATM signalling pathway is not required for the death response. Interestingly, ATR- and ATM-mediated signalling cascades overlap in response to many forms of DNA damage. For example, both pathways stimulate Cdc25A degradation following activation by DNA damage through the action of the Chk1 and Chk2 checkpoint kinases [2],[30]. Thus activation of either pathway can produce S-phase arrest that, in turn, should favour the anti-apoptotic repair of damaged DNA. However, in the absence of Chk1 or ATR, the ATM-mediated response does not appear to be sufficient to protect cells from apoptosis.
It has been reported that the pro-apoptotic protein BID participates in ATM-mediated protein kinase cascade to regulate entry into S-phase and prevent death in response to DNA replication inhibitors. BID showed nuclear localization and was phosphorylated in an ATM-dependent manner following DNA damage in myeloid progenitor cells derived from wild type mice [22]. Furthermore, cells from BID−/− mice showed a strong apoptotic response following treatment that was not evident in BID+/+ cells. In another report mouse embryo fibroblasts obtained from BID−/− mice showed a delayed entry into S-phase but not cell death following exposure to DNA damaging agents [31]. More recently these observations were disputed as several cell types from BID−/− mice generated on a different genetic background failed to show any significant change in S-phase arrest, survival, or apoptosis relative to BID+/+ cells [32]. In our knockdowns of BID in the human colon cancer cell line HCT116, no significant increase in the frequency of apoptotic cells was observed. However, BID depleted cells treated with thymidine accumulated in mid S-phase, suggesting that transition through S-phase was delayed relative to cells treated with the control siRNA.
The data reported here show that the apoptotic pathway suppressed by Chk1 in response to replication inhibitors is clearly distinguishable from both the classical intrinsic death pathway and the Chk1-suppressed IR death response (Figure 7). Unlike the intrinsic pathway, the Chk1 suppressed response to replication inhibitors does not require p53 or Chk2. The Chk1 suppressed death pathway responding to IR is not triggered following depletion of ATR and it requires ATM, ATR and caspase 2 [19]. Although this pathway was identified in screen of p53 deficient zebrafish, p53 deficient human tumour cells treated with Chk1 inhibitors also show caspase 2 cleavage and caspase 2 dependent apoptosis in response to IR. In p53 proficient tumour cells, the cleaved caspase 2 is not detected but caspase 3 is activated under these conditions. In contrast both ATR and Chk1 depleted cells undergo apoptosis in response to replication inhibitors regardless of p53 status and ATM is not required for death. Caspase 3 is clearly activated in both p53 proficient and deficient cell lines. Cleaved caspase 2 is not detected in p53 proficient HCT116 cells and is only weakly induced in the p53 deficient cells. Thus while Chk1 is required for the suppression of apoptosis in response to DNA structural alterations induced by IR or replication stress, the pathways suppressed are distinctly different.
Given the lethal effects of loss of Chk1 function on tumour cells exposed to DNA replication inhibitors, there has been interest in the use of Chk1 inhibitors in chemotherapy. Chk1 is highly expressed in some types of tumours [33]. This may confer some growth advantage to tumour cells as a result of the role of this protein in protecting cells from replication stress that may be induced by hypoxia or nutrient deprivation during tumour development [34]. However loss of Chk1 can also be lethal to some normal cell types [35] and Chk1 knockout mice show embryonic lethality [36]. Nevertheless recent work has shown that Chk1 inhibitors can be used to increase the sensitivity of tumour cells to replication inhibitors in vitro and in vivo [37]. Furthermore, Chk1 knockdown experiments suggest enhanced lethality for tumour cells may be obtained where the protein is only partially depleted, thus reducing potential lethality caused by complete loss of Chk1. Notably Chk1 inhibitors have been developed that appear to enhance the toxicity of DNA damaging agents in p53 deficient tumour cells but not p53-proficient cells [38] offering prospects for the targeted activation of the Chk1-suppressed apoptotic pathway in at least some types of tumour cells.
The HCT116 and SW480 human colon cancer cell lines were obtained from American Type Culture Collection (Manassas, VA) while the AT patient derived AT5 cells (TAT5BIVA) was obtained from the European Cell and Culture Collection. HCT116 p53−/− cells were provided by Dr. Bert Vogelstein (Johns Hopkins University, Baltimore, MD). AT-deficient (AT22IJE-T referred to as pEBs here) and corrected (YZ5) cell lines were kindly provided by Dr. Yosef Shiloh (Tel Aviv University, Tel Aviv, Israel). Nbs1 deficient (NBS1-LB1) and corrected (p95wt) fibroblasts were generously provided by Dr. Mike Kastan (St. Jude Children's Research Hospital, Memphis, TN). Cells were maintained in DMEM supplemented with 10% fetal bovine serum (FBS). For experiments using thymidine, dialyzed FBS was used to remove deoxynucleosides in the serum that might interfere in the response to this agent.
All siRNAs were obtained from Dharmacon (Lafayette, CO). The ATM and ATR siRNA consisted of a pool of four sequences designed to the relevant DNA sequence. Chk1 siRNAs were designed by J. Blackburn and C. Smythe, (unpublished data). Nbs1 siRNAs (GUCGAUCAGCCGAAAUCAU, CUCACCUUGUCAUGGUAUC, and GCUAGGUUGAUAACAGAAG) were designed by A. Ganesh and the control siRNA was obtained from Eurogentec (OR-0030-NEG). SiRNA duplexes were transfected into cells using Lipofectamine 2000 (Invitrogen, Paisley, United Kingdom) according to manufacturer's instructions. The cells were then incubated for twenty-four hours before further treatment.
After treatment, floating (obtained from the medium and a PBS wash) and adherent (obtained after trypsinization) cells were pelleted together by centrifugation. Cell pellets were washed with PBS, fixed in 70% ice-cold ethanol, and stored at −20°C for up 2 weeks. Cells were incubated overnight with Propidium Iodide as described previously [17]. Stained nuclei were analyzed on a FACScan (BD Biosciences, Franklin Lakes, NJ) using CellQuest software.
Apoptotic cells were examined using fluorescein isothiocyanate (FITC)-Annexin V and PI detection kit according to the manufacturer's instructions (BD Biosciences). The cells were analyzed by flow cytometry and the percentage of early apoptotic (% Annexin V/PI) cells is presented.
Cells were grown on glass coverslips, treated as indicated, fixed with 3% buffered paraformaldehyde for 15 minutes at room temperature (RT) and permeabilized in PBS containing 0.5% Triton X-100 for 8 minutes at RT. Cells were then incubated with 1/250 diluted anti-RPA34 (NA19L; Calbiochem) for 45 minutes at RT and 1∶500 diluted Alexa-594 conjugated anti-mouse IgG (A11005; Molecular Probes, Invitrogen) for 30 minutes at RT and in the dark. Antibody dilutions and washes after incubations were performed in PBS containing 0.5%BSA and 0.05%Tween 20. Coverslips were finally mounted in Vectashield mounting medium with DAPI (H-1500; Vector Laboratories Inc.). For fluorescent analysis, a Nikon Eclipse T200 microscope equipped with a Hamamatsu Orca ER camera and the Volocity 3.6.1 (Improvision) software was used.
Cell extracts were prepared and fractionated on SDS-PAGE gels before being blotted onto nitrocellulose (Whatman Schleicher and Schuell, Dassel, Germany) as described previously [21]. Proteins were detected with the ECL detection system (GE Healthcare, Little Chalfont, Buckinghamshire, United Kingdom) using antibodies recognizing ATM (GeneTex Inc, San Antonio, Texas), ATR, (Santa Cruz Biotechnology, Santa Cruz, CA), BID (Santa Cruz Biotechnology) Chk1 (Cell Signaling Technology, Beverly, MA), NBS1 (Cell Signaling Technology), and β-actin (Sigma-Aldrich), RPA34 (NA19L; Calbiochem), Caspase-2 (MAB3507; Millipore), cleaved caspase-3 (ab32042; Abcam).
|
10.1371/journal.ppat.1006036 | An Interactome-Centered Protein Discovery Approach Reveals Novel Components Involved in Mitosome Function and Homeostasis in Giardia lamblia | Protozoan parasites of the genus Giardia are highly prevalent globally, and infect a wide range of vertebrate hosts including humans, with proliferation and pathology restricted to the small intestine. This narrow ecological specialization entailed extensive structural and functional adaptations during host-parasite co-evolution. An example is the streamlined mitosomal proteome with iron-sulphur protein maturation as the only biochemical pathway clearly associated with this organelle. Here, we applied techniques in microscopy and protein biochemistry to investigate the mitosomal membrane proteome in association to mitosome homeostasis. Live cell imaging revealed a highly immobilized array of 30–40 physically distinct mitosome organelles in trophozoites. We provide direct evidence for the single giardial dynamin-related protein as a contributor to mitosomal morphogenesis and homeostasis. To overcome inherent limitations that have hitherto severely hampered the characterization of these unique organelles we applied a novel interaction-based proteome discovery strategy using forward and reverse protein co-immunoprecipitation. This allowed generation of organelle proteome data strictly in a protein-protein interaction context. We built an initial Tom40-centered outer membrane interactome by co-immunoprecipitation experiments, identifying small GTPases, factors with dual mitosome and endoplasmic reticulum (ER) distribution, as well as novel matrix proteins. Through iterative expansion of this protein-protein interaction network, we were able to i) significantly extend this interaction-based mitosomal proteome to include other membrane-associated proteins with possible roles in mitosome morphogenesis and connection to other subcellular compartments, and ii) identify novel matrix proteins which may shed light on mitosome-associated metabolic functions other than Fe-S cluster biogenesis. Functional analysis also revealed conceptual conservation of protein translocation despite the massive divergence and reduction of protein import machinery in Giardia mitosomes.
| Organelles with endosymbiotic origin are present in virtually all extant eukaryotes and have undergone considerable remodeling during > 1 billion years of evolution. Highly diverged organelles such as mitosomes or plastids in some parasitic protozoa are the product of extensive secondary reduction. They are sufficiently unique to generate interest as targets for pharmacological intervention, in addition to providing a rich ground for evolutionary cell biologists. The so-called mitochondria-related organelles (MROs) comprise mitosomes and hydrogenosomes, with the former having lost any role in energy metabolism along with the organelle genome. The mitosomes of the intestinal pathogen Giardia lamblia are the most highly reduced MROs known and have proven difficult to investigate because of their extreme divergence and their unique biophysical properties. Here, we implemented a novel strategy aimed at systematic analysis of the organelle proteome by iterative expansion of a protein-protein interaction network. We combined serial forward and reverse co-immunoprecipitations with mass spectrometry analysis, data mining, and validation by subcellular localization and/or functional analysis to generate an interactome network centered on a giardial Tom40 homolog. This iterative ab initio proteome reconstruction provided protein-protein interaction data in addition to identifying novel organelle proteins and functions. Building on this data we generated information on organelle replication, mitosome morphogenesis and organelle dynamics in living cells.
| Since the single endosymbiotic event leading to establishment of mitochondria approximately 2 billion years ago [1,2,3] these organelles have undergone massive changes and have evolved into highly specialized and essential subcellular compartments in all eukaryotes [4,5], with only one possible exception identified so far [6]. These changes comprise a dramatic size reduction, nuclear transfer of organelle genomes, and a renewal of the proteome, which is synthesized almost entirely as precursor proteins on cytosolic ribosomes [7,8,9,10,11,12,13,14] and imported from the cytoplasm [15]. Mitochondria have been remodeled and/or restructured to very different degrees in different species. Mitochondria-related organelles (MROs), i.e. hydrogenosomes and mitosomes [16,17,18,19,20] in some protists lacking canonical mitochondria represent extreme forms of reduction and/or divergence. The potential of highly diverged organelle-specific pathways as targets for intervention has sparked research into the evolution of MROs in single-celled organisms of all five eukaryotic supergroups [21,22]. Notably, the microaerophilic protozoan pathogens Entamoeba histolytica [20] and Giardia lamblia [23,24], as well as intracellular parasites such as Cryptosporidium parvum [25] and Encephalitozoon cuniculi [26] harbor mitosomes. Interestingly, recent investigation of MROs in Spironucleus salmonicida, a diplomonad and the closest relative of G. lamblia belonging to the Excavata super-group, revealed that these organelles are in fact hydrogenosomes [27]. Although it has been demonstrated that G. lamblia mitosomes do not produce hydrogen, this sheds a completely new light on the evolution of MROs in diplomonads.
Proliferating G. lamblia trophozoites contain 20–50 double membrane-bounded 100 nm spherical mitosomes [23,24] devoid of an organelle genome [28,29,30,31]. Although not proven experimentally, G. lamblia mitosomes are likely essential due to a subset of conserved mitochondrial proteins required for iron- sulphur (Fe-S) protein maturation [23,32,33,34,35]. Yeast genetic experiments suggested that Fe-S protein maturation, the only function currently ascribable to G. lamblia mitosomes, is in fact the minimal essential function of mitochondria [36]. Hence, these organelles have also attracted considerable interest as cell biological models to study extreme reductive evolution of MROs [23,37,38,39,40,41,42]. However, due to massive, albeit selective sequence divergence in G. lamblia, conventional data mining strategies for identification of mitosome proteins based on homology-based in silico searches fall short [26,28,32,43,44,45,46,47]. Moreover, classical, organelle enrichment-based proteome analyses approaches have had only limited success owing to the small size of the organelles and the omnipresence of contaminating endoplasmic reticulum (ER) and cytoskeleton elements in mitosome fractions [33,48,49].
Nevertheless, there is unambiguous experimental evidence for the functional conservation of the mitosomal protein import machinery [20,23,24,49]. The small subset of structurally conserved mitosome proteins such as G. lamblia IscU, ferredoxin, Cpn60, IscS and mtHsp70 are imported by transit peptide-dependent and -independent mechanisms [23]. However, the predicted components of the TOM/TIM import apparatus are diverged beyond recognition by state-of-the-art homology search tools. Indeed, the protein repertoire of the mitosomal outer membrane and its networks are scarcely characterized: only one subunit of the translocon in the outer mitochondrial (TOM) complex, a highly diverged Tom40 homologue (GlTom40), and [50] more recently a giardial Tim44 homologue [49], have been identified. Furthermore, there is no information on how mitosome homeostasis is achieved in terms of organelle size and number.
To address questions concerning protein networks at mitosomal membranes in association with mitosome homeostasis and to account for the extreme sequence divergence in G. lamblia, we implemented novel experimental approaches. We were successful to tag two outer membrane organelle proteins with GFP to show that these small organelles are immobilized, distinctive entities with no appreciable inter-organelle exchange or network character. Using a giardial TOM40 homolog as a starting bait we generated information on protein-protein interactions at the outer membrane as well as expanding the organelle proteome by identifying novel components. By using interactome targets validated by subcellular localization as baits for subsequent reverse co-IP rounds, we were able to extend this initial interactome beyond the outer membrane, including dually localized endoplasmic reticulum (ER) and mitosome proteins, as well as identifying previously described and novel imported organelle proteins. In addition to identification of two components with a role in mitosome morphogenesis and homeostasis the combined data revealed a core organelle membrane interactome composed of only 3 tightly-associated proteins. Furthermore, we tested constraints for import of nuclear-encoded mitosome proteins and could show conservation of this mechanism even in the highly diverged and reduced Giardia mitosome.
G. lamblia WB (line C6; ATCC catalog number 50803) trophozoites were grown and harvested using standard protocols [51]. Encystation was induced with the two-step method as described previously [40,52]. Transgenic parasites were generated according to established protocols by electroporation of linearized pPacV-Integ-based plasmid vectors prepared from E. coli as described in [42]. After selection for puromycin resistance, transgenic G. lamblia cell lines were cultured without puromycin.
All sequences of oligonucleotide primers for PCR used in this work are listed in S1 Table.
For cloning of C-terminally hemagglutinin (HA)-tagged proteins in Giardia, a vector PAC-CHA was designed on the basis of the previously described vector pPacV-Integ [42], where additional restriction sites were inserted [53].
A cyst wall protein 1 promoter (pCWP1)-driven G. lamblia ferredoxin (fd)-human dihydrofolate reductase (DHFR) chimeric gene was generated by fusing two genes by overlapping PCR: i) an intron-less fd mitosomal targeting signal (MTS) (MTSfdΔint) open reading frame (ORF) was generated using primer pair 33 (S1 Table) with G. lamblia cDNA as template, ii) a DHFR_HA minigene was generated using primer pair 34 (S1 Table) with a cloned human DHFR cDNA as template. The fused product was digested with SpeI and PacI and inserted in a PAC vector to yield construct pCWP1_MTSfdΔint-DHFR_HA.
A pCwp1_ MTSfdΔint-DHFR_Neomycin resistance construct (without HA tag) was generated for protein import block assays. Primer pair 35 (S1 Table) was used on pCwp1_ MTSfdΔint-DHFR_HA as a template. The amplified product was digested with NsiI and PacI and ligated into a vector containing a neomycin resistance cassette [51].
G. lamblia WBC6 and transgenic trophozoites expressing C-terminally HA tagged bait proteins were harvested and subjected to immunofluorescence assay to confirm correct subcellular distribution of bait proteins. Parasites were collected by centrifugation (900 x g, 10 minutes, 4°C), washed in 50 ml of cold phosphate buffer saline solution (PBS) and adjusted to 2 x107 cells.ml-1 in PBS (VWR Prolabo). The appropriate formaldehyde concentration for cross-linking (2.25%) was determined by a titration assay (S2 Fig). For the co-immunoprecipitation (co-IP) assays, 109 parasites were resuspended in 10 ml 2.25% formaldehyde (in PBS) supplemented with 1 mM phenylmethylsulfonyl fluoride (PMSF; SIGMA, Cat. No. P7626) and incubated for 30 minutes at room temperature (RT). Cells were pelleted, washed once with 10 ml PBS, and quenched in 10 ml 100 mM glycine in PBS for 15 minutes at RT. The collected cells were then resuspended in 5 ml RIPA lysis buffer (50 mM Tris pH 7.4, 150 mM NaCl, 1% IGEPAL, 0.5% sodium deoxycholate, 0.1% SDS, 10 mM EDTA) supplemented with 2 mM PMSF and 1 x Protease Inhibitor cocktail (PIC, Cat. No. 539131, Calbiochem USA) and sonicated twice using a Branson Sonifier with microtip (Branson Sonifier 250, Branson Ultrasonics Corporation) with the following settings: 60 pulses, 2 output control, 30% duty cycle and 60 pulses, 4 output control, 40% duty cycle. The sonicate was incubated on a rotating wheel for 1 h at 4°C, aliquoted into 1.5 ml tubes and centrifuged (14,000 x g, 10 minutes, 4°C). The soluble protein fraction was mixed with an equal volume detergent-free RIPA lysis buffer supplemented with 2% TritonX (TX)-100 (Fluka Chemicals) and 40 μl anti-HA agarose bead slurry (Pierce, product # 26181). After binding of tagged proteins to the beads at 4°C for 2 h on a rotating wheel, beads were pulse-centrifuged and washed 4 times with 3 ml Tris-Buffered Saline (TBS) supplemented with 0.1% TX-100 at 4°C. After a final wash with 3 ml PBS the loaded beads were resuspended in 350 μl PBS, transferred to a spin column (Pierce spin column screw cap, product # 69705, Thermo Scientific) and centrifuged for 10 s at 4°C. Elution was performed by resuspending beads in 30 μl of PBS. Dithiothreitol (DTT; 100mM; Thermo Scientific, Cat. # RO861) was added and samples were boiled for 5 min followed by centrifugation (14,000 x g, 10 minutes, RT).
SDS-PAGE and immunoblotting analysis of input, flow-through, and eluate fractions was performed on 4%-12% polyacrylamide gels under reducing conditions, (molecular weight marker Cat. No. 26616, Thermo Scientific, Lithuania). Transfer to nitrocellulose membranes and antibody probing were done as described previously [54], using anti-HA (dilution 1:500; Roche) followed by anti-rat antibodies coupled to horseradish peroxidase (dilution 1:5000; Southern Biotech). Gels for mass spectrometry (MS) analysis were stained using Instant blue (Expedeon, Prod. # ISB1L) and de-stained with sterile water.
Stained gel lanes were cut into 8 equal sections. Each section was further diced into smaller pieces and washed twice with 100 μl of 100 mM ammonium bicarbonate/ 50% acetonitrile for 15 min at 50°C. The sections were dehydrated with 50 μl of acetonitrile. The gel pieces were rehydrated with 20 μl trypsin solution (5 ng/μl in 10 mM Tris-HCl/ 2 mM CaCl2 at pH 8.2) and 40 μl buffer (10 mM Tris-HCl/ 2 mM CaCl2 at pH 8.2). Microwave-assisted digestion was performed for 30 minutes at 60°C with the microwave power set to 5 W (CEM Discover, CEM corp., USA). Supernatants were collected in fresh tubes and the gel pieces were extracted with 150 μl of 0.1% trifluoroacetic acid/ 50% acetonitrile. Supernatants were combined, dried, and the samples were dissolved in 20 μl 0.1% formic acid before being transferred to the autosampler vials for liquid chromatography-tandem MS (injection volume 7 to 9 μl). Samples were measured on a Q-exactive mass spectrometer (Thermo Scientific) equipped with a nanoAcquity UPLC (Waters Corporation). Peptides were trapped on a Symmetry C18, 5 μm, 180 μm x 20 mm column (Waters Corporation) and separated on a BEH300 C18, 1.7 μm, 75 μm x 150 mm column (Waters Corporation) using a gradient formed between solvent A (0.1% formic acid in water) and solvent B (0.1% formic acid in acetonitrile). The gradient started at 1% solvent B and the concentration of solvent B was increased to 40% within 60 minutes. Following peptide data acquisition, database searches were performed using the MASCOT search program against the G. lamblia database (http://giardiadb.org/giardiadb/) with a concatenated decoy database supplemented with commonly observed contaminants and the Swissprot database to increase database size. The identified hits were then loaded onto the Scaffold Viewer version 4 (Proteome Software, Portland, US) and filtered based on high stringency parameters, i.e. 95% for peptide probability, a protein probability of 95%, and a minimum of 2 unique peptides per protein. Where indicated in the text, slightly relaxed filtering parameters were applied. Proteins identified in both bait-specific and control datasets were considered of interest if they were at least 5-fold enriched in the bait-specific datasets (in terms of spectral counts) based on high stringency parameters. Access to raw MS data is provided through the ProteomeXchange Consortium on the PRIDE platform [55].
Analysis of primary structure and domain architecture of putative mitosomal hypothetical proteins was performed using the following tools and databases: MITOPROT (https://ihg.gsf.de/ihg/mitoprot.html) and PSORTII (http://psort.hgc.jp/form2.html) for subcellular localization prediction, TMHMM (http://www.cbs.dtu.dk/services/TMHMM/) for transmembrane helix prediction, SMART (http://smart.embl-heidelberg.de/) for prediction of patterns and functional domains, pBLAST for protein homology detection (http://blast.ncbi.nlm.nih.gov/Blast.cgi?PAGE=Proteins), HHPred (http://toolkit.tuebingen.mpg.de/hhpred) for protein homology detection based on Hidden Markov Model (HMM-HMM) comparison, and the Giardia Genome Database (http://giardiadb.org/giardiadb/) to extract other/organism-specific information, e.g. expression levels of the protein, predicted molecular size and nucleotide/protein sequence. For functional domains predicted by SMART we used an e-value of 10e-5 as cutoff, and for protein homologies predicted by pBLAST we accepted alignment scores above 80. However, since G. lamblia homologs for eukaryotic proteins are highly diverged, we also considered functional domain predictions associated to a lower e-value. Alignment scores between 50 and 80 were accepted only when pBLAST predictions were consistent with HHPred output.
Preparation of chemically fixed cells for immunofluorescence and analysis of subcellular distribution of reporter proteins by wide-field and confocal microscopy were done as described previously [42,54]. Nuclear labelling was performed with 4',6-diamidino-2-phenylindole (DAPI). The HA epitope tag was detected with a monoclonal anti-HA antibody coupled to FITC (dilution 1:50; Roche) whereas GlIscU was detected with a self-made antibody (dilution 1:300) followed by an anti-mouse antibody coupled to Alexafluor 594 (dilution 1:300; Molecular Probes). To avoid any possibility for cross reaction, co-labelling experiments for IFA were performed by incubating first with the anti-GlIscU antibody, followed by the AF594-conjugated anti-mouse secondary antibody, and direct detection of the HA epitope tag with a FITC-conjugated rat anti-HA monoclonal antibody as a final step.
Transgenic G. lamblia trophozoites expressing GFP-GlTom40 or Gl29147-GFP were harvested and prepared for imaging in PBS supplemented with 5 mM glucose (Cat. No. 49139, Fluka), 5 mM L-cysteine (Cat. No. C6852, Sigma) and 0.1 mM ascorbic acid (Cat. No. 95209, Fluka) at pH 7.1. FRAP and time-lapse series were performed as described previously [54,56].
Transgenic trophozoites ectopically expressing wild type G. lamblia dynamin related protein (GlDRP) (ORF Gl50803_14373) or the constitutively active (GTP-locked) GlDRP-K43E variant under the control of the CWP1 promoter [56] were harvested 3 h post induction and analyzed by transmission electron microscopy (TEM) as described previously [56].
For sub-cellular fraction experiments, 4.106 GlDRP-HA and GlDRP-K43E-HA- expressing transgenic cells were lysed by freeze-thawing and supernatant (soluble fraction) and pellet (membrane fraction) were prepared by centrifugation at 14’000 x g for 10 minutes at 4°C. The HA-tagged proteins were detected by SDS-PAGE and Western blot using a rat anti-HA mAb (clone 3F10, Roche) as described previously [54].
The MTSfdΔint-DHFR fusion (see also above under “Constructs”) was expressed under the control of the inducible CWP1 promoter in a background transgenic line constitutively expressing HA- tagged 17030 (cell line Cwp1_MTSfdΔint-DHFR/Gl17030HA). DHFR expression was induced using the 2-step method [40] for 4 h and “chased” for 24 h by placing the cells again in standard growth medium in the presence or absence of 1 μM methotrexate (MTX). Total cell lysates were separated by SDS-PAGE and Western blot to detect processed and unprocessed forms of the Gl17030HA reporter. Subcellular distribution was analyzed by immunofluorescence assay (IFA) using wide field microscopy.
Mitochondria in higher eukaryotes are highly dynamic organelle networks that move in the cell via microtubules and microfilaments and undergo constant fission and fusion to meet the energy requirements of the cell [57,58]. IFA and TEM analyses suggest that G. lamblia mitosomes are very small spherical organelles with no evidence of network formation. In addition, the mitosome population in each cell can be divided into peripheral mitosomes (PM) distributed randomly in the cytoplasm and what has been dubbed the central mitosome complex (CMC) [23]. The latter consists of a grape-like cluster of individual organelles of the size and shape of peripheral mitosomes that is closely and permanently associated to the basal body complex between the two nuclei [23]. Interestingly, these organelles remain spatially distinct despite their close proximity. The motility of this central cluster is highly constrained and restricted to ordered segregation with the duplicated basal body complex during cell division [23]. Because green fluorescent protein (GFP) imported into the mitosome matrix is not fluorescent [23], GFP-tagging of mitosomes has not been possible until now. Martinkova et al. [59] have shown that mitosomes in trophozoites can be labeled for live cell microscopy using HaloTag markers [60]. However, no quantitative information on the spatial dynamics of peripheral mitosomes in the cytoplasm was presented in this report. We investigated organelle dynamics in living cells by performing time lapse microscopy of cells expressing GFP-tagged mitosome reporters for the outer membrane. Conditional expression of N-terminally GFP-tagged GlTom40 with 3 h of induction followed by “chasing” newly-synthesized GFP-Tom40 into mitosomes over 2–3 h in normal conditions was found suitable for labeling organelles in living cells (Fig 1A and 1B). Tracking of individual organelles over a period of >30 min showed no significant cytoplasmic movement or changes in number or morphology (Fig 1C), suggesting that organelles neither move randomly nor are they transported directionally in the cytoplasm along cytoskeleton structures. To test whether mitosome outer membrane proteins are exchanged between organelles we performed FRAP experiments on cells conditionally expressing GlTom40-GFP. Since GlTom40-GFP is membrane-anchored, FRAP addresses the question whether mitosomes are isolated organelles and whether they form membrane continuities which would allow exchange of outer membrane proteins. No recovery of fluorescence in bleached CMC or PM organelles was detected (Fig 1D–1G) suggesting that peripheral and CMC organelle membranes remain distinct.
Despite intensive research in the field of MROs, little is known regarding factors required for their division. Dynamin-related proteins (DRPs) are implicated in mitochondrial and hydrogenosomal division in higher eukaryotes and in protozoa such as Trypanosoma brucei [61,62] and Trichomonas vaginalis [63]. G. lamblia harbors a single DRP (ORF Gl50803_14373) [56] with a previously documented role in trafficking of cyst wall material, and endocytic and exocytic organelle homeostasis [56]. To test for a hitherto unrecognized role of GlDRP in determining mitosome morphology and/number, we used a dual cassette expression vector [54] to express constitutive C-terminally myc-tagged GlTom40 as a reporter for mitosomes and inducible C-terminally HA-tagged wild-type (GlDRP-HA) or GTP-locked (GlDRP-K43E-HA) variants in trophozoites. In trophozoites expressing GlDRP-HA (Fig 2A–2C), IFA analyses demonstrated the typical random cytoplasmic distribution of PMs i.e. “dispersed” [23]. However, cells expressing the GTP-locked variant GlDRP-K43E-HA (Fig 2D–2F) presented a “clustered” mitosome phenotype, indicative of enlarged organelles. Consistent with this phenotype and in line with previous reports [56], the subcellular distribution of HA-tagged GlDRP remained mostly cytosolic (Fig 2B). Conversely, GlDRP-K43E-HA showed a punctate distribution (Fig 2E) and significant signal overlap with GlTom40-myc (Fig 2F), suggesting selective accumulation of GlDRP-K43E-HA on mitosome membranes. We tested whether this marked association of ectopically expressed GlDRP-K43E with organelle membranes compared to the wild type DRP variant in IFA could be corroborated in cell fractionation experiments. Separation by SDS-PAGE and immunoblot analysis revealed that GlDRP-HA was almost equally distributed between the “cytosolic” and “membrane” fraction, whereas the mutated variant GlDRP-K43E-HA was detected only in the “membrane” fraction (Fig 2G). These data were consistent with the microscopical analysis in Fig 2E and suggest increased association of GlDRP-K43E-HA with organelle membranes compared to wild-type GlDRP-HA. To characterize the nature of the GlDRP-K43E-HA-dependent phenotype in more detail, we performed transmission electron microscopy of induced transgenic cells. Cells expressing the GlDRP-K43E-HA variant frequently presented elongated and tubular mitosome structures (Fig 2I and 2J) compared to cells expressing wild type GlDRP-HA (Fig 2H).
Taken together, these data show how mitosomes are immobilized in the cell and present no measurable outer membrane exchange in the conditions tested. Their morphogenesis is perturbed following conditional ectopic expression of a dominant-negative GTP-locked GlDRP variant, suggesting a previously unappreciated role for this GTPase in the maintenance of mitosome integrity and organelle morphogenesis in G. lamblia.
The aberrant mitosome morphology after conditional expression of GlDRP-K43E points towards mitosome-associated machinery at the organelle’s surface involved in organelle homeostasis. Despite efforts aimed at defining the protein content of mitosomes in Giardia [33,49,50], the composition of this organelle’s outer and inner membrane proteome remains sparsely characterized, with the exception of a highly diverged putative Tom40 homologue (GlTom40; ORF Gl50803_17161) and a structurally-conserved Tim44 [49,50]. To generate a robust mitosome outer membrane proteome we focused on GlTom40 as a point of origin and developed a tailored co-IP protocol with an HA-tagged variant as “bait”. A transgenic line GlTom40-HA constitutively expressing the epitope-tagged bait protein was generated; exclusive mitosome localization of the bait protein in transgenic cells was confirmed by IFA in co-labelling experiments with a newly-made anti-GlIscU antibody (Fig 3A and S1A Fig). To ensure solubilization of mitosomal membranes while avoiding disruption of Tom40-associated protein complexes, we used carefully titrated, formaldehyde-based cross-linking [64] to stabilize predicted protein-protein interactions in co-IP experiments during extraction with the option to reverse covalent bonds (S2 Fig; see also in Materials and Methods). Following MS analysis and data filtration using a control dataset obtained from non-transgenic cells (ctrl.co-IP) we identified a total of 52 proteins, 46 exclusive and 6 enriched in the GlTom40 co-IP dataset (Fig 3B). This protein set was parsed and subdivided into different metabolic and/or functional categories (Fig 3C). In the mitosomal protein category few detected four previously identified mitosome proteins namely: mitochondrial HSP70 (ORF Gl50803_14581), oxidoreductase 1 (GlOR1; ORF Gl50803_91252), proteins Gl50803_9296 and Gl50803_14939, recently named MOMP35 [33,49]. We extracted additional information from the GlTom40 co-IP data by relaxing stringency parameters to (95_1_95), obtaining a total of 150 proteins (FDR 3.4%). Of these, 109 hits were exclusive to the expanded GlTom40 co-IP dataset which contained 3 additional annotated mitosome proteins namely, chaperonin 60 (Cpn60; ORF Gl50803_103891), GlQb-SNARE 3 (putative Sec20, ORF Gl50803_5161) and GlIscU (NifU-like protein; ORF Gl50803_15196).
Limited chemical cross-linking in co-IP assays expands the range of discovery beyond primary interactions with the bait. We therefore performed an initial validation of the predicted GlTom40 interacting proteins in this dataset by subcellular localization of ectopically expressed, epitope-tagged candidates to mitosomes. We selected 13 of the 109 candidate Mitosomal Outer Membrane Tom40 interacting proteins (MOMTiP; Table 1) based on their spectral counts with high stringency parameters and/or protein domains identified with HHPred (S2 Table) and engineered endogenous promoter-driven, C-terminally HA-tagged variants for all. IFA analysis of corresponding transgenic lines showed mitosomal localization for 8 candidates (Fig 4A–4H), of which 4 proteins of unknown function (MOMTiP-5 to 8) presented dual localization (mitosome and ER) (Fig 4F–4I). The five remaining proteins of this set of 13 candidates (MOMTiP- 9–13; Fig 4J–4N) showed dispersed patterns of subcellular distribution and were not considered mitosome proteins. Fig 4O shows a consolidated depiction of a first GlTom40-centered mitosomal outer membrane interactome, which includes the 8 proteins localized to mitosomes described above, as well as 4 previously identified matrix proteins and 3 newly validated hypothetical proteins comprised in the list of GlTom40 interacting proteins. Taken together, the imaging data are in agreement with the protein-protein interaction data, and support limited chemical crosslinking as a suitable method to stabilize protein complexes during co-IP.
IFA analysis of MOMTiP-1 to 13 indicated that the majority of these proteins are associated to mitosomes, thereby providing preliminary validation of the selected 13 candidates of the primary GlTom40-specific co-IP dataset. To further test the robustness of this primary interactome and expanding it beyond the mitosomal membrane, we performed a first reverse co-IP experiment using MOMTiP-1 (ORF Gl50803_29147) as bait. MOMTiP-1 was chosen because it presented the largest spectral count with high stringency parameters in the GlTom40 dataset and localized unequivocally to mitosomes (S1B Fig).
MOMTiP-1 is a Giardia-specific mitosome-localized protein of unknown function. In silico analysis using TMHMM robustly detected a 22 amino acid-long transmembrane helix in the N-terminal part of the protein followed by a large C-terminal domain predicted to be exposed to the cytosol on the mitosomal surface. To track this protein in vivo, we engineered MOMTiP-1 constructs for live cell imaging using GFP reporters. We have shown previously that GFP only fluoresces if exposed to the cytoplasm and never after import into mitosomes [23]. Therefore, the brightly fluorescing and mitosome-localized MOMTiP-1-GFP fusion supports the predicted topology for MOMTiP-1 as a type 1 transmembrane protein with respect to the outer mitosomal membrane. Surprisingly, many cells expressing MOMTiP-1-GFP showed a mitosome morphology dubbed “string” phenotype suggestive of extensive elongation of organelles to large tubules (Fig 5A; left). In many cases, virtually all PMs had been replaced by a single long organelle with a diameter that corresponded to that of an individual mitosome. Although the “string” mitosome phenotype was compatible with survival of the parasites, many trophozoites appeared to be delayed or even arrested in cytokinesis and had a typical heart-shaped appearance (Fig 5A; middle) previously observed in cells which are unable to complete cytokinesis [66]. Because the tubular organelles ran through the non-divided part connecting both daughter cells, we postulated that inability to divide mitosomes impairs completion of cytokinesis.
Co-IP with an HA-tagged variant of MOMTiP-1 yielded a large dataset of 221 exclusive hits (Fig 5B) which included GlTom40 detected at high stringency parameters, thereby confirming the strong interaction between GlTom40 and MOMTiP-1. The 221 MOMTiP-1 co-IP specific hits and an additional 20 enriched candidates were parsed according to different metabolic and/or functional categories (Fig 5B). In addition to GlTom40, the dataset contained several known mitosomal proteins, including matrix proteins HSP70 and GiOR1, cysteine desulfurase (IscS; Gl50803_14519), Cpn60, [2Fe-2S] ferredoxin (Gl50803_27266) and NifU-like protein, along with all 8 hypothetical proteins previously identified in the GlTom40 co-IP dataset and 4 additional non-annotated candidate mitosome proteins (Fig 5C–5F). Similarly to MOMTiP-1, one of these (Gl50803_17276) is also predicted to carry a TMD close to its N-terminus. Furthermore, this dataset contained two axoneme-associated GASP-180 proteins (Gl50803_137716 and Gl50803_16745) [67] detected with high stringency parameters, in line with association of the CMC to basal bodies.
Taken together, a first reverse co-IP analysis using the single-pass transmembrane MOMTiP-1 provided robust validation of the experimental approach used to identify mitosome membrane proteins, and has expanded the predicted mitosomal membrane and import machinery interactome to 22 proteins (Fig 5G).
Reverse co-IP using MOMTiP-1 as bait demonstrated that this protein and GlTom40 are strong interaction partners. We analyzed the intersection of their respective datasets to identify common candidate interaction partners and identified 27 proteins with high reliability (Fig 5H), 10 of which localized to mitosomes (Fig 4). Given MOMTiP-1’s predicted topology, strong interaction with GlTom40 and the interactome overlap, we postulated that MOMTiP-1 and GlTom40 exist in a core complex mostly likely involved in protein translocation across the outer mitosomal membrane. To characterize other components of this core interactome of the outer mitosomal membrane and to move beyond individual complexes to explore the boundaries of the growing protein interactome network (Fig 5G), we performed a series of additional reverse co-IP experiments using HA-tagged Qb-SNARE 4 (MOMTiP-7), GlIscS, protein Gl50803_9296 (MOMTiP-4) and protein Gl50803_14939 (MOMTiP-3) as baits. MOMTiP-7 (Qb-SNARE 4), MOMTiP-4 and MOMTiP-3 were chosen because they were identified either exclusively or in both the GlTom40- and MOMTiP-1 co-IP datasets, suggesting they may reside in the mitosomal outer membrane and could thus serve as tools for a lateral and outward expansion of this compartment’s interactome. On the other hand, GIIscS was chosen to extend the mitosomal proteome inwards towards the organellar matrix. Correct mitosomal localization for all 4 HA-tagged variants had been previously confirmed by IFA (Fig 4 and S1C–S1F Fig).
Evidence from extensive primary and reverse co-IP data combined with IFA analysis led us to postulate that GlTom40, MOMTiP-1 and MOMTiP-3 exist in an outer membrane core complex, likely involved in protein import. We probed the functional conservation of mitosomal import across the GlTom40 translocon with respect to the corresponding process in bona fide mitochondria by adapting the DHFR-folate analogue system [49,73] to G. lamblia. Pre-sequence directed DHFR is a classical substrate used in protein translocation studies due to its ability to fold irreversibly upon binding a folate analog, e.g. MTX. Complexed with MTX, DHFR becomes unsuitable as a substrate for import and blocks translocons, which results in a general blockage of organelle protein import [73]. Transfection of MTSfdΔint-DHFR into a Gl17030-HA background, i.e. a transgenic line expressing an HA-tagged MTS-directed mitosomal reporter, allowed testing of the general effects of MTX-induced import block. We reasoned that the presence of MTX in MTSfdΔint-DHFR expressing cells could lead to an import block due to jamming of the translocase. Localization of the reporter by IFA showed an increased cytosolic Gl17030-HA signal after addition of 1 μM MTX (Fig 7B) compared to parasites exposed to the solvent alone (Fig 7A). This suggested accumulation of the reporter in the cytosol in cells exposed to MTX as a result of a generalized import block. To test this we measured the ratio of the slightly larger Gl17030-HA reporter precursor protein and the imported and therefore processed form without the MTS by SDS-PAGE and Western blot using anti-HA antibodies. Consistent with the IFA data, unprocessed Gl17030-HA was strongly increased in the drug treated sample, whilst only the processed form was present in untreated controls (Fig 7C). Taken together the data support functional conservation of the highly diverged protein import machinery in G. lamblia mitosomes.
G. lamblia mitosomes remain the smallest known and least characterized MROs. Identification of protein components using shotgun proteomic analysis of enriched mitosome preparations has proven challenging primarily due to difficulties in isolating sufficient amounts of contaminant-free organelles [33,48]. Extensive sequence divergence prevents identification of organelle proteins via homology-based searches; a case in point is GlTOM40 whose sequence degeneration is so extensive that the identification of orthologues in Giardia, Entamoeba or Spironucleus remains tentative despite the constraints imposed by the beta barrel structure of these mitochondrial porins [44]. The function of candidate factors identified by other means and localized to organelle membranes usually cannot be deduced based on existing structural information from well-characterized mitochondrial homologs. A notable exception to this is a recently identified highly-diverged but structurally-conserved GlTim44 homologue [49]. Taken together, these challenges have frustrated attempts at systematizing intra- and inter- organelle mitosome-centered interactions, thereby limiting analysis to isolated complexes [49,50]. For example attempts at analyzing isolated GlTom40-containing protein complexes (presumably enriched mitosomal outer membrane translocons) using blue-native PAGE [50] detected no mitosomal proteins aside from GlTom40 and an unidentified 32kDa protein which could not be mapped to any known Giardia sequence. These data demonstrate how challenging it is to define novel GlTom40-interacting partners, probably due to the translocon complex being embedded in the outer membrane of the organelle. Here, we used epitope-tagged GlTOM40 [23,33] as first bait, and implemented an iterative co-immunoprecipitation approach to expand the mitosomal membrane interactome network beyond the few known components. With only 5 more bait proteins, this strategy allowed for building of a core membrane interactome and a complex interactome network extending inwards to the organelle matrix as well as outwards to components of the ER membrane, the axoneme cytoskeleton and the cytoplasm. The rationale is that with sufficient numbers of targeted reverse co-IP experiments using validated organelle proteins as baits, a comprehensive proteome interactome could be built, thereby achieving a systems-biological view of the giardial mitosome proteome. From a technical point of view, building an interactome network using a forward- and reverse co-IP approach allows isolation of “true” mutual interactions by validation in two completely independent co-IP experiments. Specifically, the 108 MS hits detected in the Tom40 co-IP dataset which include numerous non-specific interactions can be filtered with data from reverse co-IP assays to reveal actual protein-protein interactions (depicted in Fig 6E) that can be unambiguously distinguished from false-positive hits. Combined with imaging data and predicted topology this provides a robust platform to construct an integrated working model of all mitosome-associated protein interactome networks known to date (Fig 8). Blast2Go in silico enrichment analyses suggest that mitosomes may have a role beyond Fe-S protein maturation (S3 and S4 Figs) Only recently the major function of E. histolytica mitosomes was shown to be sulfate activation, and not Fe-S protein maturation as previously thought [44]. Although genes involved in this pathway are missing in other MRO-containing organisms such as G. lamblia, T. vaginalis, and C. parvum, the Entamoeba example points to a wider range of functions ascribable to mitosomes. This may even include general functions in stage-differentiation as recently shown in E. histolytica whose mitosomes are essential for the encystation process [74].
Following its identification as a prominent GlTom40 interaction partner, the single pass membrane protein MOMTiP-1 was the first bait protein selected for reverse co-IP to expand the GlTom40 interactome. MOMTiP-1 as bait pulled down GlTom40 with the most abundant peptide counts. GFP-tagging and detection of MOMTiP-1::GFP on mitosomes suggests a membrane topology in line with characterized mitochondrial receptor proteins such as Tom20 [75]. Further support for MOMTiP-1’s membrane topology may derive from a definition of membrane orientation using alternative methods such as in situ proximity ligation or protease protection assays. The latter approach has proven useful in the determination of membrane topology for other mitosomal candidate proteins in Giardia [49]. The identification of MOMTiP-1 provides an exciting lead; however, a detailed functional characterization of this protein is required to provide independent evidence for the exact nature of this interaction and to test the hypothesis that MOMTiP-1 is a component of the GlTom40 complex with a receptor function. So far 20 proteins have been validated by localization to mitosomes, allowing for a significant expansion of the GlTom40/MOMTiP-1 interactome. This protein’s predicted topology combined with its exclusive mitosomal localization and the size and composition of its interactome, supports MOMTiP-1 as a GlTom40 accessory protein with a potential receptor function for protein import. To test this hypothesis, we engineered a truncated HA-tagged version of MOMTiP-1 consisting only of the predicted C-terminal domain (residues 31–133; C-MOMTiP-1). Ectopic expression of C-MOMTiP-1 showed a distinct cytosolic localization by IFA (S5A Fig). Native co-IP of C-MOMTiP-1 and analysis of the bait-specific dataset with medium stringency parameters (95_1_95) identified only 2 mitosomal proteins (Gl50803_16424 and MOMTiP-8) (S5B Fig). These data show that the soluble cytoplasmic MOMTiP-1 variant does not recapitulate the interaction properties of the full-length membrane-anchored protein, suggesting that capture of imported matrix proteins may require incorporation of the putative receptor domain into a fully-assembled TOM complex, complete with ancillary factors.
MOMTiP-3 was exclusively identified in the GlTom40 and the MOMTiP-1 co-IP datasets, suggesting that these 3 proteins may function in a tightly-knit complex, likely involved in protein import across the outer mitosomal membrane. TMHMM predicts two TMDs at MOMTiP-3’s N-terminus, followed by a large C-terminal domain. Powerful HMMER-based searches across several eukaryotic lineages, including the closely related diplomonad Spironucleus salmonicida [76], yielded no orthologues for MOMTiP-3, neither was there any predicted functional information available. Nevertheless, analysis of protease protection assays for this protein showed that MOMTiP-3 localizes at the outer mitosome membrane with its C- terminus in the cytosol [49]. These data, in combination with data on MOMTiP-1 predicted topology and interactomes developed herein, support a model for GlTom40, MOMTiP-1, and MOMTiP-3 for a minimized mitosomal import apparatus whose core import machinery is composed of only these 3 proteins. The dramatic perturbation of mitosomal homeostasis observed when either MOMTiP-1 (this work) or MOMTiP-3 [49] were constitutively overexpressed supports the hypothesis for their belonging to the same complex. Protein translocation across the outer mitosomal membrane through this highly reduced import apparatus would be conserved in its mechanism, given that MTX-induced complexing of mitosome-targeted DHFR caused accumulation of unprocessed i.e. untranslocated mitosome reporters (Fig 7C, [49]). Incidentally, these data also confirm that mitosome membrane translocation requires pre-proteins to remain in an unfolded state [49].
Co-IP data combined with imaging of tagged variants identified 6 proteins with dual localization at mitosomes and ER (Fig 8). Contact between these organelles would serve at least two major functions, i.e. replication of mitosomes and transport associated to lipid biosynthesis. Thus far, we have identified five mitosome proteins with dual localization potentially involved in inter-organelle communication (Fig 4F). One of them is a transmembrane Qb-SNARE 4 (MOMTiP-7) [65] identified in GlTom40 and MOMTiP-1 co-IP datasets.
For their biogenesis, mitochondria and MROs rely on lipid transfer from the ER, the central site for phospholipid synthesis in the cell [77,78]. SNAREs are best known for mediating membrane fusion in vesicular transport [79] whereas in the context of mitochondria and the ER, they function as components of so called ER-mitochondria encounter structures (ERMES). In addition to being associated to mitochondrial protein import [80,81], ERMES fulfills an essential function in inter-organelle lipid transport [80]. Phosphatidylserine is shuttled from the ER to mitochondria through the ERMES complex where it is converted to phosphatidylethanolamine (PE) by a decarboxylation reaction that generates most if not all PE in mitochondria [80,82]. Unlike in the hydrogenosome-containing T. vaginalis [83], ERMES homologs have not been identified in G. lamblia, possibly due to extensive sequence divergence. Thus, whether this function is preserved in Giardia mitosomes is not known however, organelle biogenesis would necessarily depend on ER-derived lipids which are transported to mitosomes either by carrier proteins or via membrane contact sites. The latter requires tethering complexes to facilitate phospholipid exchange between the two organelles. Given that MOMTiP-7 is predicted to be a SNARE, we explored the idea that this protein is part of a larger complex mediating ER-mitosome interaction. Co-IP of MOMTiP-7 specifically detected, in addition to outer membrane proteins such as GlTom40, MOMTiP-1 and MOMTiP-3, 3 hypothetical proteins, two of which, MOMTiP-8 and MOMTiP-5 (both predicted soluble proteins), localized both to the ER and to mitosomes. In addition, a domain in MOMTiP-8 has similarity to a yeast “Maintenance of mitochondrial morphology” protein 1 (Mmm1) of the ERMES complex. Moreover, HHpred analysis revealed a link between MOMTiP-5 and a beta barrel lipid binding protein MLN64 (e-value 0.0006) in H. sapiens which facilitates cholesterol transport to mitochondria [84]. These preliminary data support the existence of an outer mitosomal membrane-associated complex in G. lamblia mitosomes possibly involved in generating ER—mitosome membrane contact sites (Fig 8).
We had previously shown that replication and inheritance of the CMC is coordinated in a cell cycle-dependent manner, whereas PMs divided stochastically [23]. The lack of a system to track organelles in living trophozoites precluded addressing the question directly whether mitosomes were motile and constituted a dynamic network of organelles with measurable exchange. Development of two GFP-tagged reporters GFP-GlTom40 and MOMTiP-1-GFP (this study) allowed for time-lapse experiments to follow individual organelles in a cell. However, we found no evidence for motility of organelles, neither in the CMC nor in PMs, even after prolonged observation (1.5 h). This is consistent with the lack of motor proteins such as kinesins and dyneins in any of the mitosomal protein interactomes we generated. Moreover, FRAP experiments revealed no exchange of GFP-tagged membrane proteins between organelles during the period of observation (Fig 7F and 7G), which further corroborated the relative isolation of mitosomes within the cytosol. The lack of mitosomal motility and contact complicates investigation of their replication and morphogenesis. The two most plausible scenarios for this are currently the following: i) PMs are released from the CMC, which continuously produces new organelles by elongation and fission to maintain a constant number of organelles in a cell-cycle independent manner; ii) PMs and the CMC organelles replicate independently in a cell-cycle independent and -dependent manner, respectively [23]. Although time-lapse microscopy experiments did not provide evidence in support of either scenario, conditional expression of a dominant-negative, constitutively active GlDRP-K43E revealed a distinct morphogenesis phenotype (see also below) indicative of an organelle replication defect. As one of the key players in the regulation of mitochondrial fission, DRPs are mechano-enzymes conserved from yeast to vertebrates [85,86,87,88]. G. lamblia harbors a single dynamin homologue GlDRP shown to play a major role in this parasite’s endocytic pathway and stage conversion [56,89,90]. Transgenic parasites expressing the GlDRP-K43E variant exhibited larger and fewer mitosomes, compared to cells expressing the wild type GlDRP variant(Fig 6). This is in line with the dominant-negative effect on mitochondrial fission elicited by the corresponding mutation in DRPs in other organisms. To our knowledge, this is the first report on the involvement of GlDRP in mitosome homeostasis, supporting the (at least partial) functional conservation of mitochondrial and MRO fission [91,92,93,94]. The notion that G. lamblia mitosome fission is functionally conserved is further substantiated by the identification of MOMTiP-6 which presents dual localization to mitosomes and the ER. HMMER-based predictions relate MOMTiP-6 to human mitochondrial fission protein (Fis1, e-value 6.3E-05) which participates in the recruitment of dynamin-related protein 1 (Drp1) to the mitochondrial surface for organelle fission [95,96]. The distinctive “string” mitosome phenotype in cells expressing MOMTiP-1-GFP clearly demonstrated that mitosomes can assume an elongated, tubular morphology, which is a prerequisite for organelle division and replication. The implication is that G. lamblia mitosomes retain at least the machinery for fission in which the mechano-enzyme GlDRP and outer mitosomal membrane elements such as MOMTiP-1 and 3 [49] play central roles.
We used an iterative approach based on co-IP experiments to generate a GlTom40-centered interactome network. Ultimately this strategy should allow building a combined proteome, which delineates the full complement of organelle proteins, peripherally associated factors, as well as interfaces with the ER and the cytoskeleton. Although this strategy requires numerous rounds of sequential co-IP and validation, it is highly informative because it produces interaction data in addition to identifying novel organelle proteins. Combined with testing of epitope-tagged variants of candidate proteins for organelle localization as a straightforward validation criterion, serial co-IPs allow for unambiguous definition of the organelle-specific proteome, as well as interfaces with other cellular structures. This strategy also led to the discovery of MOMTiP-1, a strong GlTom40 interaction partner which plays a role in mitosomal morphogenesis. Together with GlDRP (this work) and MOMTiP-3 (MOMP35; [49]), these are the only proteins so far known to affect mitosomal homeostasis in G. lamblia.
|
10.1371/journal.pntd.0003405 | Long-term Survival and Virulence of Mycobacterium leprae in Amoebal Cysts | Leprosy is a curable neglected disease of humans caused by Mycobacterium leprae that affects the skin and peripheral nerves and manifests clinically in various forms ranging from self-resolving, tuberculoid leprosy to lepromatous leprosy having significant pathology with ensuing disfiguration disability and social stigma. Despite the global success of multi-drug therapy (MDT), incidences of clinical leprosy have been observed in individuals with no apparent exposure to other cases, suggestive of possible non-human sources of the bacteria. In this study we show that common free-living amoebae (FLA) can phagocytose M. leprae, and allow the bacillus to remain viable for up to 8 months within amoebic cysts. Viable bacilli were extracted from separate encysted cocultures comprising three common Acanthamoeba spp.: A. lenticulata, A. castellanii, and A. polyphaga and two strains of Hartmannella vermiformis. Trophozoites of these common FLA take up M. leprae by phagocytosis. M. leprae from infected trophozoites induced to encyst for long-term storage of the bacilli emerged viable by assessment of membrane integrity. The majority (80%) of mice that were injected with bacilli extracted from 35 day cocultures of encysted/excysted A. castellanii and A. polyphaga showed lesion development that was similar to mice challenged with fresh M. leprae from passage mice albeit at a slower initial rate. Mice challenged with coculture-extracted bacilli showed evidence of acid-fast bacteria and positive PCR signal for M. leprae. These data support the conclusion that M. leprae can remain viable long-term in environmentally ubiquitous FLA and retain virulence as assessed in the nu/nu mouse model. Additionally, this work supports the idea that M. leprae might be sustained in the environment between hosts in FLA and such residence in FLA may provide a macrophage-like niche contributing to the higher-than-expected rate of leprosy transmission despite a significant decrease in human reservoirs due to MDT.
| Leprosy is a progressive disease of the skin and nervous system caused by the bacillus, Mycobacterium leprae. Implementation of multiple drug therapy (MDT) for leprosy has significantly reduced the global cases of leprosy. Currently, only a few endemic countries remain where relatively high number of cases persists. Despite global reduction of leprosy and the concomitant decrease in human reservoirs, leprosy transmission and incidence have not declined as expected, suggesting a possible extra-human or environmental source of the bacilli. In the current study, we demonstrate that M. leprae can survive long-term within cysts of common environmental free-living amoebae. M. leprae residing in amoebal cysts for over 30 days remain fully capable of transferring disease to mouse footpads and retain viability phenotypes after several months residence within amoebal cysts. It is hypothesized that these protozoa provide an intracellular refuge for M. leprae in environments for which they would otherwise seem ill suited. Traits allowing bacilli to survive in macrophages may likely be acquired via an evolutionary response against predation by amoebae. The results from this work suggest alternative non-human reservoirs for M. leprae exist fostering further study to determine the role of amoebae in the transmission of this Mycobacterium to humans.
| Human beings have been afflicted by leprosy for over a millennium. Leprosy is a chronic granulomatous infection of skin and peripheral nerves caused by the bacillus Mycobacterium leprae. The bacilli are slow growing obligate intracellular organisms trophic for macrophages, dendritic cells (DC) and Schwann cells in peripheral nerves. The scientific community has reached a generally accepted consensus that M. leprae is principally a parasite of humans and is spread primarily thereby [1]. In addition, there have been autochthonous cases of leprosy among native-born Americans in the southern region of the United States with no prior history of foreign exposure. In the same regions, wild armadillos are infected with M. leprae. A unique M. leprae genotype had been found in the majority of armadillos that was identical to U.S. patients who resided in areas where exposure to armadillo-born M. leprae was possible [2]. This is highly suggestive of the fact that armadillos are a significant natural reservoir for the bacilli and, leprosy might be a zoonosis in the these areas. There has also been a substantial history of studies, anecdotal evidence, rationalizations and opinions that argue in favor of additional non-human sources of the bacillus [3]. What is more intriguing is that, despite many years of using multidrug therapy (MDT) resulting in a significant reduction in disease prevalence, transmission remains stubbornly high implicating among other issues, ineffective detection of early infection, case reporting deficiencies or a lack of a thorough examination of potential environmental sources of the bacillus [3], [4].
M. leprae is an extremely fastidious organism that, despite over 100 years of endeavor, has not been successfully cultured in artificial medium [5]. It is, thus, classified as an obligate intracellular organism with an evolutionarily minimized genome that is believed to have constrained its growth to the intracellular niche. With such a stringent requirement for survival, several questions remain as to how the bacillus remains viable and infectious between human hosts. Are there environmental elements that are capable of sustaining viable M. leprae for long periods or are these bacilli dependent on close-quartered conditions necessary for aerosol transmission from human to human? is M. leprae harbored in soil and water niches? is the bacillus sheltered and capable of surviving intracellularly in ubiquitous protozoa such as free-living amoebae (FLA) that provide similar micro-niches as human macrophages? Evidence of an environmentally sustainable entity for M. leprae would certainly explain, in part, the apparent lack of reduction of the rate of transmission of leprosy in spite of successful MDT [6], [7].
The nature of the relationship between most intracellular organisms and host FLA is currently not defined. The terms “endosymbionts”, or “symbionts” fail to adequately describe these complicated interactions. It is currently proposed to define intracellular microorganisms that associate with FLA without any known directional host/bacterial benefit as “endocytobionts” [7] [8] [9]. Over the past three decades, numerous studies have reported that microorganisms can survive as endocytobionts in FLA. It was reported in 1980 that Acanthamoeba harbored Legionella pneumophila and that the bacterium resisted phagosome-lysosome fusion and multiplies within the amoebae [8], [9]. This latter work implicated infected amoebae as a source of Legionnaire's Disease. Additionally, there are numerous reports describing infection of Acanthamoeba FLA with both pathogenic and environmental mycobacteria such as M. avium subsp. paratuberculosis, M. avium-intracellulare, and M. bovis [10]–[13]. In a study involving hospital networks, FLA such as A. polyphaga, and Hartmannella vermiformis were associated with many species of mycobacteria in water specimens including M. gordonae, M. xenopi, M. avium and M. kansasii subtype 1 lending to much circumstantial speculation regarding the means to which mycobacteria have adapted to environmental persistence [14], [15].
The evolutionary response to amoebal predation is the acquisition of traits that confer resistance to digestion in food vacuoles of amoebae [16]. Many Mycobacterium species survive and even thrive intracellularly in protozoa [16], [17]. As has been known for many years, mycobacteria have a rich hydrophobic cell wall and, as such, lend themselves quite well to attachment to cellular surfaces and are efficiently phagocytized by macrophages [18] and protozoa [19]. Many elements of the mycobacterial cell wall contribute to efficiently enable an active entry of the bacterium into phagocytes [20] [21] [22]. Furthermore, protozoa possess the remarkable ability to transform into cysts protecting them from harmful and often times rapidly fluctuating environmental influences such as extremes in temperature, drought and a spectrum of biocides [23]. Mycobacteria, in turn, can use the nutrients of protozoa as a food source and their intracellular life offers protection against the potentially harmful extracellular milieu. This poses the interesting question as to whether amoebae provide an environmental niche simply for persistence or are a selective proving ground enhancing virulence. Additionally, the dual lifestyles of amoebae (trophozoite vs. cyst) likely provides a survival niche to fragile, fastidious microbes such as M. leprae when the bacillus is subjected to relatively harsh environments such as those between hosts.
Few studies have investigated whether mycobacteria infect amoebae in their natural environment. Thus, an inherent resistance to predation by amoebae likely has important consequences since bacteria that infect and evade amoebal digestion might exploit these traits to enter and resist destruction within macrophages or dendritic cells (DCs) thus thwarting or altering innate immune responses [16], [24]. FLA are environmentally ubiquitous and most are non-pathogenic to immune-competent humans. Delivery of pathogenic mycobacteria within non-pathogenic amoebae to cells of the innate immune system will likely elicit alternative host immune response in comparison to that generated against the Mycobacterium alone. This endocytobionic relationship between the somewhat weakly pathogenic bacteria and ubiquitous amoebae and the potential to aid transmission to susceptible host is of great concern to human, animal and ecosystem health.
In the present study we show that M. leprae remains viable up to 8 months as determined by the accepted criteria of assessment of membrane integrity by viability staining in 3 species of Acanthamoeba (A. lenticulata, A. castellanii and A. polyphaga) and 2 strains of Hartmannella vermiformis. Additionally, M. leprae extracted from cocultures of A. castellanii and A. polyphaga that were induced to encyst with the phagocytosed bacilli for 35 days remained viable causing infections and M. leprae proliferation in Foxn1nu/Foxn1nu (nu/nu) mouse footpads (FP). This works shows for the first time that cysts from amoebae representing species from both Acanthamoeba and Hartmannella genera are capable of supporting the viability of M. leprae, a bacillus so fastidious that it has never been successfully cultivated axenically. The implications of this work relate to the environmental sustainability of M. leprae in the context of persistent transmission despite a vastly reduced human reservoir of infection.
All mouse work was conducted according to relevant U.S. and international guidelines. The procedures for isoflurane anesthesia, infection of nu/nu mouse FPs with M. leprae and fine needle aspirate (FNA) biopsy are Institutional Animal Care and Use Committees (IACUC)-approved protocols (protocol # 12-3613A and 11-3037A) that are approved/renewed yearly by an institutional review board of certified veterinarians and selected faculty. The mice are monitored twice weekly by trained animal laboratory technicians employed by our Laboratory Animal Resources (LAR) center. Any maladies, whether directly, indirectly or unrelated to the protocol are reported immediately to both the attending veterinarian and the PI (WHW) holding the approved protocol. The committee is in compliance with the U.S. Public Health Service Policy on Humane Care and Use of Laboratory Animals.
Stocks of axenic Acanthamoeba lenticulata ATCC 30841, Acanthamoeba castellanii ATCC 30232, Acanthamoeba polyphaga CCAP 1501/18, Hartmannella vermiformis ATCC 50237 and Hartmannella vermiformis CHUV 172 were obtained from the American Type Culture Collection (Manassas, VA) and STERIS SA R&D Fontenay-aux-Roses, France. Amoebae stocks were derived from several sources as diverse as ATCC and hospital and city water supplies and were cultivated to axenic stocks using standardized methods [14], [25]. Acanthamoeba trophozoites were axenically maintained in culture in 1X PYG medium which consists of Page's amoebae saline (PAS) [60mg NaCl, 2mg MgSO4·7H2O, 68mg KH2PO4, 71mg NaHPO4 and 2 mg CaCl2 in 500 ml dH2O (pH = 6.9)] to which 1/10 volume of 10XPYG solution [50 g Proteose Peptone (Difco); 5 g yeast extract (Difco); 2.45 g MgSO4·7H2O; 2.5 g Sodium citrate·2H2O; 0.05 g ammonium iron sulfate (NH4)2Fe(SO4)2·6H2O; 0.85 g KH2PO4; 0.89 g Na2HPO4·7H2O; 22.5 g α-D-glucose; 0.295 g CaCl2 in 250 ml dH2O] was added [26] [14]. Hartmannella trophozoites were cultured in modified PYNFH (ATCC medium 1034) medium.
Viable M. leprae was obtained from the National Hanson's Disease Programs, Baton Rouge, LA. Trophozoite monolayers of A. lenticulata, A. castellanii, and A. polyphaga, were maintained at 28°C and passaged in 1X PYG. H. vermiformis str. ATCC 50237, and H vermiformis str. 172 were maintained at 28°C and passaged in PYNFH medium. Amoebae were infected with viable M. leprae at a bacilli:amoebae ratio of 5–10∶1 and incubated for 48 hr at 32°C. Extracellular bacilli were removed by centrifugation at 600xg and washing the amoebae pellet in HBSS (Hank's Balanced Salt solution) 3 times. For some smaller scale experiments (e.g., for phagocytosis assays), infections were carried out at M.O.I. of between 1 and 100 as well as some cocultures kept at 4°C and aliquots withdrawn every hour to determine adsorption of PHK26-labeled bacilli (see below) to FLA by flow cytometry. FLA- containing M. leprae were induced to encyst by pelleting the cultures and subsequently suspending in encystment buffer (0.1M KCl, 0.02M Tris-HCl pH 8.0, 8 mM MgSO4, 0.4 mM CaCl2 and 1mM NaHCO3). Intracellular M. leprae was extracted from amoebae cysts maintained at 32°C at various times (one week, two weeks, 35 days, 45 days, 3 months and ≥6 months). Prior to extraction of bacilli, long-term encysted cocultures were induced to transform back to trophozoites in complete growth media at each of the above time points. Bacilli extracted from excysted trophozoites by suspending the pellet in 100 µl of sterile PBS containing 0.5% SDS, vigorously vortexing and washing three times with PBS were then processed for viability using BacLight staining procedure (Molecular Probes; Life Technologies, Grand Island, NY), fluorescence microscopy or injection into mouse FPs.
Athymic FoxN1nu/FoxN1nu (designated as “nu/nu” throughout this manuscript) mice, five in each group, were challenged in the plantar surface of the left hind foot with M. leprae bacilli extracted from A. castellanii or A. polyphaga cysts as described [27]. Mice were injected a total of 3 times every other week for one month. All bacilli used in experimental FP injections were extracted from 35-day encysted A. castellanii or A. polyphaga cocultures. This three-time injection scheme was performed because the bacillary yield from the extraction process seemed rather low and would ensure a relatively timely appearance of FP induration. The inocula were estimated based on direct counting of bacilli. [28].
Mice were anesthetized by inhalation of 5% isoflurane. Once fully anesthetized, infected mouse FPs were aspirated using a 0.5 cm, 23-ga needle syringe inserted subcutaneously into the infected area of the FP. Portions of the samples were prepared for microscopy by acid-fast staining or for nucleic acid extraction for PCR analysis. This procedure was performed monthly for 6 months.
The Thai-53 isolate of M. leprae was maintained in the footpads of athymic nu/nu mice infected for approximately 6 months, and then harvested as described previously [27]. Extracted bacilli were washed by repeated (2X) suspension and centrifugation in RPMI-1640 (Gibco) containing 10% fetal bovine serum ((FBS) Gibco). Bacilli were enumerated by direct counting according to Shepard's method [29]. M. leprae suspensions were purified by NaOH treatment as described [27]. Briefly, 1 X109 fresh M. leprae were suspended in 1 ml of 0.1N NaOH and incubated for 3 minutes at room temperature to remove animal tissue. The bacteria were subsequently washed 3X in Hanks Balanced Salt solution (HBSS) and suspended in a final volume of appropriate medium. Freshly harvested viable bacilli were consistently used in experiments within 24–32 hr of harvest. nu/nu mice, five in each group, were challenged in the plantar surface of the left hind foot with 107 M. leprae harvested from passage animal FP as described [27].
Concentrates of M. leprae are separated from infected livers or spleens of 9-banded armadillos (Dasypus novemcinctus). The tissues were collected aseptically and kept frozen at −80°C. Briefly, the procedure for preparation of M. leprae has been described earlier [30], [31] and is carried out at 0 to 2°C. The tissue is homogenized and separated by density gradient centrifugation in sucrose and KCl. The bacilli are disrupted by ultrasonic oscillation. Tissues were treated with trypsin, chymotrysin, collagenase and 0.1N NaOH to remove any host tissue. The bacilli are intact but presumed nonviable due to prolonged storage of the tissue at below freezing temperatures.
For conventional and confocal microscopy and phagocytosis assays M. leprae freshly harvested from FPs were stained with the vital fluorescent red PKH26 dye (Sigma-Aldrich) following the manufacturer's protocol. Briefly, bacilli were stained for 3 minutes at RT in a 1∶250 dilution of dye in Diluent-C (Sigma-Aldrich). The staining suspension was washed three times in PYG containing 5% bovine serum albumin. Bacilli were counted by fluorescence microscopy by averaging several fields counted using a hemocytometer.
Healthy actively dividing amoebae trophozoite cultures were seeded in 6-well plates at 3×106/ml containing appropriate growth medium (above). Amoebae were infected with viable M. leprae, or M. leprae isolated from armadillo tissues (presumed non-viable) that were first stained with the red fluorescent vital PKH26 membrane dye. Triplicate M. leprae-infected amoebae cultures were prepared at bacilli∶amoebae ratios of 1∶1, 5∶1, 10∶1, 50∶1 and 100∶1. Cultures were maintained either in a humidified incubator at 32°C or in a cold room at 4°C. Amoebae were harvested at 0 time (at the time of M. leprae challenge), 2 hr, 3 hr, 4 hr, 5 hr and 6 hr post-infection, washed twice to remove extracellular M. leprae and suspended in 300 µl of FACS buffer (PBS +1%BSA) prior to analysis by flow cytometry using a Becton Dickinson FACS Cantos II instrument. Results were gathered from gates of uniform size as determined from uninfected samples. The resulting mean fluorescence intensities (MFI) were acquired as one-color histograms and increases in MFI were plotted against time using GraphPad Prism software. Results are shown as the average MFI of triplicate cultures.
The viability of M. leprae was determined by assessing membrane integrity using the LIVE/DEAD BacLight bacterial viability kit (Molecular Probes; Life Technologies, Grand Island, NY). Bacilli extracted from amoebae cyst cultures were washed in normal saline (0.90% NaCl w/v) and incubated for 15 min at RT with a final concentration of 1.67 mM Syto9 and 18.3 mM propidium iodide (PI). The bacilli were subsequently washed twice in normal saline (NS) and the pellet was suspended in 25 µl of NS and 5 µl was spotted on a glass slide and mounted on a #1.5 cover glass using BacLight mounting oil. The dead and live bacteria were assessed by direct observation of fluorescent red (PI+) and green (Syto9+) bacilli respectively under a fluorescence microscope using appropriate single bandpass filter sets [FITC filter (480 nm excitation/500 nm emission for Syto9); TRITC filter (488 excitation/653 excitation for PI)]. In cases of nuclear staining of amoebae or mouse tissues, a DAPI filter (358 nm excitation/461emission) was utilized.
Auramine/rhodamine was used to visualize acid-fast bacilli (such as mycobacteria) using fluorescence microscopy. Staining was performed as is routine in the laboratory. Briefly, aliquots of M. leprae-infected or uninfected cysts or trophozoites were transferred to microscope slides and heated to 78°C for 30 min. Slides were then stained for 30 min at RT with auramine/rhodamine (Becton Dickenson, Franklin Lakes, NJ). Slides were rinsed with acidified-alcohol (5% HCl/70% isopropanol) followed by staining with Hematoxylin QS (Vector Laboratories, Burlingame, CA) for 5 sec. Slides were rinsed with dH2O and stained with DAPI (200 µg/ml) for 20 min. The slide were washed with dH2O and mounted to cover glasses with Prolong Gold (Life Technologies, Grand Island, NY). Slides were visualized using a fluorescence microscope within 24 hr of preparation/staining.
Both fluorescence and confocal microscopes were used to visualize extracted and internalized M. leprae by all amoebae spp. studied. Fluorescence microscopy was performed with an Olympus IX71 microscope (Center Valley, PA) using Retiga 2000R (Qimaging, Surrey, BC, Canada) and Qcolor3 (Olympus) cameras. Qimaging and Slidebook software (Intelligent Imaging Innovations, Inc., Denver, CO) were used for image acquisition and analysis on a Macintosh G5 dual processor computer (Apple Computer, Cupertino, CA). Confocal microscopy was performed on a Zeiss LSM 510 confocal microscope. To determine the spatial occupancy of M. leprae within amoebae, serial optical sections were imaged of infected amoebae and were taken at 0.2 nm intervals using a 514 nm excitation laser and 560±20 nm emission filters.
Nucleic acid extraction from amoebae cocultures from fine needle aspirate tissue samples was performed using the Qiagen DNeasy kit and PCR amplification was performed on 50 ng extracted DNA using primers that amplify the M. leprae-specific repetitive element (RLEP) [32]. Amplified PCR samples positive for the presence of M. leprae produced a 129 bp product.
To assess growth and counting efficiency of M. leprae in FPs real-time TaqMan PCR assays were performed. M. leprae genomic DNA was obtained from FP tissue homogenates as described elsewhere [33]. Briefly, 200 µl aliquots of tissue homogenates were subject to 3 freeze/thaw cycles, and proteinase K was added to 10 mg/ml and the sample were incubated at 56°C for 2 hrs. The genomic DNA was processed and purified using the DNeasy Kit (Qiagen, Inc, Valencia, CA) according to the manufacturer's directions. Molecular enumeration of M. leprae was determined using purified DNA fractions from each specimen via TacMan technology using primers and a probe for a common region of the RLEP family of dispersed repeats in M. leprae as previously described [33] [32]. The specific sequences of the primers and probe have been described elsewhere [34]. All reagents used in the TaqMan assay were recommended by the manufacturer (PE Applied Biosystems), including AmpErase UNG enzyme and AmpliTaq Gold DNA polymerase. PCR cycling conditions were 40 cycles with 60°C annealing/extension temperature for 60 seconds and 95°C denaturing temperature for 15 seconds. PCR and data analyses were performed on a 7300 RealTime PCR System (Applied Biosystems, Foster City, CA).
To show that amoebae are capable of phagocytosing M. leprae bacilli, we infected amoebae trophozoite cultures with M. leprae at an M.O.I. of 5. To facilitate infection, the trophozoite cultures were shifted from optimized medium to 1/10 the optimum nutrient concentration allowing for a parallel shift from a primarily pinocytotic nutrient acquisition mode to a macro-phagocytic mode that effectively optimizes the trophozoites to take up the bacilli [35]. Fig. 1 shows light microscopy of acid-fast staining of M. leprae in cocultures established by infecting three species of Acanthamoeba with freshly harvested viable M. leprae. Greater than 95% of the amoebae were observed to be internally occupied (Fig. 1; Panels A–C) with at least one acid-fast bacillus residing in the amoebic trophozoites. At relatively low M.O.I. (1∶1 to 5∶1), ingestion of live M. leprae did not exert any observable adverse effect on amoebae that divided normally over several days. At higher M.O.I. (>5∶1), however, the bacterial burden negatively affected the growth of trophozoites and the cocultures showed low-level lysis of amoebae and considerable detachment from plate wells. M. leprae bacilli were also readily taken up by two strains of H. vermiformis (str. ATCC 50237 and str. 172) as well. Fluorescence microscopy of PKH26-labeled M. leprae in cocultures of amoebae that were stained by the DNA-specific dye, DAPI, showed that the bacilli were taken up into areas that were mostly exclusive to nuclear staining indicative of primarily cytoplasmic staining (Fig. 1 D–F).
In order to determine whether the M. leprae bacillus is phagocytosed by amoebae as opposed to being merely adsorbed to the protozoan surface, we prepared cocultures as above and examined the fluorescently labeled bacteria by confocal microscopy following 16 hr of culture (Fig. 2). Confocal microscopy was utilized to resolve the physical location of the M. leprae bacilli within infected amoebae at various focal planes. Layered focal resolution of M. leprae-infected A. polyphaga showed that the best optical and fluorescent resolution of the bacilli was well within the interior of the amoebae suggesting that the bacilli resided within the amoebae interior as opposed to their external surface. Similar resolution was obtained for A. castellanii, A. lenticulata as well as both strains of H. vermiformis.
Amoebae were cultured with live PKH26-stained M. leprae in appropriate amoebae medium for 16 hr to allow for complete envelopment of the bacilli. Following uptake of M. leprae, the cultures were pulsed at 32°C for 2 hr with 100 mM Lysotracker Green-DND-26 (Molecular Probes, Life Technologies, Grand Island, NY) in order to fluorescently stain the acid-rich organelles such as lysosomes residing within the amoebic cytoplasm. Fig. 3 shows that the fluorescently labeled M. leprae infecting either A. castellanii or A. polyphaga resided primarily within acid-rich organelles (i.e. lysosomal compartments) of amoebae similar to what is observed in macrophages, DCs and Schwann cells. The bacilli were similarly located within the cytoplasmic regions of A. castellanii, A. lenticulata, H. vermiformis str. ATCC 50237 and H. vermiformis str. 172.
To determine whether uptake of M. leprae by amoebae requires metabolic viability of either amoebae and/or bacilli, phagocytosis assays were performed that measure the extent of uptake of PKH26-labeled M. leprae by A. lenticulata, A. castellanii or A. polyphaga and the two strains of H. vermiformis. Fluorescently-labeled M. leprae were introduced to actively growing amoebae trophozoites and the extent of acquisition of fluorescence over a 6 hr period was determined by flow cytometry as measured by gain of mean fluorescence intensity (M.F.I.) by the amoebae (Fig. 4). Trial assays showed that the maximal extent of M.F.I. was routinely achieved at 6 hr post-challenge for all amoebae tested. Fig. 4 indicates such for infected axenic cultures of A. castellanii and A. polyphaga (and all amoebae tested (S1 Fig.). Infections of a M.O.I. greater than 100 proved detrimental to the amoebae and demonstrated lower overall M.F.I. per unit time. The acquisition of red fluorescence by the amoebae as a function of time at 32°C is shown in Fig. 4. In all cases, the best acquisition of red M.F.I. occurred if the temperature was 32°C and the infecting M. leprae were viable (i.e. from passaged nu/nu mouse FP). M. leprae freshly isolated from mouse FP and deemed viable by both radiorespirometry and viability staining (>90% viable) proved to be optimally phagocytosed. M. leprae harvested from armadillo tissues, has very low or no viability, [30], [31]. Armadillo-derived M. leprae was not capable of transferring to amoebae the level of red fluorescence achieved by their mouse FP extracted counterparts, achieving only approx. 10% of the maximal M.F.I. (Fig. 4). In addition, assays performed at the sub-physiological temperature of 4°C showed that the amoebae achieved only about 1% of the M.F.I. of viable M. leprae and 10% of M. leprae from armadillo tissue when compared to their respective counterparts at 32°C (Fig. 4). Furthermore, most surface adsorption of the fluorescent M. leprae to the amoebae at 4°C was removed by rigorous washing of the cells. Collectively, these data suggest that uptake and internalization of M. leprae by amoebae optimally requires active amoebae metabolism driving phagocytosis of viable bacilli.
To determine whether the viable bacilli extracted from 35-day amoebae cocultures were capable of causing characteristic M. leprae-induced FP indurations in infected mice, bacillary extracts were injected directly into the left FP of athymic nu/nu mice, and FP pathology was monitored over a period of 8 months. All experimental bacilli that were injected into nu/nu FPs were extracted from cocultures from either A. castellanii or A. polyphaga that remained encysted with M. leprae bacilli at 32°C in amoebae encystment medium for 33 days followed by excystment for 2 days as described above. The appearance of FP lesions in the nu/nu mouse model for leprosy is typically very slow with measurable swelling appearing only after 4–5 months post challenge (using an infecting dose of 2–5×107 bacilli/FP)[37]. Since the number of bacilli that were extracted from excysted cultures were considerably lower than the amount extracted directly from FP, we chose to inject FPs with coculture-extracted bacilli every other week for a total of three times in order to decrease the time of emergence of FP symptoms. As positive controls, five mice were challenged in the FP with 107 M. leprae bacilli freshly extracted from infected nu/nu FPs in a manner that is routinely performed to passage M. leprae in the laboratory (Fig. 9A, panel 1). Mouse FPs were also injected with M. leprae kept in amoebae medium alone for identical periods of temperature and days (Fig. 9A, panel 2). FP swelling was assessed monthly using a Vernier digital caliper and plotted as illustrated in Fig. 9 (Fig. 9A; panels 1–4). During the 6.5 months post-challenge, measurable swelling of the left FP was consistently evident in the positive control animals. In contrast, measurable swelling of the FPs challenged with bacilli extracted from A. castellanii and A. polyphaga cocultures was not detectable until 7.5–8 months post-challenge but the rates of swelling were similar to early stages of the positive controls (Fig. 9A; compare panels 1 with 3 and 4). There was no detectable FP swelling in animals that were injected with M. leprae kept for 35 days at 32°C in amoebae medium alone (Fig. 9A, panel 2; compare photo insets). Any increase in measurement of right FPs or FPs injected with M. leprae in medium alone was due to increasing size of the FP because of the overall growth and development of the animal over the duration of the experiment. These results thus indicate that bacilli extracted from long-term (35 days) cocultures of A. castellanii and A. polyphaga are capable of growth in nu/nu mice FP (albeit with a 2 month delay) similar to M. leprae extracted conventionally from donor nu/nu mouse FPs.
To show that the swelling measured above contained a considerably high burden of acid-fast bacilli, small samples of tissue were extracted by FNA biopsy from challenged FPs. Smears were stained for subsequent fluorescence microscopic analysis with DAPI for cell nuclei and auramine/rhodamine for acid-fast bacilli. Fluorescence micrographs show considerable (red staining) acid-fast bacilli in FP tissue obtained from all of the mice that were challenged with M. leprae derived from freshly harvested passage mice (positive controls) (Fig. 9B; panel i) and in 4 out of 5 of those from each category challenged with bacilli extracted from 35 day cocultures of A. castellanii or A. polyphaga (Fig. 9B, panels iii and iv). There was no evidence of acid-fast bacilli in FNA-extracted tissue from mice FPs challenged with M. leprae from axenic cultures without amoebae (Fig. 9B, panel ii). These results suggest further that FP lesions in these mice were the direct result of viable M. leprae extracted from long-term amoebic cocultures.
To further confirm that the acid-fast bacilli observed in tissue was indeed M. leprae, nucleic acid was extracted and tested in PCR analysis for amplification of the M. leprae-specific RLEP sequence as in Fig. 8 above. Positive PCR signals were obtained from FNA tissue from all mice challenged with M. leprae directly from passage mouse FPs and 4 out of 5 of the mice challenged with bacilli extracted from cocultures of A. castellanii or A. polyphaga (Fig. 10). The one mouse FP in each group that was negative by PCR was also negative by auramine-rhodamine staining for acid-fast bacteria. By contrast, after 8 months post-challenge, there was no evidence of the M. leprae RLEP PCR signal in any of the FPs from mice challenged with M. leprae maintained in axenic cultures of amoebae medium. PCR analysis of FNA tissue for amplification of Acanthamoeba-specific 18S rRNA sequences [38] was negative as well, suggesting that either the extraction method for obtaining the bacilli effectively killed the amoebae or the mice successfully resolved any residual amoebae infection over the long-term experimental period. This confirms that the acid-fast bacilli observed in FP lesions produced by challenge with bacilli extracted from amoebae are most likely M. leprae and that the bacilli remain viable and capable of transmitting FP pathology for up to 35 days in cocultures of two different Acanthamoeba spp.
The observation that there is a considerable increase in the number of acid-fast bacilli in FPs (Fig. 9) coupled with the strong PCR signal obtained from primers for the M. leprae-specific RLEP (Fig. 10), strongly suggested the presence and growth of M. leprae in FPs challenged with amoebae-derived bacteria. However, to confirm that the 35 day-amoebae cocultured M. leprae are indeed capable of replication within the FP, we performed a TaqMan quantitative PCR analysis for the RLEP region of M. leprae to compared the number of M. leprae present in FPs challenged with bacilli derived from amoebic cocultures vs. the number from FPs challenged with M. leprae grown alone for 35 days in amoebae medium. Fig. 11 shows the results of the Taqman analysis. The amount of M. leprae extracted from FP challenged with coculture-derived bacilli exhibited a significant increase of the RLEP signal when compared to those challenged with M. leprae kept in medium for 35 days. The data represents a 3–3.5 log increase M. leprae in FPs challenged with M. leprae from amoebae cocultures compared to those challenged with M. leprae maintained in axenic cultures medium. These data confirm that M. leprae extracted from 35 day encysted amoebae cultures are capable of replication in the nu/nu mouse footpad.
The precise manner in which leprosy is transmitted is unknown. Until recently it was widely believed that the disease was transmitted by proximal contact between untreated or asymptomatic cases of leprosy and healthy people. Currently, the possibility of transmission by the respiratory aerosol route has gained considerable interest [39]. Other means such as transmission through insects [40] [41] has been considered but there has not been any substantial evidence supporting that claim. The possibility of discharge of M. leprae from the nasal mucosa begs the question of how the discharged organism remains viable in between hosts. Since M. leprae is a fastidiously obligate intracellular bacterium, it would be reasonable to assume that it could find safe refuge in the environment by interaction with ubiquitous free-living organisms with physiological semblance to human phagocytes. Recently, it was shown that M. leprae could be taken up by FLA, survive and remain viable intracellularly in these protozoa for a period of at least 72 hr [42]. In the current study, we demonstrate that M. leprae can survive and remain virulent for at least 35 days within amoebal cysts from both A. castellanii and A. polyphaga as determined by their ability to transfer infection to recipient nu/nu mouse FPs. Furthermore, we show that acid-fast bacilli extracted from M. leprae/amoebae cocultures with A. lenticulata, A. castellanii, A. polyphaga, H. vermiformis str. ATCC 50237 and H. vermiformis str. 172 remain viable for over 8 months in encysted amoebae as determined by viability staining of bacilli in situ within cysts or from the those extracted from the cysts. These data provide a proof of concept that M. leprae can be phagocytized and lysosomally occupy common environmental FLA trophozoites, survive encystment while remaining viable and are fully capable of infectivity under suboptimal conditions endured by the amoebic cyst. Although M. leprae has been shown to be approximately 30% viable in terms of membrane integrity by BacLight after two weeks in optimized medium [36], survival in either amoebae medium described here is very detrimental to axenic M. leprae and necessitates refuge within amoebae. It can be reasonably argued that possible environmental reservoirs for this fastidious bacillus are common FLA.
M. leprae cannot be cultured, and therefore there are limited means available to ascertain its viability in long-term amoebae culture. No single method of investigation can be utilized to confirm its viability in amoebic cysts. For example, the specificity of serological techniques and PCR can be impaired by antigenic cross-reactivity and PCR contamination, respectively. Therefore, we embarked on an exhaustive set of experiments in order to be absolutely certain that M. leprae can be engulfed and remain viable within the amoebae examined and to be confident that the acid-fast organisms detected in FPs from mice challenged thereof were indeed M. leprae from amoebae cocultures. We built upon this conclusion by first demonstrating that M. leprae is phagocytized by amoebae as determined by microscopic and flow cytometric analysis revealing that optimal uptake requires active temperature-dependent metabolism and viability of both organisms. Subsequently, we assessed bacillary viability by use of a two-color, Syto9/propidium iodide fluorescence staining that scores for membrane damage in individual bacilli and is a proven technique that correlates well with other viability measures for M. leprae such as radiorespirometry [36]. Virtually all of the bacilli extracted from long-term cocultures at 32°C were propidium iodide negative (PI-) and Syto9+ indicative of viable organisms. Control cultures containing M. leprae alone in amoebae medium showed considerable degradation of the bacilli that was virtually 100% propidium iodide positive after two weeks and were undetectable after 35 days at 32°C (Fig. 5). We also observed occupancy of acid-fast organisms within amoebal cysts for 8 months post culturing. The extracted bacilli emerged either as intact extracellular bacilli or residents of acid-rich organelles of recently excysted trophozoites (Figs. 5 and 7). Most emergent bacilli were deemed viable by Syto9+/propidium iodide negative staining. PCR analysis of nucleic acid from encysted cocultures amplified the M. leprae-specific RLEP element strongly supporting the fact that the acid-fast bodies observed in cysts were indeed M. leprae. The loss of RLEP PCR signal in 35 day axenic M. leprae cultures is curious since M. leprae DNA has been known to persist in tissues for very long times after host death. The loss of signal in these cultures is likely due to release of soluble nucleic acid from the dead bacilli in the liquid medium that is lost when washing. In addition, detection of RLEP in archeological samples is performed using the more sensitive TaqMan PCR methodology. Most importantly, M. leprae survival and retention of virulence in cocultures were confirmed by transference of extracted bacilli from 35-day cocultures of A. castellanii and A. polyphaga into nu/nu mouse FPs. 80% of the mice (4 out of 5) challenged with either coculture developed FP swelling with histological evidence of acid-fast bacilli. Collectively, these data confirm that M. leprae can indeed survive for extended periods of time in encysted FLA cultures and is capable of growth in nu/nu mice FP.
The number of viable M. leprae extracted from these cocultures was significantly less than the initial number used to infect the trophozoites. 1.5×107 bacilli were used to infect 3×106 (MOI = 5) amoebae and, based on microscopic field counts, the estimated number of M. leprae harvested from amoebic cysts and injected into FPs was between 105–106 per injection. This may be due to several reasons: i) Only approximately 30% of the trophozotic amoebae were observed to be capable of encystment as the shift in the transcriptional program necessary for this transformation is considerably complicated and incompletely understood [35]. Since incomplete transformation to cysts might impose a restriction on the actual numbers of bacilli housed therein there would be a considerable culling and reduction in the numbers of protected and viable M. leprae. ii) To facilitate processing, bacilli were extracted from cysts that were first induced to excyst. Studies have shown that bacteria residing in Acanthamoeba cysts are generally housed both within the cytoplasm as well as within the cyst walls between the endo- and ectocyst shells as is the case for Acanthamoeba spp. [35]. The extrusion process of emerging trophozoites from cyst wall pores known as ostioles has the potential of leaving a considerable number of bacilli in the cyst wall remnants that may be either unavailable for infection or are pelleted in the slow speed centrifugation steps used in the purification of the extracted bacilli [43]. iii) The extraction process of recently emergent trophozoites in this study involved treatment with SDS. Due to their unique cell wall, mycobacteria can survive relatively long exposures to detergents [44] but the effect of SDS treatment on long-term viability of M. leprae has, to our knowledge, not been determined and may be a factor in reduced viability of extracted bacilli. However, there was no indication of membrane damage in extracted bacilli as assessed by viability staining. iv) There may be a slow loss of viability over time in the amoebae cyst cocultures if the cultures are unable to optimally support the bacilli and the cysts are simply “buying time” for M. leprae. Also it is possible that not all the extracellular bacteria were capable of being endocytosed by the amoeba. It has been shown that the inevitable clumping that occurs in cell-free Mycobacteria suspensions contain aggregates that are not efficiently taken up by either amoebae or macrophages [11], [16], [45]. Regardless, the encystment of the bacilli prolongs viability empirically. Molecular enumeration, and analysis of viability transcripts by reverse transcriptase-based quantitative real-time PCR will likely provide some insight into the longevity of the bacteria in cysts. Recently, transcripts encoding the M. leprae-specific ESAT-6, heat-shock protein 18, superoxide dismutase A and the 16S rRNA subunit have been determined to be sensitive viability indicators for M. leprae [28], [34]. Many of these approaches are planned in future endeavors.
Prior observations that amoebae can house and transport L. pneumophila and can serve to increase the virulence of M. avium have raised concern that protozoa have the potential to be general environmental reservoirs or vectors of human pathogens [11], [45]. It has been considered that adaptation of organisms to parasitism, commensalism or simply endocytobiontism of FLA might have molded environmental microorganisms to infect and persist in human phagocytes. That is to say, the process of the selection of environmental microorganisms for resistance to digestion by predatory FLA behaving as feral macrophages might be a driving force in the evolution of pathogenic environmental bacteria. Such a process may likely be the “missing link” between ecology and pathology [46]-[48]. FLA are present worldwide [49] and have been isolated from soil [50]–[53], water [54]–[58], air [59], and the nasal mucosa of otherwise healthy human volunteers [60]–[62]. The fact that there are repeated observations of clinical leprosy in those that appear to have no history of exposure to known cases [63]–[66] and that leprosy tends to cluster in areas proximal to water sources [67], [68] strongly suggest that M. leprae has extra-human environmental sources [69], [70] and those environs are also compatible with the globally ubiquitous FLA.
The natural environmental landscape for amoebae (as is for most organisms) is not static and, as such, various adaptive genetic programs have evolved to survive dynamic and potentially detrimental conditions. Exposure to suboptimal conditions such as starvation, extreme temperatures, excessive UV light, radiation, pH changes, as well as exposure to biocides induce amoebae trophozoites to undergo encystation [16]. Amoebae exist in the environment cyclically transforming from free-feeding trophozoites to highly dispersible and resilient cysts. M. leprae, by virtue of having a slow generation time can likely withstand the confines of the amoebal cyst allowing bacillary viability to persist longer in this manner. Moreover, the coculture of M. leprae with Acanthamoeba or Hartmannella spp. is particularly suited for both bacteria and amoebae since their temperature optima are compatible for both initial infection of trophozoite and long-term “storage” of cysts.
It has been shown that M. leprae can remain viable if lyophilized in the presence of 10% skim milk-water [71]. In addition, viability was preserved up to 4 years at 4°C. This work also demonstrated clearly that the composition of the solution for suspending the bacilli was critical for the maintenance of M. leprae viability by lyophilization—with skim milk being 100-fold more effective that water or water with 10% fetal calf serum. With respect to viability in amoebic cysts, these results are intriguing. While desiccation may provide a means of survival/viability to the bacillus, it is unlikely that drying per se is a “natural” means of persistence since most if not all of the remaining endemic areas are those of high humidity and abundant water. The amoebic cyst (in particular the acanthamoeba cyst) is a very efficient desiccant that is essentially devoid of water. The natural "arid" environment inside of cysts allows long-term survival (years) of the amoebae in the face of drought etc. by virtue of its impermeable cellulose wall [35]. Could the same mechanisms (dehydration etc.) that provides viability to the amoebae be "hijacked" by the captured M. leprae to provide long-term viability to the bacillus? Future experimentation will likely reveal answers to these rather intriguing questions.
It would be intriguing to determine whether M. leprae, by virtue of residing in cysts, has evolved its own dormancy program in order to persist and maintain or enhance viability or virulence. Also, active prompting of protozoan encystment by bacteria has thus far only been demonstrated for L. monocytogenes suggesting that these bacteria have a selective advantage of exploiting the cysts' ability to serve as vehicles and to assume dormant stages that aid dispersal in the environment [48], [72]. Whether this is the case for M. leprae as well awaits further investigation. Other outstanding questions include the determination of (A) whether M. leprae, by virtue of being transmitted via amoebae, can enter host macrophages via a Trojan horse mechanism thereby changing the overall pathogen-associated molecular pattern (PAMP) presented to innate immune cells and subsequently altering innate and adaptive responses to the benefit of the pathogen; (B) whether M. leprae is capable of multiplying within amoebae or is simply maintaining survival therein. The results described in this current work do not demonstrate that M. leprae is capable of multiplying inside amoebae but are suggestive of a role for FLA providing sustenance to maintain viability of the bacilli; (C) whether the leprosy bacillus requires periodic excystment as is likely the case in the natural world in order to re-infect emergent trophozoites or human host cells; and, finally (D) whether the virulence of M. leprae is affected either positively or negatively by its passage through amoebae. Future experimentation including testing for appearance of disease by transferring M. leprae from our other existing cocultures of A. lenticulata and H. vermiformis strains to nu/nu mouse FPs and determining whether mice challenged directly with M. leprae-infected amoebae (cysts or trophozoites) display any differences in progress to disease should resolve some of these issues.
In summary, we show that M. leprae is capable of prolonged survival in three common and ubiquitous species of Acanthamoeba and two strains of Hartmannella. At this point we are unsure of whether this endocytobiotic relationship in nature serves to allow some FLA to function as transmission vehicles/vectors, a Trojan horse and/or biological reservoirs for M. leprae. It will be fascinating to determine whether FLA in general provide an environmental sanctuary possibly facilitating virulence and contributing to microbial survival in harsh conditions along with aiding transmission to susceptible hosts. Future experimentation will clearly unravel these issues.
|
10.1371/journal.pcbi.1000940 | Differentially Expressed RNA from Public Microarray Data Identifies Serum Protein Biomarkers for Cross-Organ Transplant Rejection and Other Conditions | Serum proteins are routinely used to diagnose diseases, but are hard to find due to low sensitivity in screening the serum proteome. Public repositories of microarray data, such as the Gene Expression Omnibus (GEO), contain RNA expression profiles for more than 16,000 biological conditions, covering more than 30% of United States mortality. We hypothesized that genes coding for serum- and urine-detectable proteins, and showing differential expression of RNA in disease-damaged tissues would make ideal diagnostic protein biomarkers for those diseases. We showed that predicted protein biomarkers are significantly enriched for known diagnostic protein biomarkers in 22 diseases, with enrichment significantly higher in diseases for which at least three datasets are available. We then used this strategy to search for new biomarkers indicating acute rejection (AR) across different types of transplanted solid organs. We integrated three biopsy-based microarray studies of AR from pediatric renal, adult renal and adult cardiac transplantation and identified 45 genes upregulated in all three. From this set, we chose 10 proteins for serum ELISA assays in 39 renal transplant patients, and discovered three that were significantly higher in AR. Interestingly, all three proteins were also significantly higher during AR in the 63 cardiac transplant recipients studied. Our best marker, serum PECAM1, identified renal AR with 89% sensitivity and 75% specificity, and also showed increased expression in AR by immunohistochemistry in renal, hepatic and cardiac transplant biopsies. Our results demonstrate that integrating gene expression microarray measurements from disease samples and even publicly-available data sets can be a powerful, fast, and cost-effective strategy for the discovery of new diagnostic serum protein biomarkers.
| Protein biomarkers in the blood are urgently needed for the diagnosis of a wide variety of diseases to improve health care. We aim to find a fast and cost-effective strategy to discover diagnostic protein biomarkers. Hundreds of diseases have already been investigated using microarray technology, measuring the mRNA expression of all genes in the disease-damaged tissues. We analyzed biopsy-based microarray data for 41 diseases in the public repository, identified genes with dysregulated mRNA expressions and detectable-protein abundance in the blood, and predicted them as candidate diagnostic protein biomarkers. We found that clinically and preclinically validated diagnostic protein biomarkers were significantly enriched in our predicted protein candidates for 22 diseases. We then measured the concentrations of ten predicted protein biomarkers in the serum samples from 39 renal transplant patients. Three of them were confirmed to be diagnostic of acute rejection after renal transplantation. All three proteins were further confirmed to be diagnostic of acute rejection in 63 cardiac transplant recipients. Our results show that publically available genome-wide gene expression data on disease-damaged tissues can be effectively translated into diagnostic protein biomarkers.
| The utility of serum and plasma proteomic techniques to find diagnostic biomarkers has received considerable attention and investment in recent years. However, the limited sensitivity of mass spectrometers, the dynamic range of protein concentrations, and the presence of high abundance proteins in blood samples are major challenges in the identification and verification of potential protein biomarkers in peripheral blood [1].
Since the development of gene expression microarrays more than a decade ago [2], [3], many microarray studies have been used to study changes in mRNA transcripts in disease-related tissues. Considerable microarray data have been deposited into international repositories including the Gene Expression Omnibus (GEO) [4] and ArrayExpress [5], with at least 30% of US mortality already covered [6]. Integration of publicly-available microarray data has been used to show commonalities across cancers [7], suggest candidate gene variants associated with disease [8], associate relations with studied phenotypes [9], and even to validate gene-expression-based diagnostics for US Food and Drug Administration approval [10]. Although microarrays measure the relative abundance of mRNA transcripts, their translated proteins are also likely to be differentially present in diseased tissue and possibly even secreted or detectable in the blood. Rhodes et al first proposed to predict serum protein biomarkers by integrating cancer gene expression data from Oncomine and filtering the list with Gene Ontology annotation of “extracellular”, “extracellular matrix”, and “extracellular space” [11]. They predicted ten serum protein biomarkers for ovarian cancer and found that PRSS8 had previously been known to be an accurate biomarker for ovarian carcinoma.
We have previously developed a methodology to determine gene expression signatures across 238 diseases from GEO. We have found that the molecular signature of disease-specific RNA across tissues is more prominent than the signature of tissue-specific expression patterns [12]. We hypothesized that RNA measurements in diseased tissue could be used to identify candidate serum protein biomarkers of disease. We also hypothesized that integrating data sets from similar conditions could be used to find protein biomarkers applicable across the related conditions. Finally, we tested whether focusing only on those genes that code for proteins known to be detectable in serum and urine (here termed biofluids), using previously published resources might improve our specificity [13]. We then evaluated the general and specific performance of this Integrated RNA Data Driven Proteomics (IRDDP) method to suggest protein candidates across hundreds of diseases.
One field in urgent need of non- or minimally invasive protein biomarkers is solid-organ transplantation [14]. The diagnosis of acute allograft rejection (AR) is currently based on functional and histological grounds. The latter approach requires an invasive procedure in order to obtain sufficient representative tissue for pathology [15], [16], though blood-based RNA diagnostics are successfully being validated [17]. Additional serum biomarkers are still needed to reduce or avoid invasive diagnostic procedures for AR [16], [18], [19]. Many efforts have been made to identify such biomarkers [20]. For example, in renal transplant rejection, significantly increased protein concentrations of VEGF have been observed in serum and urine [21], [22], [23]. Increases in the following other entities have also been observed: CXCL9 in urine [24], soluble CD44 in plasma [25], ADL in serum [26], and TNF-alpha in serum [27]. HLA class I (ABC) protein levels were recently found to be elevated on the surfaces of peripheral blood CD3+/CD8+ T lymphocytes in AR at 14 and 21 days after renal transplantation [28]. However, there is still no reliable blood-based protein test to diagnose AR and none of these biomarkers has been shown to be universally present across all transplanted organs[14].
At the same time, a previous study also showed that there are similarities in the biology of the processes involved in the rejection of different transplanted solid organs [29]. Informed by these previous successes and efforts, we applied the IRDDP method here to search for blood-detectable proteins for acute rejection across different transplanted organs.
Our first goal was to test the hypothesis that blood- and urine-detectable protein biomarker candidates could be identified by using tissue-based gene expression microarray data. Using previously described methods [12], we acquired gene expression data sets representing 41 diseases, as well as control tissue samples for each from GEO [4], the largest international repository for gene expression microarray data with over 400,000 samples at the time of this writing.
We applied our IRDDP methodology to each disease. First, we calculated a set of differentially expressed genes for each disease using the RankProd meta-analysis package at a percentage of false prediction (pfp) ≤5% [30]. For diseases with multiple microarray data sets, we included genes that were differentially expressed in at least one of the data sets. We then filtered the gene sets through a list of 3,638 proteins with known detectable abundance in serum, plasma, or urine. The list was created from public sources [31], [32], [33], [34] and has been described [13]. This effort yielded a set of candidate protein biomarkers for each disease (Dataset S1).
For each disease, we then compared our candidate biomarkers with known diagnostic protein biomarkers in the GVK BIO Online Biomarker Database (GOBIOM). GIOBIOM is an independent manually curated knowledge base taken from global clinical trials, annual meetings, and journal articles [35]. As of this writing, GOBIOM contains 6,098 known biomarkers for 368 therapeutic indications with 23,166 unique references. For 22/41 diseases, known diagnostic protein biomarkers were enriched in our predicted protein sets (p<0.05, Fisher's exact, Table 1). In 9/11 diseases for which at least three data sets were available, known diagnostic protein biomarkers were even more significantly enriched in our predicted protein sets. The -log(p-value) in diseases with three or more data sets (n = 11) was significantly higher than those in diseases with fewer than three data sets (n = 30; p = 0.004, Fisher's exact, Fig. 1). Of the remaining 19 diseases, 11 were represented by only a single gene expression data set. Therefore, we concluded that the more gene expression datasets for a disease, the more likely known biofluid protein biomarkers are going to be significantly differentially expressed across any one of those data sets, suggesting the likelihood of finding new biomarkers increases with more available data sets. While this finding is not at all surprising, we were able to conclude that joining as few as three experiments could statistically significantly improve the performance to rediscover clinically validated protein biomarkers across 41 diseases.
We then applied IRDDP to the specific problem of finding serum biomarkers for the diagnosis of transplant acute rejection (AR). We integrated three biopsy-based gene expression microarray studies from pediatric renal, adult renal [36], and adult cardiac [29] transplantation, identified genes commonly upregulated in AR compared to stable graft function, and then measured the abundance of proteins encoded by these genes in serum to identify cross-organ AR protein biomarkers (Fig. 2). The first of the three studies was performed in pediatric renal transplantation. It compared gene expression profiles in biopsy samples from 18 AR patients and 18 patients with stable graft function (STA) at the absence of AR and any other substantive pathology (Table S1). Using Significance Analysis of Microarrays [37], we found 2,805 genes with increased expression in AR biopsies (q-value ≤0.05; fold change ≥2).
We combined the results of this study with data from two other transplant studies that we retrieved from GEO. One study compared biopsy samples from 13 AR patients with 19 STA samples after adult kidney transplant (GEO dataset GDS724 [36]). The study yielded 2,316 upregulated AR genes with q-values ≤0.05. The second study compared 12 AR biopsy samples with 13 non-rejection samples after cardiac transplant (GEO series GSE4470 [29]). It yielded 283 upregulated AR genes with q-values ≤0.05. By intersecting the three data sets, we identified a gene expression signature containing 45 genes in common, irrespective of the specific studies or transplanted organs (Table S2). These genes are hereafter referred to as the “common-AR” set of genes.
To evaluate the significance of finding 45 genes in common, we shuffled the gene labels across the three data sets and repeated the entire analysis 100,000 times. In random performance, the number of intersecting genes was normally distributed around n = 9 (Fig. S1), suggesting a false discovery rate of 20%. This result also suggested that the probability of finding 17 or more commonly dysregulated AR genes by chance was less than 1%, and that the probability of finding 24 or more of them by chance was less than 1×10−5.
We next retrieved mRNA expression data for each common-AR gene across 74 tissue and cell types from SymAtlas [38], and identified the cell type with the highest expression. Surprisingly, our common-AR genes were most enriched in CD14+ monocytes (p = 0.003, Fisher's exact). Seven of the 45 common-AR genes had their highest expression levels in CD14+ monocytes: they were CD44, IL10RA, S100A4, IGSF6, CTSS, CASP4, and SCAND2. Our results suggest an important role for activated pro-inflammatory monocytes in transplant rejection. This finding is consistent with recent reports that monocyte/macrophage activation might induce inflammation, leading to impairment of graft function in renal transplant patients [39].
We then analyzed the functions of the 45 common-AR genes using Ingenuity Pathway Analysis. As expected, 28 of the 45 common-AR genes were involved in the inflammatory response (p = 3.37×10−17, Fisher's exact; p<3.56×10−3 after Benjamini-Hochberg multi-test correction). Furthermore, 23 common-AR genes were involved in cell-mediated immune responses, (p = 3.34×10−15; p<2.97×10−3, Benjamini-Hochberg correction). Finally, 23 common-AR genes were involved in a single pathway associated with inflammatory responses, antimicrobial responses, and cellular movement regulated by STAT-1 (Fig. S2).
ELISA kits were available for ten of the 45 candidate proteins, including six proteins known to be in biofluids and four outside. We measured all ten proteins in a pilot study of serum samples collected within 24 hours after biopsy from an independent set of 19 patients with biopsy-proven AR and 20 patients with absence of AR or any other substantive pathology (STA). The patients were from a pediatric and young adult renal transplant study. No patients were positive for BK virus infection, and no patient samples in the ELISA study were matched with samples used in the microarray study. The AR/STA samples were matched for recipient and donor gender, age, type of immunosuppression, time post-transplant, race, and type of end stage renal disease (Table S3).
Three of the ten proteins were statistically significantly upregulated in the AR serum samples compared to the STA samples after renal transplantation (Fig. 3). They were PECAM1 (also known as CD31 antigen, or platelet/endothelial cell adhesion molecule), CXCL9 (MIG, chemokine ligand 9), and CD44 (hyaluronic acid receptor). Mann-Whiney U test for significant differences yielded p-values of 1×10−3, 1×10−4, and 5×10−3, respectively. Receiver Operating Characteristics (ROC) curves showed the ability of each individual protein to distinguish AR from STA (Fig. 3d). The areas under the ROC curves (AUC) were 0.811, 0.864, and 0.761 for PECAM1, CXCL9, and CD44, respectively. At optimal performance, PECAM1 distinguished AR from STA with 89% sensitivity and 75% specificity; CXCL9: 78% sensitivity and 80% specificity; CD44: 80% sensitivity and 75% specificity.
We then measured the concentration of these proteins in a second pilot study on plasma samples of cardiac allograft recipients to identify cross-organ AR biomarkers. We compared samples from 32 AR patients and 31 STA patients. The samples were matched for demographic characteristics (Table S4). None of them was infected with CMV. Interestingly, all three markers were upregulated in AR compared to STA. Mann-Whitney U test for significant differences yielded p values of 3×10−3 (PECAM1), 0.019 (CXCL9), and 4×10−3 (CD44) (Fig. 4). The areas under the ROC curves for distinguishing AR from STA were 0.716, 0.672, and 0.711 for PECAM1, CXCL9, and CD44, respectively.
We evaluated the performance of a combined panel of PECAM1 and CXCL9 using a three-fold cross-validation. We randomly selected two thirds of the samples, trained a multinomial logistic regression model, and calculated the predictive performance on the remaining one third of samples. After repeating the process 1000 times, the average ROC curves showed an improvement on cardiac AR diagnosis and no additional improvement on renal AR diagnosis (Fig. S3), suggesting a large clinical trial combining PECAM1 and CXCL9 with other previously found protein biomarkers would be needed to evaluate the predictive diagnosis of AR. Adding CD44 did not improve the regression models.
We performed an immunohistochemistry study on our best-performing marker, PECAM1. The goal of the study was to compare its protein expression in AR and STA samples from renal, hepatic and cardiac allograft biopsies (Fig. 5). In STA kidney tissue, PECAM1 staining was mainly observed in the endothelial cells of glomeruli, in peritubular capillaries, and in large blood vessels. In contrast, examination of staining patterns in AR biopsies revealed dense infiltrates of PECAM1, as well as positive lymphocytes and mononuclear cells in the interstitium. Similarly, dense endothelial PECAM1 staining was observed in the hepatic and cardiac transplant AR tissues, along with staining in infiltrating mononuclear cells. We observed only minor endothelial staining in hepatic and cardiac STA tissues. These immunohistochemistry results showed significantly increased PECAM1 protein expression in the AR tissues compared to STA tissues across transplanted organs.
Furthermore, our studies showed that PECAM1 protein was also significantly upregulated in the serum samples from AR patients compared with samples from patients with BK virus infection (n = 10, p = 0.001, Mann-Whitney U test) and chronic allograft injury (n = 10, p = 6×10−5, Mann-Whitney U test) after renal transplantation (Fig. S4). Analysis across hundreds of diseases using our GeneChaser tool [40] showed that the mRNA expression of PECAM1 is significantly upregulated in various cancers, but not in other potential confounding conditions, such as infection and hypertension (http://tinyurl.com/yhq9h3k). These results suggest that PECAM1 is a serum marker specific for allograft acute rejection, irrespective of the transplanted organ.
Finally, as mentioned above, 23 of our 45 common-AR genes were involved in a single pro-inflammatory pathway regulated by STAT-1 (Fig. S2). Among the ten proteins we tested by ELISA, five were within this pathway and five were outside of it. All five proteins outside the pathway failed validation, while three of the five proteins inside it were validated as AR markers. The 60% success rate from within this single pathway suggests that it is likely to represent a common functional pathway in AR across transplanted organs. Other novel AR protein markers are likely to be found from the remaining 18 common-AR genes/proteins inside this pathway that have not yet been tested by ELISA (Fig. S2). These proteins include CD2, Cathepsin S, and SH2D2A.
We developed an Integrated RNA Data Driven Proteomics (IRDDP) method, which exploits the link between RNA changes in disease-affected tissue with serum detectable proteins coded by those RNA, yielding candidate proteins diagnostic for those diseases. We have demonstrated that this approach could be used to suggest candidate protein biomarkers for 22 diseases, and have shown the enrichment of known clinically and pre-clinically validated protein biomarkers in these candidate biomarkers. We applied our method to new and publicly-available microarray measurements on solid-organ transplantation, and identified and validated three cross-organ serum protein biomarkers for transplant rejection. Our results demonstrate that the integration of gene expression microarray measurements from disease samples, and even publicly-available data sets, can be a powerful, fast, and cost-effective strategy for discovering diagnostic serum protein biomarkers.
We found that PECAM1, CXCL9 and CD44 proteins were significantly upregulated in the serum/plasma samples of both renal and heart transplant patients with acute rejection compared with patients with stable graft function. The abundance of CXCL9 in urine [24] and that of soluble CD44 in plasma [25] have previously been shown to increase in renal AR compared with STA. In addition, macrophage surface PECAM1 can distinguish lung transplant rejection [41] but is not diagnostic in mouse models of cardiac transplant rejection [42]. But to our knowledge, this study is the first to show all three markers as cross-organ AR protein biomarkers in human serum or plasma. Our best marker, serum PECAM1, identified renal AR with 89% sensitivity, 75% specificity, 26% PPV, and 99% NPV at 9% prevalence[43], suggesting its potential clinical usage to monitor transplant patients to decrease the number of biopsies. We have focused on the biomarkers that were upregulated in AR because that was what most clinic tests are using. Proteins downregulated in AR could potentially be used for diagnosis as well, and we have predicted both up and downregulated proteins for 44 diseases in Table 1 and Dataset S1.
We found that the likelihood of finding protein biomarkers indicative for a disease increases with the number of available gene expression datasets. Many meta-analysis methods have been shown to improve the identification of differentially expressed genes [44]. The identification of protein biomarkers might be improved through more sophisticated meta-analysis methods, such as the Rank Product method [30], measurements of concordance among data sets [45], and the identification of common features for diagnostic protein biomarkers. Filtering differentially expressed genes through known proteins detectable in biofluids may also improve specificity.
Future work will involve taking markers validated in our pilot studies of cross-organ AR and testing their clinical utility in blinded prospective studies. These studies might also elucidate the prognostic value of these markers. Given that hundreds of thousands of microarray measurements are now publicly available and that this number is growing, RNA data-driven proteomics could provide hundreds of serum and urine biomarkers for other diseases.
This study was approved by the Stanford University Institutional Review Board. Written informed consent was obtained from all the subjects.
As previously described [12], we identified microarray experiments containing both disease and normal control tissues for 280 diseases from the NCBI Gene Expression Omnibus (GEO) [4], calculated differentially expressed probes with percentage of false prediction (pfp) ≤5% using the RankProd R package [30], and converted probe IDs to Entrez Gene IDs using AILUN [46]. For genes with multiple probes, the probes with the most significant pfp values were used. For diseases with multiple data sets, we used genes that were differentially expressed in at least one data set.
We have previously constructed a human biofluid proteome database [13] with known serum- and urine- detectable proteins containing data from the HUPO Plasma Proteome Project [31], a non-redundant list from the Plasma Proteome Institute [32], the MAPU Proteome database [47], and the Urinary Exosome database [48]. We filtered the differentially expressed gene sets with our human biofluid proteome database to yield potential protein biomarkers for each disease.
We downloaded all diagnostic protein biomarkers from the GVK BIO Online Biomarker Database (GOBIOM) [35] with a selection of Biochemical in Nature and Diagnosis in Application, and limited the retrieval to those with valid Entrez Gene IDs annotated as Protein in Chemical Nature. We mapped clinical indications in this database to disease concepts represented in the Unified Medical Language System (UMLS) [49], and matched them to the disease concepts curated in our microarray data. We used the microarray data to identify 41 diseases with predicted biomarkers and are known protein biomarkers in GOBIOM. For each disease, we calculated the enrichment p-values between the predicted and known protein biomarkers using Fisher's exact test in R.
We collected 18 acute rejection (AR) and 18 stable (STA) biopsy samples from pediatric renal allograft recipients at the Stanford Hospitals, and measured gene expression profiles by microarrays. AR and STA samples were matched for recipient and donor gender, age, donor source, race, time post-transplant, HLA matches. Furthermore, all patients were under the same double (Tacrolimus and MMF) or triple immunosuppression protocols (Tacrolimus, MMF and steroid), and all had received Daclizumab induction therapy [50]. Mean and standard deviation data for patient demographic and clinical variables are provided in Table S1. The difference in the sample collection time between AR and STA was caused by two AR samples collected at 69 and 97 months after transplant. The remaining 16 AR samples were collected at 9±7 months after transplant, the same as that of stable patients. Removing the two late-stage AR samples only caused minor changes in the AR signature. A sample was categorized as AR with biopsy proven according to the Banff classification [51] on tubulitis, interstitial inflammation, glomerulitis, and vasculitis (n = 18, Banff grade samples were IA, IB, and IIA not including border line). Samples were categorized as STA (n = 18) if AR and any other substantive pathologies were absent. We also required stable graft function on protocol biopsy, which we conducted at 3, 6, 12, and 24 months after transplantation and for graft dysfunction [50], [52]. None of patients was infected with BK virus. All pathology analyses were performed by a single blinded pathologist (NK) at Stanford University.
For ELISA experiments on renal transplant serum samples, we used previously collected serum samples from 19 AR and 20 STA patients who were not infected with BK virus. All serum samples were obtained within 24 hours of a clinically indicated or protocol graft biopsy, and each sample was matched with the patient's biopsy. AR samples were biopsy-proven according to the Banff classification (IA, IB, IIA, IIB, not including border line). For specificity testing, an additional 10 samples were collected from patients with chronic allograft injury, who were defined as having an IFTA score ≥1 [51]. Ten samples were collected from renal transplant patients with BK virus infection. We also collected plasma samples from 32 AR patients and 31 STA patients without CMV infection after cardiac transplant at Stanford Hospitals. To minimize loss in sample processing, plasma was directly used in the ELISA study. Acquisition of samples in both studies was approved by the Stanford University Institutional Review Board. All AR samples were graded as ISHLT grade 3A or 3B. Stable samples showed an absence of AR and any other substantive pathology.
Total RNA used for first-strand cDNA synthesis using a T7 promoter-linked oligo(dT) primer following the standard protocol for the Affymetrix One-Cycle cDNA Synthesis Kit (Affymetrix, Part. 900493). After second strand cDNA synthesis, biotin-labeled cRNA was prepared in an in vitro transcription reaction using a GeneChip IVT Labeling Kit (Affymetrix). Ten micrograms of fragmented cRNA was used for hybridization on Affymetrix Human Genome U133 Plus 2.0 microarrays according to the manufacturer's instructions. The raw and processed data have been deposited into GEO (accession ID; GSE14328).
All three datasets (pediatric renal, adult renal, and adult heart) were normalized by the quantile-quantile method using dChip software[53]. Probes significantly upregulated in AR versus STA were identified using Significant Analysis of Microarray (SAM; q ≤0.05) [37]. All probes associated with AR were linked to Entrez Gene IDs using AILUN [46]. We limited AR genes as significantly upregulated in AR compared to STA. We found 9,086 genes associated with pediatric renal AR, 2316 in adult renal AR, and 283 in heart AR.
The number of heart AR genes was significantly less than those of kidney AR genes due to different platforms and organs. Publicly available heart AR data came from studies that used a 70mer spotted array from NIH/NIAID (GEO accession numbers GPL1053 and GSE4470). The array contained 8972 probes that corresponded to 8437 Entrez Gene Ids. This array was smaller than the Affymetrix U133 plus 2.0 array used for the pediatric renal study and the Affymetrix U95 array used for the publicly available adult renal study (GEO accession numbers GPL91 & GDS724).
To make the number of AR genes comparable between pediatric and adult renal studies, we added an extra filter. We included only genes with a fold change ≥2 in the pediatric renal study. We obtained 2,805 pediatric renal AR genes, 2,316 adult renal AR genes, and 283 heart AR genes (Fig. 2). When we intersected these three AR gene lists, we found 45 common upregulated AR genes irrespective of transplanted organs.
Ten proteins in serum were measured by using commercial ELISA kits. ELISA kits for PECAM1 (Cat. No. ab45910), CD44 (Cat. No. ab45912), and SELL (Cat. No. ab45917) were purchased from ABCam Inc (Cambridge, MA); an ELISA kit for SA100A4 (Cat. No. CY-8059) was purchased from MBL International (Woburn, MA); ELISA kits for CCL4 (Cat. No. DMB00), CXCL11 (cat. No. DCX110) and CXCL9 (cat. No. DCX900) were purchased from R&D Systems (Minneapolis, MN). An ELISA kit for STAT-1 (cat. CBA034) was purchased from Calbiochem (Gibbstown, NJ); an ELISA Kit for BIRC5/Survivin (Cat. No. 900-111) was purchased from Assay Designs (Ann Arbor, MI), and an ELISA assay for CCL8 was developed using the DuoSet ELISA Development System for human CCL8/MCP-2 from R&D Systems (Cat. No. DY281).
Sample, reagent, and buffer preparation were done according to manufacturer manuals, and the assay was performed by following manual instructions exactly. Microwell plates were read by a SPECTRAMax 190 microplate reader (Molecular Devices, Sunnyvale, CA). Protein concentrations were determined from a standard curve generated from standards supplied with the kits. Protein concentrations of PECAM1, CXCL9 and CD44 in the plasma samples of heart transplant patients were also measured with the ELISA kits specified above.
Immunohistochemical staining was performed on 4 µm sections obtained from formalin-fixed paraffin embedded tissues using mouse monoclonal anti-human antibodies directed against PECAM-1 (DAKO, Carpinteria, CA; Catalog # M823; dilution 1∶150). Heat induced antigen retrieval was performed with Ventana Benchmark Autostainer. The staining was optimized using appropriate positive and negative controls.
T-tests and chi-square tests were used to compare continuous and categorical clinical variables in patient demographics using SAS 9.1.3 (SAS Institute Inc., Cary, NC). Protein concentration data from ELISAs were compared between AR and STA using the Mann-Whitney U test in R. P-values ≤0.05 were considered statistically significant. The enrichment of known protein biomarkers in differentially expressed genes was calculated using Fisher's exact test in R.
|
10.1371/journal.pntd.0004775 | The Nature of Exposure Drives Transmission of Nipah Viruses from Malaysia and Bangladesh in Ferrets | Person-to-person transmission is a key feature of human Nipah virus outbreaks in Bangladesh. In contrast, in an outbreak of Nipah virus in Malaysia, people acquired infections from pigs. It is not known whether this important epidemiological difference is driven primarily by differences between NiV Bangladesh (NiV-BD) and Malaysia (NiV-MY) at a virus level, or by environmental or host factors. In a time course study, ferrets were oronasally exposed to equivalent doses of NiV-BD or NiV-MY. More rapid onset of productive infection and higher levels of virus replication in respiratory tract tissues were seen for NiV-BD compared to NiV-MY, corroborating our previous report of increased oral shedding of NiV-BD in ferrets and suggesting a contributory mechanism for increased NiV-BD transmission between people compared to NiV-MY. However, we recognize that transmission occurs within a social and environmental framework that may have an important and differentiating role in NiV transmission rates. With this in mind, ferret-to-ferret transmission of NiV-BD and NiV-MY was assessed under differing viral exposure conditions. Transmission was not identified for either virus when naïve ferrets were cohoused with experimentally-infected animals. In contrast, all naïve ferrets developed acute infection following assisted and direct exposure to oronasal fluid from animals that were shedding either NiV-BD or NiV-MY. Our findings for ferrets indicate that, although NiV-BD may be shed at higher levels than NiV-MY, transmission risk may be equivalently low under exposure conditions provided by cohabitation alone. In contrast, active transfer of infected bodily fluids consistently results in transmission, regardless of the virus strain. These observations suggest that the risk of NiV transmission is underpinned by social and environmental factors, and will have practical implications for managing transmission risk during outbreaks of human disease.
| Nipah viruses cause outbreaks of severe human disease with high fatality rates. Different patterns of transmission have been associated with recurrent outbreaks caused by Nipah virus (NiV) in Bangladesh, where person-to-person transmission is a major pathway for human infection, compared to an outbreak in Malaysia, where pig-to-human transmission accounted for virtually all human infections. To date, there have been limited comparative studies that address the question of whether this difference may be attributed to differences between geographical isolates of NiV at the genome level, or to other factors in play during human outbreaks of disease. In this report, we employed the ferret, a surrogate human model, to compare features of infection and transmission of NiV isolates from Bangladesh and Malaysia. Our findings indicate that, although differences in levels of shedding are seen between the virus isolates, transmission risk is more likely determined by interactions between infected patients and at-risk individuals; factors that are driven by the social and environmental context within which human disease events occur. Our observations in the ferret have important implications for the implementation of strategies to mitigate transmission risk during outbreaks of NiV infection in people.
| Nipah viruses (NiV) isolated from Malaysia (NiV-MY) and Bangladesh (NiV-BD) each cause human disease characterized by febrile encephalitis with high case fatality rates, but they exhibit different epidemiological features [1]. Person-to-person transmission is an important pathway for NiV-BD infection of people [2], while for NiV-MY most patients acquired their infection from domestic pigs [3, 4]. It has been suggested that a greater prevalence and severity of respiratory disease signs in patients infected with NiV-BD may facilitate person-to-person transmission by this strain [2, 5], and exposure to respiratory secretions from patients is reported to be a major risk factor for onward NiV-BD transmission [6, 7]. However, NiV-MY has also been isolated from respiratory secretions [8], and so the factors responsible for different strain attack rates remain poorly understood. In particular, viral transmission occurs within a social and environmental framework that includes co-morbidities such as malnutrition and pre-existing respiratory disease [2], or exposure factors such as levels of patient care and interactions between patients and at-risk individuals [1, 9]; each of these may also play an important and differentiating role in NiV transmission rates.
Oronasal exposure of ferrets to human isolates of NiV results in fulminant infection that recapitulates key features of the human disease [10, 11]. Additionally, patterns of NiV shedding in ferrets [12] are consistent with those observed in people, with virus having been recovered from respiratory secretions and the urine of each species. Furthermore, the ferret is a social animal, making conspecific housing studies both feasible and meaningful in the context of virus transmission mechanisms.
In earlier work conducted in the ferret, we found significantly higher levels of viral RNA in oropharyngeal secretions of animals infected with NiV-BD by comparison with NiV-MY [12], suggesting that the two strains may differ in their replication efficiency at sites of potential relevance to transmission. In a new time-course study, we have specifically assessed the relative tropism and replication efficiency of NiV-MY and NiV-BD for the pharynx and the upper and lower respiratory tract tissues of ferrets exposed to virus via a plausible natural route. We observed more rapid onset of productive infection and also higher levels of virus replication in respiratory tract tissues of ferrets infected with NiV-BD compared to NiV-MY. The observation of strain differences in replication efficiency in tissues of relevance to transmission risk suggests a contributory mechanism for the observed increased incidence of NiV-BD transmission between people compared to NiV-MY.
We then assessed the relative propensity of NiV-BD and NiV-MY for ferret-to-ferret transmission under differing environmental conditions of viral exposure. Firstly, opportunity for natural transmission of infection was provided by cohousing naïve ferrets with animals in which NiV infection had already been established by experimental exposure. Cohousing permitted direct contact of naïve animals with environmental saliva, urine and feces of infected cage mates; interaction through play including wrestling and mouthing; mutual grooming including licking; and sharing of sleeping quarters, food and water containers, with most contact occurring during the incubation period or disease prodrome. Secondly, we provided an alternative transmission opportunity by the addition of direct transfer of oronasal fluids between experimentally-infected and naïve ferrets during known shedding periods, including the time of advanced clinical disease.
Infection of naïve ferrets was not identified for either NiV-BD or NiV-MY after they were cohoused with experimentally-infected animals as described above. However, all naïve ferrets developed acute NiV infection following their assisted and direct exposure to oronasal fluid from animals that were shedding either NiV-BD or NiV-MY.
Our findings for this infection model indicate that, although there may be higher levels of viral shedding with NiV-BD compared to NiV-MY, the attack rate for both viruses may be equivalently low under exposure conditions provided by cohabitation alone. In contrast, the attack rates for each virus are similarly high when there is active transfer of infected bodily fluids. These observations have practical implications for the management of transmission risk with NiV-BD and NiV-MY in the field.
The studies were approved by the Commonwealth Scientific and Industrial Research Organisation, Australian Animal Health Laboratory Animal Ethics Committee, in accordance with the National Health and Medical Research Council Australian code for the care and use of animals for scientific purposes 8th edition (2013).
Outbred ferrets, 12–18 months of age, were randomly assigned to virus infection groups (NiV-BD or NiV-MY) which were then maintained under separated biosafety level-4 (BSL-4) conditions. Animal housing, husbandry, and handling for sample collections were as previously described [10, 13]. Before each study, ferrets were implanted with a LifeChip Bio-Thermo (Destron Fearing, Eagan, MN, USA), and baseline subcutaneous and also rectal temperatures, body weight data, and serum samples were obtained.
After exposure to viral inoculum or potentially infectious secretions, animals were assessed daily for signs consistent with acute NiV infection such as reduced play activity, including by remote scanning of microchip temperature. Ferrets were also anesthetized at selected time-points for collection of clinical samples as outlined below. Animals that reached a predetermined humane endpoint for disease were euthanized as previously described [10, 11], irrespective of when their euthanasia had been scheduled within the study design. Clinical samples and tissues were collected from all animals at euthanasia to assess for viral shedding, viremia, and virus replication.
Virus inoculums used were low-passage human isolates that had been cultured in Vero cells (passage 2 or 3). The NiV-BD isolate originated from the oropharynx of one of 12 people infected during an encephalitis outbreak in Rajbari district, Bangladesh (Nipah Bangladesh/Human/2004/Rajbari, R1; [14, 15]). The NiV-MY isolate was Nipah virus/Malaysia/Human/99, originating from the cerebrospinal fluid of an encephalitic patient; a comparison of the NiV-BD and NiV-MY strains at the nucleotide and amino acid levels is provided elsewhere [15].
Ferrets in the time-course study and donor animals in the transmission studies were anesthetized and exposed to 5,000 TCID50 NiV-BD or NiV-MY in 1 mL phosphate-buffered saline (PBS) via the oronasal route as previously described [11, 12]. In each instance, inoculums were back-titrated on Vero cells to confirm the administered dose.
Clinical samples comprised whole EDTA-treated blood, nasal washes, oral and rectal swabs and, when available, urine; samples were analyzed for viral load by RT-PCR and virus isolation. Whole blood was also collected from recipient ferrets at euthanasia for measurement of serum antibody against NiV. Tissues collected at post mortem were analyzed by RT-PCR, virus isolation, histopathology, and immunohistochemistry (IHC) using routine methods [13, 16]. For the time-course study, respiratory tissue sampling was more extensive. This included: upper respiratory tract (URT) comprising extra- and intrathoracic sections of the trachea, hard and soft palate (oral and nasal mucosal surfaces), nasal turbinate, pharynx, larynx, tonsil, and the base of the tongue (for pharyngeal mucosa); and lower respiratory tract (LRT) comprising hilar and peripheral regions of each major lung lobe plus hilar samples from intermediate lung lobes. All respiratory tissues were assessed by histopathology and IHC. Lower respiratory tract tissues, pharynx, nasal turbinates and trachea were also assessed by RT-PCR.
Extraction of viral RNA from nasal washes, oral and rectal swabs, whole blood and urine was carried out as described [13], or using a magnetic bead-based robotic liquid handling platform (Janus Integrator Platform, PerkinElmer, MA, USA) and MagMAX Express-96 Magnetic Particle Processor as per the manufacturer instructions (Applied Biosystems, CA, USA). Evaluation of RNA samples by multiplex RT-PCR, using primers targeting the NiV N gene and host 18S ribosomal RNA (rRNA), was performed as described [12]. Samples were considered positive for NiV N gene by RT-PCR if the mean cycle threshold value of duplicate reactions was <38.14. Virus isolation by titration on Vero cells was attempted on clinical and tissue samples positive for NiV by RT-PCR [10].
Sera from recipient ferrets were tested for the presence of a specific (binding) antibody response against NiV glycoprotein G by multiplex bead-based Luminex array (Bio-Plex, Bio-Rad Laboratories, CA, USA), and for the presence of NiV-neutralizing antibodies by serum neutralization testing (SNT; [17]). Selected sera were also assessed for differential IgM and IgG antibody responses to NiV-BD infection. To this end, binding antibody Luminex analysis was performed using methodology described by Bossart et al. (2007), except that instead of protein A/G, each sample was tested using biotinylated anti-ferret IgM and anti-ferret IgG (1:50,000 and 1:10,000 in PBS, respectively; Rockland Immunochemicals, PA, USA).
Ferrets were considered to have become infected if viral RNA was detected in any tissue sample by RT-PCR at post mortem, or if anti-viral antibody was found in serum. Because analysis of tissues by RT-PCR did not allow differentiation between viral genome and mRNA representing viral transcription, viral replication was considered to have occurred within a tissue if both host and viral elements were seen: viral RNA was recovered and NiV antigen was identified within cells; or, virus was re-isolated and antigen was identified within cells; or, virus was re-isolated and histopathologic lesions typical for NiV infection (for example vasculitis, necrosis, or syncytia) were identified; or, antigen was identified in association with lesions typical for NiV infection. Respiratory secretions were defined as nasal wash and oral swab samples; viral shedding in respiratory secretions was defined as agent detection by both RT-PCR and isolation from either sample. Fever was defined as a rectal or subcutaneous temperature >40°C.
For samples assessed by RT-PCR, data were expressed as NiV N gene copy numbers relative to 18S rRNA copies, which were calculated as previously described [12]. Statistical analyses were carried out using IBM SPSS statistical software (version 21.0, IBM, Foster City, CA, USA), to a significance level P ≤ 0.05 for all tests.
All ferrets were successfully infected with either NiV-BD or NiV-MY, with one animal exposed to NiV-BD (Ferret 11) reaching its humane endpoint on d6pi, prior to its scheduled euthanasia of d7pi. Clinical disease was not observed in ferrets euthanized on days 1 to 3pi. Fever was recorded from d4pi in 6/8 remaining ferrets given NiV-BD and from d5pi in 6/6 remaining ferrets given NiV-MY. Other clinical signs included reduced play activity in NiV-MY ferrets from d5pi and agitation, disorientation, ataxia, facial edema, hunched posture, tachypnea/ dyspnea and straining to defecate in NiV-BD ferrets from d6pi.
Virus replication and shedding data are summarized in Tables 1 and 2.
Virus replication was detected in the LRT of NiV-BD animals from d1pi, but not in NiV-MY ferrets until d3pi. From d3pi, NiV-BD ferrets shed virus in respiratory tract secretions and replicated virus in the URT, and both groups replicated virus in lymph nodes. NiV-MY was identified in respiratory tract secretions (but not URT tissues) on d4pi and 5pi but not thereafter, in spite of ongoing URT replication in the same animals. Overall, there was a trend towards earlier recovery of NiV-BD from respiratory secretions, URT, and LRT compared to NiV-MY (P = 0.051). Among ferrets that were febrile at euthanasia, the proportion of animals shedding virus in respiratory tract secretions was significantly higher for those exposed to NiV-BD, compared to NiV-MY (P = 0.030), despite similar rates of detection of replication of the two virus strains in the URT and LRT tissues. For both strains, detection of viral RNA in blood coincided with replication of virus in spleen from d4 to d5, followed by widespread viral replication in liver, kidney, and/or gonads. In this study, virus replication in brain was only found in ferrets infected with NiV-BD.
Once developed, the features of LRT pathology were comparable between NiV-BD and NiV-MY. Early changes were confined to accumulation of viral antigen in alveolar walls and vascular endothelium without other lesions, progressing to focal or multifocal bronchoalveolitis (Fig 1) followed by necrosis and vasculitis. Viral antigen was detected in bronchoalveolar and glandular epithelium and luminal airway debris, as well as within alveolar walls, endothelial cells, and epithelial and endothelial syncytia.
Similarly for the URT, acute viral rhinitis and/or nasopharyngitis were manifest on d3pi in the NiV-BD group and from d5pi in the NiV-MY group (Fig 2). Epithelial cells positive for viral antigen were also seen in the hard and/or soft palates and trachea of animals infected with NiV-BD (Ferrets 2, 8, 11, 13) and NiV-MY (Ferret 24) late in the disease course. Interestingly, at the later time points there was also viral antigen in, variously, nasopharyngeal epithelium; nasopharyngeal submucosa including connective tissues (Ferret 21) and vascular endothelium; mucosal lymphoid tissue; and tracheal and soft palate stroma of animals infected with NiV-BD (Ferrets 8, 11, 13, 14) and NiV-MY (Ferrets 15, 20, 21, 24, 25).
Viral loads in URT and LRT tissues, calculated from RT-PCR data, are presented in S1 Fig; outcomes of REML analyses of respiratory and lymphoid tissue and respiratory secretions are presented in S1 Table. Ferrets exposed to NiV-BD had significantly higher predicted mean viral RNA levels in airways, lungs and lymphoid tissues over time compared to those exposed to NiV-MY (Fig 3). For both groups, there was also a strong trend of increasing viral RNA levels in these samples as dpi progressed. Ferrets that had fever at euthanasia had significantly higher viral loads in lymphoid tissues for both viruses, a finding that was not observed for other tissues once the strong effect of day was accounted for in the statistical models (S1 Table). This finding was of unclear pathogenic significance but may reflect the extent of macrophage/ lymphocyte activation and production of pyrogenic cytokines with increasing viral loads.
In the lungs, for both viral strains significantly higher levels of viral RNA were recovered from hilar lung tissue, compared to peripheral lung tissue samples (Fig 4).
Viral RNA was detected in nasal wash and oral swab samples from animals in which there was no evidence of viral replication in tissues of the URT, and was considered reflective of virus with a LRT origin; accordingly, nasal wash and oral swab samples were analyzed together. Compared to tissues, viral loads in respiratory secretions did not show the same strong effect of day on increasing viral load, with an increase observed to day 5 followed by a plateau or decrease in levels (Fig 5). However, as for respiratory tract tissues, predicted mean viral loads in respiratory secretions over time were significantly higher for NiV-BD-infected ferrets (Fig 5).
Levels of viral RNA in blood and in other major organs (kidney, adrenal gland, liver, thymus and brain) were comparable for the virus strains (blood, Fig 5; major organs, S2 Fig).
Directly exposed donor animals were observed to interact closely with in-contact animals for the first two to three days of cohabitation. They engaged in play (wrestling and mouthing) and mutual grooming (including licking); slept together with shared bedding; and shared toileting spaces, food and water bowls, food items, and chew toys. With the progression of infection, each donor exhibited increasingly reduced play activity and reluctance to leave the bedding area.
All ferrets directly exposed to NiV developed clinical disease consistent with acute, systemic NiV infection and were euthanized at humane endpoint on d7pi and d8pi (NiV-BD) and d7pi and d8pi (NiV-MY). NiV was re-isolated from respiratory secretions and urine samples collected from each donor at euthanasia and histopathologic lesions consistent with acute, systemic NiV infection were found in each animal [10–12]. Virus replication was detected in URT and LRT tissues.
For NiV-BD, all four in-contact ferrets remained clinically well over the four week observation period. Viral RNA was not detected in clinical samples from any in-contact ferret at any sampling time after commencing cohabitation with a donor ferret. At post mortem examination, all tissues were negative for viral RNA, antigen, and histological lesions, including the olfactory pole and occipital region of the brain; serum was negative for anti-NiV antibody. It was concluded that NiV-BD infection did not occur in any in-contact ferret.
For NiV-MY, all four in-contact ferrets remained clinically well over the four week observation period. Viral RNA was detected in nasal wash samples of three of four ferrets on one of either 2, 4 or 6 days after commencing cohabitation with a donor ferret. The in-contact ferret which had viral RNA in its nasal wash at day 6 following cohabitation also had viral genome in a rectal swab sample at day 2. NiV was not re-isolated from RNA-positive specimens. All other clinical samples from in-contact ferrets were negative for viral RNA. At post mortem examination, all tissues, including the brain, were negative for viral RNA, antigen, and histological lesions, and serum was negative for anti-NiV antibody by Luminex assay and by SNT. Opportunity for contact with donor ferrets continued for at least 4 to 5 days (depending on the time of donor euthanasia) after cohabitation commenced. Accordingly, viral RNA in day 2 and day 4 nasal wash samples were attributed to sub-infectious viral doses or non-viable viral genome derived from donor secretions. It is possible that the in-contact ferret with viral RNA in a rectal swab at day 2 and nasal wash at day 6 (one day after donor euthanasia) experienced transient URT infection that was rapidly cleared by innate immune mechanisms. However, in the context of the current work, it was concluded that NiV-MY infection did not occur in any in-contact ferret.
Directly exposed donor animals were observed to interact closely with in-contact animals for the first day of cohabitation. With the progression of its infection, each donor exhibited increasingly reduced play activity and reluctance to leave the bedding area over the following one to three days.
All donor ferrets directly exposed to NiV developed clinical disease consistent with acute, systemic NiV infection and were euthanized at humane endpoint on either d7pi (NiV-BD) or d8pi and d9pi (NiV-MY). NiV was re-isolated from the respiratory secretions and urine samples collected from each donor at euthanasia and histopathologic lesions consistent with acute, systemic NiV infection were found in each animal [10–12]. Virus replication was detected in URT and LRT tissues.
One donor ferret in the NiV-BD group was febrile when recipient animals were introduced at d6pi; the other three donors (one NiV-BD and both NiV-MY donors) were not. The two virus exposure doses for each recipient pair are presented in Fig 6; titres were comparable for the virus strains and, as expected, viral titre in secretions collected at the time of donor euthanasia were equivalent to or higher than those from samples earlier in the donor infection course.
All in-contact ferrets developed clinical disease consistent with acute, systemic NiV infection reported after routine experimental exposure; they were euthanized at humane endpoint on days 6 to 9 after first exposure to donor-derived inoculum and the commencement of cohabitation. All in-contact ferrets shed virus in nasal washes and oral swabs and, variously, in rectal swab samples (Fig 6): the shedding pattern was comparable to those observed in ferrets receiving routine experimental exposure to NiV [12]. In one NiV-MY in-contact ferret (Ferret 17), respiratory shedding was detected as early as day 2 after the first exposure and prior to the second exposure to donor inoculum: infection of this animal was therefore known to have been initiated by the first donor inoculum which had a virus titer of 63 TCID50; the donor ferret was asymptomatic at that time.
Levels of viral RNA in respiratory secretions increased across time for both viruses (Fig 7) and on this occasion were found to be comparable between NiV-BD and NiV-MY; similar viral loads were also found in lung tissues of NiV-BD and NiV-MY infected ferrets (S1 Table).
At ferret euthanasia, all donor and recipient serum samples were negative for neutralizing antibody by SNT. Sera from NiV-BD in-contact ferrets were also analyzed by Luminex using biotinylated IgG and IgM anti-ferret antibodies (Fig 8). Three of the four animals had developed an anti-NiV IgM response between day 4 and day 8 after the first exposure to donor inoculum, accompanied by a less marked rise in IgG.
There were no significant differences between NiV-BD and NiV-MY in the overall proportion of transmission events but, for each virus, the proportion of transmission events was significantly higher when direct transfer of oronasal fluids took place in addition to cohousing from late in the incubation period (P = 0.029).
In spite of clinical similarities in disease–but important differences in the incidence of person-to-person spread–between NiV-BD and NiV-MY, there are few reports of strain comparison studies in laboratory animals that have employed plausible natural routes of virus exposure, and reported transmission studies have been limited to NiV-MY [16, 18]. In one comparative study in which hamsters received intranasal exposure to virus, more severe lesions were identified with NiV-BD at d2pi in both URT and LRT but, by d4pi, similar degrees of rhinitis and broncho-interstitial pneumonia were present with both virus strains [19]; comparative analysis of viral loads was not described. In mice, those receiving intranasal NiV-BD were more likely to have viral antigen detected in the lungs compared those given NiV-MY [20]. However, NiV infection of mice is also subclinical and so its application to NiV strain comparisons in the context of human disease is limited, because a typically infected mouse does not exhibit equivalent illness behaviors that may influence its infectivity for other mice.
We employed an oronasal route of exposure for direct ferret challenge in this NiV strain comparison study, and it is likely that a modicum of virus inoculum was delivered directly to the ferret LRT. However, the URT was also exposed to virus at this time, providing equivalent opportunity for early replication at that site. For both NiV-BD and NiV-MY, the trend for earlier detection of infection in LRT tissues, as well as the recovery of significantly higher levels of viral RNA in hilar lung tissue suggest that under these conditions of exposure, infection is first established at the level of the bronchi or larger bronchioles. Thus, while viral replication in the URT may contribute to increasingly infectious respiratory secretions as the disease progresses, it is possible that LRT exudates may play an important role in infectivity particularly during the early stages of productive infection. Viral antigen has been noted in the bronchiolar epithelium of a human NiV case, and in vitro studies have demonstrated that undifferentiated primary human epithelial cells derived from bronchi and small airways were permissive to NiV infection and supported replication to high titers [21, 22]. Overall, these observations align with the demonstrated epidemiological link in NiV-BD-infected patients between coughing and an increased risk of human-to-human transmission, with coughing generating particles small enough to be inhaled into the distal airways of in-contact individuals [23].
We have shown that oronasal exposure of ferrets to comparable doses of NiV from either Bangladesh or Malaysia induced similar systemic diseases, albeit in small cohort sizes. However, we also observed a trend for earlier viral replication in the URT and LRT tissues and earlier detection of viral shedding in respiratory secretions of NiV-BD compared to NiV-MY, as well as higher levels of NiV-BD RNA in respiratory tissues and secretions over the infection course. Differences have been described for geographically-distinct NiV isolates at the genome and putative protein levels, including for the NiV-encoded nonstructural proteins V, C and W [15] which, together with P, are key viral determinants of disease outcomes in ferrets [24] and hamsters [25] infected with NiV-MY, through antagonism of host interferon-mediated innate immune responses to infection [26]. The editing site on the P gene, which prompts an open reading frame shift facilitating translation of the accessory proteins, also differs between NiV-MY and NiV-BD [15]. However, the greatest heterogeneity between NiV isolates from Bangladesh and Malaysia lies in the untranslated regions (UTRs) of the P gene encoding these accessory proteins [15, 27], and it is possible that heterogeneity within this region may give rise to differential regulation of viral gene transcriptional and translational efficiency. This has been described for mutations in gene UTRs for other paramyxoviruses [28–30] and may result in differential replicative ability of NiV at sites relevant to transmission. There is also in vitro evidence that viral replicative kinetics and induction of innate immune factors may differ between NiV-BD and NiV-MY, in that in hamster kidney cells NiV-MY caused more rapid and severe cytopathology and replicated to higher titers over time compared to NiV-BD [31]. However, our observations in ferrets of similar viral loads between NiV-BD- and NiV-MY in blood and other major organs, in the face of contemporaneous differences in respiratory and lymphoid tissues, suggest that functionally significant differential replicative ability of NiV strains during natural infection may be peculiar to the intact respiratory and lymphoid systems. In any event, our observations of a trend for earlier onset of respiratory tract shedding, a higher rate of respiratory tract shedding during clinical disease, and a greater magnitude of viral shedding from NiV-BD infected ferrets constitute a virus-dependent contribution to the basic reproduction number (R0) which might result in a difference in R0 between the two virus strains.
However, viral factors are only one determinant of R0; social factors–for example, those which define the interactions between affected individuals and the other members of the population–also play an important role in the likelihood of transmission. Accordingly, we adapted studies designed to assess the transmissibility of influenza virus strains in ferrets [32–34] to the operational constraints of the BSL-4 environment and completed two small scale comparisons of NiV strain transmissibility, providing different contact exposure opportunities which were duplicated for each virus strain.
In the first cohabitation study, we simulated such features of NiV-BD human infection clusters as: being in the same room or having close physical contact with an infected patient; sleeping in the same bed as a symptomatically-infected patient; and sharing food, cutlery or crockery with an infected patient [2, 7]. But we also observed that towards the time of peak virus replication and with the onset of clinical illness, infected ferrets reduced their play activity, food and water intake, and withdrew to their sleeping quarters from which they had to be encouraged to rise. Under these conditions of contact exposure opportunity, not only was transmission to cage contacts not observed for either NiV-BD or NiV-MY, but the data also indicate that any impact on transmission likelihood which might have been conferred by enhanced shedding of NiV-BD was not able to be shown in a study of this size.
Importantly, the interactions between sick and healthy ferrets did not adequately mirror human behaviors, as ferrets that became socially withdrawn as they became more infectious were increasingly ignored by their cage-mates. Specific activities identified as risk factors for human NiV infection in Bangladesh included direct contact with respiratory secretions from a patient, receiving a cough or sneeze from a patient directly to the face, and interaction with moribund patients through activities such as force-feeding or cradling the head of a patient [2, 9]. People who died as a result of their infection were more likely to transmit NiV, and contact with patients with advanced clinical disease was associated with a high transmission risk [2, 7]. In contrast, patients with NiV-induced disease in Malaysia received their care within advanced tertiary healthcare settings and, following detection of virus in secretions from patients early in the outbreak, hospital hygiene and infection control measures were scaled up to include rigorous barrier nursing practices to protect in-contact health care workers and patients’ visitors [35]. Considered together, these factors suggest that extremely close interactions with infected patients, in particular the direct and gross exposure of in-contact subjects to infectious material from the terminally ill, may be critical drivers of transmission.
Accordingly, in the second cohabitation study we not only simulated contact opportunity during the disease prodrome and terminal illness of infected ferrets–providing overlap with the first study in which transmission of infection was not recorded–but we also ensured the repeated exposure of in-contact ferrets to late-stage infectious secretions. This was done in order to mimic close and repeated patient contact which occurs in Bangladesh during advanced disease, intensifying as illness progresses when care-givers and family members provide increased hands-on care [9]. Under these modified conditions of contact exposure opportunity, and in contrast to the outcome of the first cohabitation study, virus transmission to all in-contact ferrets was observed for both NiV-BD and NiV-MY and with a uniformly lethal outcome. In one case, there was evidence of infection in a ferret following its first exposure to secretions from a donor that was asymptomatic, suggesting that onward transmission prior to the onset of clinical disease in patients is plausible under certain exposure conditions.
In ferrets, we found that NiV-BD replicated to higher levels than NiV-MY at anatomical sites relevant to transmission, but under the exposure conditions provided by our first cohabitation study neither NiV-BD nor NiV-MY successfully transmitted from infected to in-contact animals. More highly powered studies would be required to demonstrate smaller impacts of viral factors on R0 that may be of significance at the population level. However, infection is a comparatively low frequency event for both NiV-BD and NiV-MY, and the relative risk of transmission of Nipah viruses from Bangladesh and Malaysia may never be able to be experimentally defined. We also recognize that reported studies of NiV-BD to date have used a human isolate that was not associated with onward transmission in the field, and the possibility that future comparisons of additional field NiV isolates from Bangladesh may reveal even more significant differences in respiratory tract replicative efficiencies cannot be discarded. Interestingly, human-to-human transmission was reported in a recent outbreak of fatal encephalitis in the Philippines that was attributed to infection with a putative henipavirus more closely related to NiV-MY than to NiV-BD [36]. However, no virus isolate was obtained from this outbreak, and agent characterization has thus far been limited to partial genome sequencing.
Findings from the studies presented herein are consistent with the view that the risk of human-to-human transmission of certain isolates of NiV is more plausibly underpinned by environmental factors in play with each outbreak event, than by inherent differences in tissue tropism or pathogenicity between NiV strains. In particular, we have demonstrated that under experimental conditions that mimic high-risk human exposure events as reported in Bangladesh, NiV-BD and NiV-MY cause similar outcomes of infection in ferrets. Our observations suggest that a critical driver of onward NiV transmission is the nature of the interaction between subjects and that its impact may be irrespective of virus strain. In particular, the social and cultural contexts in which infection events occur, and how these influence the management of the typical infected individual, are the pivotal determinants of the likelihood of onward transmission.
|
10.1371/journal.pcbi.1002755 | Determinants of Translation Elongation Speed and Ribosomal Profiling Biases in Mouse Embryonic Stem Cells | Ribosomal profiling is a promising approach with increasing popularity for studying translation. This approach enables monitoring the ribosomal density along genes at a resolution of single nucleotides.
In this study, we focused on ribosomal density profiles of mouse embryonic stem cells. Our analysis suggests, for the first time, that even in mammals such as M. musculus the elongation speed is significantly and directly affected by determinants of the coding sequence such as: 1) the adaptation of codons to the tRNA pool; 2) the local mRNA folding of the coding sequence; 3) the local charge of amino acids encoded in the codon sequence. In addition, our analyses suggest that in general, the translation velocity of ribosomes is slower at the beginning of the coding sequence and tends to increase downstream.
Finally, a comparison of these data to the expected biophysical behavior of translation suggests that it suffers from some unknown biases. Specifically, the ribosomal flux measured on the experimental data increases along the coding sequence; however, according to any biophysical model of ribosomal movement lacking internal initiation sites, the flux is expected to remain constant or decrease. Thus, developing experimental and/or statistical methods for understanding, detecting and dealing with such biases is of high importance.
| Gene translation is the process by which ribosomes translate mRNA molecules to proteins, a central process in all living organisms. Thus, understanding the biophysics of gene translation and the way its efficiency is encoded in the different features of the coding sequence has ramifications to every biomedical discipline. Recently, a new large-scale experimental approach named ‘ribosomal profiling’, has been developed for monitoring the ribosomal density at a resolution of single nucleotides. In this study, we analyzed ribosomal profiling data of mouse embryonic stem cells. These data enabled us to directly show that translation velocity is affected by the adaptation of codons to the tRNA pool, local mRNA folding of coding sequence, and local charge of the amino acids encoded in the coding sequence. In addition, our analyses suggest that ribosomal speed tends to be slower at the beginning of the coding sequence. Finally, we report possible biases in the ‘ribosomal profiling’ procedure that should be considered in future studies utilizing this method.
| Gene translation is the second major step of gene expression and thus has ramifications related to every biomedical discipline including human health [1], [2], biotechnology [3], evolution [4]–[6], functional genomics [7], [8] and systems biology [9], [10]. One of the open questions in the field is related to the way translation efficiency is encoded in the transcript.
The most promising approach for studying gene translation is the ribosomal profiling method [11]. This approach was introduced only a few years ago but has already been successfully employed for answering various fundamental biological questions [12]–[19]. Specifically, ribosomal profiling has been used for: 1) understanding the mechanism of gene expression down-regulation by microRNAs [13], 2) understanding the dynamics of translation in mouse embryonic stem cells [12], 3) showing that the anti-Shine–Dalgarno sequence drives translational pausing and codon choice in bacteria [14], 4) studying the yeast meiotic program [15], 5) showing that miR-430 reduces translation before causing mRNA decay in zebrafish [17], and 6) to reveal the co-translational chaperone action of trigger factor in vivo [16].
In the current study we analyzed ribosomal profiles of mouse embryonic stem cells measured in a previous experiment [12]. The experiment output included ribosomal density measurements along hundreds of genes at a few time points, after preventing translation initiation. These data enabled us to infer the translation elongation speed in different genes, allowing us for the first time to study several biophysical aspects of translation elongation in mouse embryonic stem cells.
To study the kinetics of translation elongation in M. musculus, ribosome footprint profiles of isoforms expressed in embryonic stem cell were reconstructed based on a previous study [12]. Briefly, translation was halted by applying cyclohexamide. Fragments covered by ribosomes were mapped to the transcript and a baseline ribosomal read counts profile for each expressed isoform was created (see Methods). Let us denote these created profiles by , where is the index of the analyzed isoform. In addition, to estimate the elongation speed of ribosomes, in three additional experiments harringtonine was used to stop translation initiation, while allowing ribosomes that already started translating the mRNA to continue their movement on it. Cyclohexamide was again applied 90/120/150 seconds after applying harringtonine to stop translation. In this work, the time difference between applying harringtonine and cyclohexamide for creating depleted profiles is named the ‘run-off’ time.
Let us denote the ribosomal read counts obtained in each of these three experiments by accordantly. The estimated Starting Location of the depleted ribosomal profile (SL) was defined as the point where the ribosomal read counts profile of gene at time point ( profile) reached half of the original ribosomal read counts profile (Methods). Using these SL points, local translation elongation velocities were estimated for each analyzed isoform. Figure 1 outlines a schematic description of the method used to estimate the SL points, demonstrated on the uc007gge.1 isoform (see also Figure S7).
In the original work 4,994 isoforms with good read counts were found [12]. The authors noticed that the effect of harringtonine was best observed for genes longer than 750 codons, as for shorter genes the ribosomes managed to exit the mRNA for the used run-off times. Thus, only genes that were long enough (at least 750 codons) were used to infer the position of the SL points. In the current work, the same isoforms satisfying these conditions were analyzed, resulting in 785 processed isoforms (see Table S10). Let us define the three estimated SL points by corresponding to time points 90/120/150 respectively. Let us mark with the segments defined by , accordantly, and the ribosomal average translation velocity in these segments by and . The average translation velocity of a segment was estimated by dividing the segments' length by 30 seconds. For each gene and time point, various quality checks were performed to reliably estimate the position of the SL points (see more technical details in the Methods section). Eventually, only isoforms with SL points satisfying were selected, resulting in 692 valid isoforms out of the 785 processed isoforms (88%).
Analysis of the data indicated that the median length of was 128 codons (130±77 codons) while the median length of was 184 codons (181±75 codons). Therefore, although the mean translation velocity of all genes is around 5.5 codons/second [12] (see Figure 2A and Tables S4, S6), the average translation velocity along the second segment is larger than the average translation velocity along the first segment (6+/−2.5 codons/second vs. 4.3+/−2.6 codons/second, Wilcoxon test p = 2.2*10−26 Figure 2A). This result remains significant under various estimations methods of these velocities.
We performed additional analyses to support the conjecture that translation elongation velocity is not similar among genes: first, the standard deviation of the estimated SL points was between 17% and 49% (Figure 2A–B, Table S4, S6, columns 1, 2, 3). Second, the relative difference between the two estimated velocities (calculated using ) resulted in a median value of 0.82 while the median value of the ratio resulted in a value of 1.37 (see also Figure 2C–D). To compare the attained results to simulated genes with uniform translation elongation rate, we simulated 692 synthetic genes with 1) lengths distribution identical to the lengths distribution of the analyzed genes, and 2) with constant codons translation efficiency (see Methods). The ribosomal profile of these genes was simulated with a biophysical model (see Methods), resulting in a much smaller difference between the calculated velocities , (median = 0.01; KS-test: p-value <1.81*10−271), as seen in Figure 2C. The ratio between the velocities was also much more moderate when calculated on these simulated ribosomal profiles (0.99+/−0.03, KS-test p-value <1.56*10−295), as seen in Figure 2D. This comparison supports the claim that there is a high variance in the elongation speed of the analyzed genes.
In order to explain the high variability among segments length, those were analyzed with respect to different features of the coding sequence, such as the adaptation to the tRNA pool (e.g. the tAI [20] and the CAI measure [21]), local mRNA folding energy [22] and local charge of the translated amino acids [22], [23]. Specifically, codons recognized by more abundant tRNA molecules increase the tAI measure, therefore we expect longer segments to positively correlate with this measure [24]. The CAI index, which measures the frequency of codons in a segment relatively to their appearance in highly expressed genes, is also expected to positively correlate with the segment length.
In addition, it was suggested that strong local mRNA folding tends to slow down ribosomal translation elongation as it increases the time it takes the helicases to unfold the mRNA molecules [24]. Therefore, segments more strongly folded (i.e. with lower folding energy (FE)) are expected to be shorter. Finally, the polypeptide must traverse two negatively charged regions to exit the ribosome [22], [23], [25], thus charged amino acids (specifically positively charged amino acids [23]) that are encoded in the codons preceding (upstream) the currently translated codon should have electrostatic interactions with the ribosome exit tunnel [22], [23], [25]. Therefore, segments more positively charged are expected to be shorter. More details about the calculation of these measures appear in the Methods section.
To estimate the distinct contribution of each of the coding sequence features to the elongation speed, we calculated the correlation between the length of the segments and each of these features, when controlling for the other two features, and after binning the data (details in the Methods section). Spearman correlation between the segments length and the genes' tAI/CAI when controlling for charge and folding energy of the segments resulted in a correlation coefficient of r = 0.29/0.21 (P<0.00615/0.049) accordantly. Spearman correlation between the segments' length and their mRNA folding energy when controlling for charge and gene tAI was r = 0.42 (P<4.72*10−5). The correlation between the segments' length and their charge when controlling for folding energy and the genes' tAI was r = −0.21 (P<0.046) (additional analyses appear in the supplementary). Thus, the results reported in the current subsection support the conjecture that the translation elongation speed is independently affected by each of the following features of the ORF: the adaptation of the ORF codons to the tRNA pools, local mRNA folding and local amino acids charge.
As mentioned in the previous section, the speed of translation elongation tends to increase along the coding sequence. Aiming at explaining this phenomenon, features measured on the first and second segment were also compared using a paired Wilcolxon test, resulting in significant values for folding energy (Wilcolxon test: P<1.04*10−3) but not for tAI/CAI and charge. This suggests that in mouse, a possible explanation of the increase in translation speed along the coding sequence is the decrease in the strength of the mRNA folding along the coding sequence. Finally, a weak but significant correlation between the average and translation speed and the average transcripts length was observed in mouse (Spearman correlation: r = −0.05, p = 0.022), supporting the conjecture that shorter genes are more efficiently translated.
According to the accepted biophysical model of translation, during the elongation step ribosomes move along the coding sequence, translating each codon with a speed related to the features of the coding sequence in its vicinity and according to cellular factors such as concentrations of elongation factors and tRNA molecules. In addition, a ribosome may be delayed if a ribosome is located downstream in front of it [26]. It is also assumed that in general, ribosomal abortion during translation is relatively rare and that initiation usually occurs at the 5′UTR (i.e. ribosomes do not appear in the middle of the coding sequence [26]).
According to the protocol of the experiment (e.g. see [11], [12]), ribosomal footprint reads of a certain codon are generated when the codon is covered by ribosomes. From a biophysical perspective, slower codons are covered by ribosomes for a larger amount of time (relatively to other codons in the mRNA), creating a higher number of reads (for an illustration see Figure 3A).
In this study, for each analyzed isoform, both and segments were assumed to be translated in an equal time interval of 30 seconds, therefore according to the above assumption, on average, it is expected for the sum of read counts in the and segments (measured on the baseline profile ) would be equal. Therefore, in each isoform the shorter segment is expected to have a higher ribosomal read count per nucleotide in comparison to the longer one.
Let us mark the sum of read counts in intervals and by , accordantly. Let us define the percentage difference between and (relatively to the minimum of and ) byThis measure is invariant to the genes' various mRNA levels and translation initiation rates, therefore enabling comparison between all analyzed isoforms. Using the above assumption, we expect this measure to be close to zero. Figure 3B shows the histogram of the measure calculated both on the real ribosomal profiles and on the simulated ribosomal profiles created using the TASEP biophysical model for various initiation rate values (see Methods). However, in contrast to the made biophysical assumptions, the results indicate that for a substantial part of genes, the measure is abnormally high (median value of 88 vs. 1–6 for simulative data of different levels of noise; KS test, all p-values <3.97*10−215).
In addition, the ribosomal flux at a certain codon along the coding sequence is defined as the multiplication of the translation velocity and density at this point . Therefore, according to any biophysical model with negligible amount of initiation events inside the ORF, we expect the flux to be constant (i.e. for different ) or decrease (due to ribosomal abortion);
Let us mark the mean ribosomal read counts measured in the first and second segments by and respectively and the average velocity in the first and second segment by and . If we assume that the local flux remains constant, we also expect that . Given that the average velocities of , in both the first and second intervals were measured during the same time intervals, we can rewrite this relation as
Thus ifwe would expect the correlation between and to be negative. Intuitively, for a given gene, longer segments should have relatively lower mean read counts. Indeed, the calculated ratios for the simulated densities resulted in a negative correlation (Figure 4B, Spearman correlation of R = −0.9, P<10−291; R = −0.91, P<10−294; R = −0.91, P<10−297; for low/high/proportionate initiation rates). However, when measured on real ribosomal read counts profiles, the correlation between and achieved a significant positive value (R = 0.13, P<0.00082; Figure 4A), contradicting the accepted translation model. Finally, the flux itself is expected to remain constant or decrease (due to ribosomal abortion), i.e. . Yet, we found that this ratio tends to increase (.
Next, we calculated the values of all presented measures on the simulated ribosomal profiles for different initiation rate regimes (see Methods) and compared them to the values obtained when calculating them on the real ribosomal profiles. This analysis resulted in significantly different values: the measure calculated on the simulative data resulted in a median value of 1.01 (KS test in comparison to the measured data: p-value <9*10−95; Figure 4C, Table S9), while the difference between the velocities and resulted in a median value of 0.06 (KS test: p-value <6.18*10−153; Figure 2B). In addition, the ratio between the velocities resulted in median values of 1–1.01 (KS test: p-value <5.67*10−250) (Figure 2C). Overall, the comparisons between all measures calculated on the experimental data and on the simulative ribosomal profiles created by the biophysical model point on the existence of substantial biases in the data produced by the ribosomal profiling procedure.
In this study, we reanalyzed the ribosomal profiling data of mouse embryonic stem cells that was generated in a previous study [12]. Our analysis demonstrates that even for relatively long analyzed genes, that are not expected to be under strong selection for translation efficiency [27], in unusual tissue/conditions such as embryonic stem cells, translation elongation speed is affected by features such as the adaptation of codons to the tRNA pool, local mRNA folding, and charge.
In addition, our analysis directly shows for the first time that the translation elongation speed tends to increase along the coding sequence. The reasons for this phenomenon may be related to the fact that at the beginning of the coding sequence features such as adaptation to the tRNA pool and mRNA folding strength tend to slow down ribosomal movement (see, for example, [22], [24]). This may also be related to the fact that there is a selection for lower codon bias at the beginning to reduce the costs of both missense and nonsense translational errors [28], [29]. The statistical analysis performed in this study support the conjecture that the slower speed at the beginning of the coding sequence is due to stronger mRNA folding in this region. This phenomenon, however, may also be related to yet unknown properties of this process or to biases of the ribosomal profiling methods.
Finally and importantly, at least in the reported study, our analysis demonstrates the existence of some unexplained deviations between the output of the ribosomal profiling approach and any of the accepted models of translation elongation, which assume that the rate of initiation from sites inside the ORF is negligible. This discrepancy may be explained by the fact that current models of translation elongation are inaccurate and, for example, initiation does tend to occur from sites inside coding sequences. However, the most plausible explanation is that ribosomal profiling approach, as in the case of the more traditional approaches for studying mRNA levels (e.g. [30]), includes experimental biases that should be further explored. Another bias of the ribosomal profiling approach which is related to the increased ribosomal density at the beginning of the ORF has been suggested recently in [12].
We also suggest a few explanations for these observed biases, while taking into consideration that there might be additional sources of bias in the ribosomal profiling protocol that are not mentioned here. For example, an insufficient number of mRNA molecules could increase the estimation errors and bias all the presented measures. Specifically, the ribosomal profiling approach produces for each gene the ribosomal positions along the mRNA molecules that have been transcripted from it and that are present in the cell at the time of the experiment. As the read counts per location of a single mRNA are stochastic, averaging them over many mRNA molecules of a gene should theoretically produce a profile that is similar to the stationary density profile of the gene. Thus, the number of mRNA copies affects the averaged profile and eventually the quality of the estimated measures mentioned in this study. In practice, genes with a relatively low number of mRNA molecules can result in highly biased profiles. Indeed, when we modified our computational simulation of the experiment to simulate a low number of mRNA molecules per gene (see Methods), the correlation between and decreased (Figures S12, S13, S14) while the DSRC measure increased (Figures S9, S10, S11), contrary to the expected trend.
Another source of bias may be related to the fact that the current ribosomal density protocol involves filtering some of the reads, distorting the resultant ribosomal density profiles. Specifically, by the protocol of the experiment, only short mRNA fragments that are covered by exactly one ribosome (i.e. monosomes) are purified for further analyses [11], [31], while mRNA segments covered by polysomes are discarded. Thus, it is also possible that the reported biases are, at least partially, due to the fact that fragments that origin from ribosomes located very close to each other on the mRNA are filtered and not analyzed, creating deviated ribosomal profiles. Indeed, cases of fragmented mRNA covered with more than one ribosome as a result of very close ribosomes were reported in a previous study [32]. In addition, when only monosomal footprints were considered in the simulation (see Methods), we obtained a decrease in the correlation between the and ratios and a major increase in the DSRC measure (see Figures S9, S10, S11, S12, S13, S14).
The deviations from the accepted biophysical model could also be explained by the non-uniform effect of the harringtonine/cyclohexamide substances on the different mRNA molecules, causing uneven run-off times, and distorting the location of the SL points. The simulation of this possible experimental bias (see details in Methods) also resulted in an increased DSCR and a decrease in the correlation between and (Figures S15, S16, S17).
Finally, complex relations between the sequence features, their effect on ribosomal density and on the output of the ribosomal profiling approach may also contribute to the deviation from the biophysical model. For example, it was suggested that elongation speed and ribosomal density are affected by the strength of the local folding of the mRNA (stronger folding→slower elongation speed→high ribosomal density) [22]. However, it is also possible that stronger mRNA folding decreases the efficiency of footprint production in the ribosomal profiling protocol (e.g. the efficiency of RNase activity decreases for mRNA fragments with strong folding; e.g. see [33]), contributing to a distorted ribosomal density profiles.
Nonetheless, currently, the ribosomal profiling approach is the major method for studying gene translation, therefore understanding these biases and accurately correcting them should significantly affect studies in various biomedical disciplines. As was demonstrated in this study, one possible direction for detecting such biases is by comparing the ribosomal profiling outcome to the computational biophysical models using statistical analysis. We believe that such approach will be used in the future for employing filters and normalization procedures that are inversed to the noise/bias obtained in the experimental procedure and for adjusting the experimental procedure itself.
Sequencing data were downloaded from the GEO database (accession number GSE30839) [12]. We analyzed all data related to the study of the kinetics of translation elongation. The specific processed files are summarized in Table S1.
Sequenced reads comprise short RNA fragments of different lengths; therefore, a generated linker sequence (CTGTAGGCACCATCAATTCGTATGCCGTCTTCTGCTTGAA) was attached to enable the recovery of the original fragment. More details of this method appear in the original work [12]. In this study, linkers were first detected and removed from the published fragments and only then aligned to transcripts. The start location of the linker was estimated to be between the 20–36 nt of the RNA fragment. Next, the distance between the estimated linker and the published linker was calculated (in terms of number of different nucleotides); a valid linker was accepted if this distance differed by up to two nucleotides. If no valid linker was found, the fragment was rejected. Table S2 summarizes the number of fragments published by Ingolia et al. (see Table S1, column 2) and the percentage of processed fragments after removing the attached linker (column 3).
Aligning the fragments directly to the genome resulted in a high number of ambiguous matches. Therefore, fragments were aligned to known transcripts (exons) and spliced junctions. The M. musculus transcripts were derived from the UCSC Genes data set [34] and the alignment was performed using the Bowtie software [35], allowing up to two mismatches.
As mentioned by Ingolia et al., fragments of different lengths tend to have slighter different A site locations, therefore the beginning of the A site for fragments of 29–30/31–33/34–35 nt was defined to begin +15/+16/+17 nt relatively to the 5′ end of the fragment. Additional details about this topic appear in the original work [12].
As summarized in Table S2, part of the processed fragments matched to more than one location. To overcome multiple mapping of a single fragment, we performed the following procedure: first, only fragments aligning to a single location were mapped. In the second iteration, for all fragments aligning to more than one location, the mean read counts in the region of the possible locations was calculated (10 nt before and after the location of the A site for each possible location). These mean read counts defined the probability of an ambiguous fragment to be aligned to only one of the locations.
For each isoform, nucleotide read counts profiles were reconstructed by assembling read counts of relevant exons and spliced junctions. Codon reads were calculated by averaging the obtained reads of each three non-overlapping consecutive nucleotides.
In the original work, the profiles were smoothed using an averaging window of five codons and normalized by the average read counts of codons 800–1000. This normalization assumed that read counts in regions not affected by harringtonine (codons 800–1000) have a similar value (for each one of the run-off profiles apart). When assuming the experiment is reproducible, i.e. ribosomal read counts of all profiles are similar after the first 750 codons (the harringtonine effect did not extend beyond this point for any isoform in the experiment [12]), it is possible to estimate the Starting Location (SL) point of a depleted profile by comparing it to the baseline profile, . The SL of the depleted ribosomal profile of each isoform was defined as the position beyond the first 40 codons, where the normalized ribosomal density profile exceeded a value of 0.5. In this work, this parameter is defined as the recovery factor. Isoforms with SL points not satisfying (see Table S3) were discarded. When smoothing the profiles with longer averaging windows, the number of isoforms with non-physical SL points reduced to 141 (out of 785, see also Table S3).
Further study of the nature of ribosomal profiles revealed that the original SL estimation method suffers from some difficulties: the results presented in Figure S3-A show that read counts in regions not affected by harringtonine (beyond the 750th codon, excluding the last 20 codons) have a high variability, therefore their average read count value cannot be used for normalizing the ribosomal profiles. In addition, in the original method the SL point was defined as the location where the run-off profile exceeded the threshold value 0.5. This criterion assumes again that profiles are relatively homogenous, and small spikes caused by noises can be filtered by first smoothing them. However, the results in Figure S3-B show that different profiles have a high read counts variability, also suggesting that ribosomal read counts could be position dependent, making the comparison of the run-off profile to a static threshold of 0.5 problematic.
To overcome these issues, in the current work we suggested scaling each run-off profile to the baseline profile by a dynamic factor that derives from the read counts beyond the 750th codon of both profiles (excluding the last 20 codons). This factor is set to minimize the distance between these regions. In the current study, we also tested the effect of the smoothing window size (10/15/20/25/30 codons) on the number of genes with physical SL points, as presented in Table S5. The SL location of each isoform was defined as the position beyond the first 40 codons, where the ribosomal density profile exceeded the value of the profile multiplied by the recovery factor. This created a dynamic threshold for the run-off profiles to be compared to. The influence of the recovery factor on the number of genes with physical SL points was also evaluated, as presented in Table S6. In addition, to improve robustness of the method to local bursts of noise, an SL point was defined to be valid if 50% of the next 20 points could also exceed the dynamic threshold. The optimal smoothing window size and recovery factor were selected to maximize the number of genes whose SL points were physically estimated (), resulting in a window size of 30 codons and a recovery factor of 0.5 (see Table S3, S4, S5, S6).
To compare between the methods' ability to correctly estimate SL points in a noisy environment, both the original [12] and the newly suggested methods were also evaluated on synthetic data created using the TASEP model (e.g. see [22]). SL points were estimated for different run-off times and different levels of additive noise (see Methods, evaluating the error rate of the SL points). Figures S4, S5, S6 show the mean and standard deviation estimation error as function of noise level and size of the smoothing window for both estimation methods. As seen from the results, on the simulative data the newly suggested method achieved a lower estimation error for all levels of noise and smoothing window sizes.
For comparison, in this work, the various tested measures were calculated based on SL points estimated using both methods. The smoothing window size was set to 30 codons and the recovery factor was set to 0.5. The figures in the main text were generated using the new method with these parameters. More details appear in Text S1.
Folding energy (FE) of a nucleotide was defined as the folding energy of a 40 nt segment, starting from the current nucleotide. The segment's FE was calculated using the rnafold Matlab function [36]. The FE of a gene (segment) was defined as the average folding energy of its nucleotides.
Codon tAI values were calculated according to [20], using tRNA copy numbers published in http://gtrnadb.ucsc.edu/Mmusc10/. The tAI value of a segment was calculated using:Where is the relative adaptiveness of codon of type , the index of the codon and the number of codons in segment . Let be the copy number of the anti-codon that recognizes the codon, and let be the selective constraint of the codon/anti-codon coupling efficiency. Then, the absolute adaptiveness value of a codon is defined byThe relative adaptiveness value of a codon is obtained by normalizing with the maximal value among its 61 values (for specific values see Table S10).
To calculate CAI of a segment, codons were ranked according to their usage in ribosomal proteins (). Using these frequencies, the CAI of a segment was similarly defined in the following manner:
For each gene, a vector of charges was defined by assigning +1 to positively charged amino acids (Arg and Lys) and −1 to negatively charged amino acids (Asp and Glu). The charge of other amino acids was set to 0. A sliding window of 40 codons was applied on the charge vector to smoothen the charge effect on the mRNA. The overall charge of a segment was defined as the sum of its charges.
To enable analysis of various features in a simulated environment, ribosomal densities of the analyzed isoforms in this work were calculated using the TASEP biophysical translation model, previously used in different studies (e.g. [22], [37]). The mRNA was modeled using a lattice of sites, representing the number of codons of the isoform. Each ribosome was defined to cover 11 codons and the A site was located at the sixth codon. During translation, any codon could be covered at a time by a single ribosome at most. In each step of the simulation, a single ribosome was allowed to attach itself to the lattice or advance to the next codon if the first/next six codons were not occupied. The time between initiation attempts was set to be exponentially distributed with a constant rate . Similarly, the time between jump attempts from site to site was assumed to be exponentially distributed with rate .
The time between events, (initiation or jumping between sites) is therefore exponentially distributed (minimum of exponentially distributed random variables) with rate:where describes the site (codon) number on the lattice and if codon is being translated, otherwise . Therefore the initiation probability is given by and the probability of a ribosome to jump from site to is given by .
The parameter was determined for each codon type according to its translation efficiency, estimated by the tAI measure (for specific values see Table S10). The initiation rate was studied for different values, depicting different initiation rate regimes.
To achieve an initial scattering of the ribosomes on the mRNA, 106 simulations steps (events) were performed. This number of steps was selected to enable full initial steady state ribosomal cover for the analyzed genes. In general, longer genes or genes with low initiation rates (relative to the gene's codon translation efficiencies) require a higher number of simulation steps to achieve this condition.
To calculate ribosomal density profiles, we simulated another 107 steps. In each step, the simulation updated the time each site was translated by a ribosome. The final vector of times representing the total time a site was translated by a ribosome was then normalized by the total time of the simulation. In addition, the final scattering location of the ribosomes on the mRNA was saved.
Simulated ribosomal profiles were created by using three different initiation rate regimes : low, high and proportional to the genes' mean ribosomal read counts. The low initiation rate was set to be 10% of lowest codon translation rate (based on the tAI measure), while the high initiation rate was set to be twice the value of the highest codon translation rate (based on the tAI measure).
Proportional initiation rates were set for each isoform according to its measured mean ribosomal read counts (excluding the first 60 and last 40 codons). This initiation rate type assumed that in general, genes with higher mRNA and ribosomal densities levels (thus higher ribosomal read counts) are more highly expressed, therefore their initiation rate should be higher. Thus, for this regime initiation rate of the isoform with the lowest mean read counts was set as half of the slowest codon translation rate, while the initiation rate of the isoform with the highest mean ribosomal read counts value was set to twice the value of the highest codon translation rate. Initiation rates for the rest of the genes were set with equal distance between these two extremes, according to the genes' mean ribosomal read counts.
To simulate ribosomal profiles for different run-off times, the TASEP model was run 106 simulations steps to achieve a steady state ribosomal spread on the mRNA. Initiation halting was simulated for 100 different run-off times, defined bywhere was defined to be the maximal translation time of a codon (based on the tAI measure).
To simulate numerous mRNA copies per gene, for each run-off time and analyzed gene, 500 ribosomal density profiles were calculated and those were averaged with equal weight to obtain a representative ribosomal profile for each gene and run-off time. More details appear in Text S1.
In the original work, it was claimed that translation elongation is constant throughout the translation of the mRNA. To test this hypothesis, we created synthetic genes using the length of the analyzed genes in this work, but with codons of equal translation efficiency, which was set as the mean tAI value of the codons calculated in M. musculus. Using the TASEP model, the ribosomal profile of each one of the synthetic genes was created for different run-off times for low and high initiation rates. More details appear in Text S1.
To allow accuracy evaluation of the original and new method for estimating SL points, ribosomal density profiles with specific run-off times were created, as previously described. To test the robustness of the estimation method for different levels of noise, additive uniformly distributed noise of different levels was added prior to estimating the SL points of each analyzed gene. The noise level added to each gene was selected to be proportional to its maximal simulated ribosomal density, such thatLet us mark by the estimated SL location for a noise level characterized by . The estimation error is then defined byThe SL points for all simulated genes for run-off times of were calculated for the above noise levels. The general estimation error for a given noise level was defined as the average estimation error for all tested genes and run-off times. More details in Text S1.
For each simulated ribosomal profile (based on the real analyzed genes) and various initiation rates (low/high/proportional) the estimated SL points were calculated for run-off times of . These points were selected to resemble the real aggregated profiles (see Figures S1, S2).
These SL points were used for calculating the ratio between the estimated velocities and , analysis of the measure and correlation between the ratio of the mean read counts and the ratio of the segments length.
In addition, these measures were also calculated for the simulated ribosomal profiles of genes composed of codons with equal translation efficiency (same run-off times as described above), for low and high initiation rate. More details appear in Text S1.
To simulate ribosomal densities profiles obtained after filtering long fragments (created by adjacent ribosomes), for each simulated mRNA copy, ribosomal read counts were considered only for fragments covered by ribosomes that had a least one codon gap between themselves and their neighboring ribosomes, on both sides (using the final ribosome scattering on the mRNA). More details appear in Text S1.
To simulate a non-uniform effect of the propagation time of harringtonine, the analyzed isoforms were simulated using the TASEP model for low initiation rate (this regime results in profiles similar to the real measured profiles, see Figure S1, S2). For each gene, initiation halting was calculated for the following run-off times when using 500 mRNA copies per gene. Let us denote the ribosomal profile of gene calculated for the mRNA copy and run off time by Let us denote the aggregated profile of gene for the run-off time by . The non-uniform effect of harringtonine was simulated for each gene by aggregating different run-off profiles is the following manner:Where is the number of mRNA copies simulated per gene and is a random variable, such that . The simulation was calculated for . The higher the value, the more prominent the effect of the non-uniform propagation time of haringtonine. More details appear in Text S1.
The comparison between the translation velocity and was done using the paired Wilcoxon test, as supplied in the Matlab 2011b software. The comparison between the , , , measures, calculated on the real ribosomal profiles and on the simulated ribosomal profiles, was done using the two samples Kolmogorov-Smirnov (KS)-test. The correlation between the segments' length and their tAI/CAI/FE/charge properties was calculated using partial Spearman correlation, as supplied in the Matlab 2011b software. The comparison between the translation velocities of segments with top/bottom 20%–50% of the tAI/CAI/FE/charge properties that appear in the supplementary results was calculated using the unpaired t-test and the two samples KS-test. The value of the tAI/CAI/FE/charge in the first and second segment ( and ) was also compared using a Wilcolxon test. The correlation between the tAI/CAI and gene length was calculated using Spearman correlation.
Before we performed partial correlation between tAI/CAI/FE/charge measurements and segment length we binned the data in the following manner: first, segments were sorted according to their length and then divided into bins of 15 samples. For each bin, the average length/tAI/CAI/folding energy/charge was calculated in order to reduce noise.
|
10.1371/journal.pgen.1003901 | The Maternal-to-Zygotic Transition Targets Actin to Promote Robustness during Morphogenesis | Robustness is a property built into biological systems to ensure stereotypical outcomes despite fluctuating inputs from gene dosage, biochemical noise, and the environment. During development, robustness safeguards embryos against structural and functional defects. Yet, our understanding of how robustness is achieved in embryos is limited. While much attention has been paid to the role of gene and signaling networks in promoting robust cell fate determination, little has been done to rigorously assay how mechanical processes like morphogenesis are designed to buffer against variable conditions. Here we show that the cell shape changes that drive morphogenesis can be made robust by mechanisms targeting the actin cytoskeleton. We identified two novel members of the Vinculin/α-Catenin Superfamily that work together to promote robustness during Drosophila cellularization, the dramatic tissue-building event that generates the primary epithelium of the embryo. We find that zygotically-expressed Serendipity-α (Sry-α) and maternally-loaded Spitting Image (Spt) share a redundant, actin-regulating activity during cellularization. Spt alone is sufficient for cellularization at an optimal temperature, but both Spt plus Sry-α are required at high temperature and when actin assembly is compromised by genetic perturbation. Our results offer a clear example of how the maternal and zygotic genomes interact to promote the robustness of early developmental events. Specifically, the Spt and Sry-α collaboration is informative when it comes to genes that show both a maternal and zygotic requirement during a given morphogenetic process. For the cellularization of Drosophilids, Sry-α and its expression profile may represent a genetic adaptive trait with the sole purpose of making this extreme event more reliable. Since all morphogenesis depends on cytoskeletal remodeling, both in embryos and adults, we suggest that robustness-promoting mechanisms aimed at actin could be effective at all life stages.
| Every embryo develops under its own unique set of circumstances, with variable inputs coming from mother, father, and the environment. To then ensure a reliable outcome, mechanisms are built into development to buffer against challenges like genetic deficiency, maternal fever, alcohol exposure, etc. This buffering, called “robustness”, can be overwhelmed, ending in miscarriage, pre-mature birth, and structural and functional birth defects. Thus, we need to understand how robustness arises in order to define an embryo's susceptibilities to genetic background and environment; and to ultimately promote healthy reproduction. In this work we provide new insight into how morphogenesis, the process of tissue building in embryos, is made more robust. First, we show that early gene expression by the embryo, or zygote, supplements the stockpile of proteins already supplied by the mother to ensure the robustness of early morphogenesis. Specifically, our data suggests that a specific gene, sry-α, and its expression by the embryo at the maternal-to-zygotic transition, is a genetic adaptation with the sole function of making the first tissue building event in the fruit fly more robust. In addition, we show that the robustness of this morphogenetic event is promoted by mechanisms regulating the actin cytoskeleton.
| Every embryo develops under its own unique set of circumstances and challenges. To then ensure a reliable outcome, mechanisms are built into development to buffer against fluctuations in genetic, biochemical, and environmental inputs [1]. This buffering, called “robustness”, can be overwhelmed, ending in miscarriage, shortened gestation, and structural and functional birth defects [2]. Thus, we need to understand how developmental robustness arises in order to define an embryo's susceptibilities to genetic/epigenetic background and environment; and to ultimately promote healthy reproduction.
Many mechanisms are used to buffer biological systems against fluctuating inputs, including redundant protein function [3], [4], secondary or “shadow” enhancers [5], [6], smart network design [7]–[10], and chaperone activity [3], [11]–[14]. Among developmental systems, a rigorous quantitative understanding of these mechanisms has been largely limited to examples where cell fate decisions are made, and robustness is fostered at the level of gene expression [5], [6], [10], [15] or signaling [9], [16], [17]. For morphogenesis, which translates cell fate decisions into embryonic form, the detailed characterization of specific buffering mechanisms has been slower to come. Morphogenesis requires activities that span nuclei, cytoplasm, and whole tissues, and is driven by cell shape change [18], [19]. So robustness could be promoted at many levels (e.g. gene expression, signaling, membrane dynamics, cytoskeletal remodeling, and cell adhesion). But we do not know enough about the molecular and mechanical underpinnings of morphogenesis to predict where its greatest susceptibilities are, or where buffering mechanisms would be most effective. Specifically, we lack a comprehensive understanding of how cell biological steps convert gene expression into reliable tissue-building events. Actin and microtubules seem like good targets for robustness-promoting mechanisms during morphogenesis because they drive cell-shape change and modulate the mechanics of cells and tissues [20], [21]; however, experimental support for this is lacking.
In order to identify the mechanisms that promote robustness, the outcomes of the process in question must be quantifiable [1]. For example, we know of robustness-promoting mechanisms for cell fate decisions in development because the outcomes are binary, and typically happen along well-separated spatial dimensions so that fidelity can be readily tracked and quantified over a range of perturbations [5], [6], [9], [10], [15]–[17]. For morphogenesis, which is spatially complex, challenging to image, and has long been scored by qualitative rather than quantitative methods, fidelity is not easily measured. Consequently, the number of well-tested examples that show how robust morphogenesis is achieved remains low, in the context of both individual cell shape change and whole tissue remodeling [22], [23]. To address this gap in our knowledge, we are using the first tissue-building event in the fly embryo, cellularization, as a simple, quantifiable model to study robustness. Fly embryos first develop as a syncytium, passing through 13 mitoses with no intervening cytokinesis. Then, during cell cycle 14 the embryo undergoes cellularization, during which ∼6000 cortically-anchored nuclei are simultaneously packaged into a sheet of cells that will be the primary epithelium (Figure 1A) [24]. Cellularization takes ∼60 minutes and plasma membrane furrows ingress 35 µm, cutting straight between adjacent nuclei to form mononucleate cells. This simple architecture allows unambiguous quantification of the fidelity of cellularization, where furrow failures or regressions show up as multinucleate cells, and hundreds to thousands of packaging events can be assayed per embryo to generate a ratio of mononucleate cells-to-nuclei (ratio = 1 in wild-type embryos; Figure 1A).
Fly embryos cellularize just after the maternal-to-zygotic transition (MZT) [24], when transcription from the zygotic genome starts to maximally impact the developmental program [25]. Thus, the few zygotic genes that are required for cellularization have long been thought of as “switches” to control this morphogenetic event [26], [27]. However, we now report a new role for one of these long-supposed switches, Sry-α. We identify Sry-α as a zygotic gene product that is expressed at the MZT, not to control cellularization, but rather to make it robust in the face of both environmental and genetic perturbations. We find that Sry-α acts together with its maternally provided paralog Spt to reinforce the actin cytoskeleton and so promote robust cellularization. Our data provides a clear example of how zygotic contributions, made at the MZT, not only instruct development, but also supplement the maternal machinery to ensure the fidelity of specific morphogenetic events. What's more, our data suggests that this robustness is fostered via regulation of the actin cytoskeleton.
At the start of cellularization, actin filaments (F-actin) accumulate at incipient furrow tips, which in surface views form furrow “canals” around the nuclei (Figure 1A) [24]. Furrow canals are then maintained throughout cellularization, and are required for stable furrow ingression (Figure 1A) [28]–[30]. We previously showed that mutations or drug treatments that reduce F-actin levels in all furrow canals, precipitate the regression of a fraction of furrows (Figure 1A) [29], [30], consistent with multinucleation phenotypes reported for actin regulators like Rho1 GTPase, RhoGEF2, and the Formin, Diaphanous [28], [31], [32]. In an effort to identify other actin regulators that are required for the fidelity of cellularization, we examined a poorly characterized mutant, serendipity-α (sry-α), that similarly displays a multinucleation phenotype (Figure 1B) [26], [27]. The sry-α gene was previously mapped, and is expressed at the MZT just prior to cellularization [27]. Consequently, sry-α has long been thought of as a developmental cue that provides some new activity to trigger cellularization [26], [27]. We found that all furrow canals in sry-α null mutants (sry-α−/−) have significantly reduced levels of F-actin compared to wild-type, throughout cellularization (Figure 1C, 1D). In addition, incipient furrow canals in sry-α−/− mutants display an increased number of Amphiphysin tubules (Figure 1E), which indicates promiscuous endocytosis upon F-actin reduction [29], [33]. These results show that Sry-α regulates F-actin levels in furrow canals during cellularization.
Based on remote homology searches, including PHYRE [34] and I-TASSER [35], we found that Sry-α is a novel member of the Vinculin/α-Catenin Superfamily (Figure 2A, 2B) [36]. Our analysis also identified a Sry-α paralog in the D. melanogaster genome that we called Spitting Image (Spt, CG8247). Sry-α and Spt align with the middle sequences of Vinculin and α-Catenin, including the Vinculin-Homology 2 domain (VH2; Figure 2A, Figure S1) [37], [38]. Like α-Catulin, Sry-α and Spt represent a distinct clade of the Vinculin/α-Catenin Superfamily (Figure 2B, Figure S2) [39], [40]. Based on a “roll-call” analysis of orthologs in organisms with fully sequenced genomes, sry-α and spt co-exist in all Drosophilids, while spt alone is present in other insects (see Table S1). In higher metazoans, PHYRE analysis also identified other uncharacterized proteins, like Sry-α and Spt, which share remote homology with the middle sequences of Vinculin and α-Catenin (Figure S1, S2). Members of the Vinculin/α-Catenin Superfamily peripherally associate with the plasma membrane and interact with the actin cytoskeleton [36]. Thus, this evolutionary relationship is functionally consistent with a role for Sry-α and Spt at the actin cortex during cellularization.
To examine the relationship between Sry-α and Spt, we asked if these paralogs have unique or overlapping functions. Both Sry-α and HA-tagged Spt localize to F-actin rich furrow canals in cellularizing embryos (Figure 2C, 2D). In addition, we used an F-actin co-sedimentation assay to show that both Sry-α and Spt bind F-actin directly. Recombinant Sry-α and Spt proteins were purified from insect cells and mixed with F-actin. Upon centrifugation, F-actin and its interacting proteins pellet (α-Actinin; Figure 2E, 2F), while unbound proteins remain in the supernatant (GST; Figure 2E, 2F). For both Sry-α and Spt, we detected a significant fraction of F-actin bound protein (Figure 2E, 2F). Thus, the co-localization of Sry-α and Spt, as well as their biochemical activity, suggest that they could act redundantly during cellularization.
To look for overlapping in vivo functions, we first used RNAi [41] and found that spt knockdown (sptRNAi) causes multinucleation during cellularization, which is qualitatively indistinguishable from either the sry-α−/− genetic mutant or sry-α RNAi (sry-αRNAi; Figure 3, Figure S3). Additionally, embryos with double sptRNAi plus sry-αRNAi knockdown display a strongly enhanced multinucleation phenotype (Figure 3B, 3C). Together, the localization, biochemistry, and RNAi phenotypes suggest that Sry-α and Spt share redundant functions during cellularization. To then confirm this, we tested whether Spt overexpression (sptOE) can rescue sry-α−/− multinucleation. We used Sry-α immunostaining to genotype embryos (Figure S4A), and found that 100% of sry-α−/− mutants are rescued by sptOE (Figure 4A, 4B, Figure S4). Rescue of sry-α−/− multinucleation by sptOE is equivalent to that accomplished with a genomic construct encoding sry-α itself (Figure S4B). We also confirmed that sptOE restores F-actin to wild-type levels in sry-α−/− mutants (Figure 4C). Thus, Sry-α and Spt share an overlapping actin regulating function during cellularization.
These results challenge a long-standing idea that sry-α is zygotically expressed at the MZT to supply some new activity that instructs cellularization to proceed [26], [27]. In fact, the Sry-α activity may already be available via Spt. We compared the developmental expression profiles of endogenous Sry-α and Spt. As previously shown, sry-α transcript and protein are expressed at the MZT, in a pulse that just coincides with cellularization (Figure 5A, 5B) [27]. Conversely, spt transcript and protein are provided maternally, and Spt levels persist throughout and far beyond cellularization (Figure 5A, 5B, Figure S5). That is, a pulse of zygotically expressed Sry-α adds to a pool of maternally provided Spt during cellularization. (i.e. Sry-α expression only boosts an already existing Spt activity in the embryo.) Given that Spt can completely replace Sry-α during cellularization (Figure 4, Figure S4), Sry-α does not supply some unique activity that triggers the process. Instead, the expression profiles suggest that the level of Spt plus Sry-α is somehow critical for the successful progression of cellularization.
It was previously shown that the co-expression of paralogs in worms and yeast promotes biological robustness [3], [4]. Presumably, these “gene duplicates” provide overlapping functions, and so replace or supplement each other in the face of internal and external perturbations [3], [4]. Hence, we hypothesized that there is a threshold level of Sry-α plus Spt activity that is required for cellularization to proceed with high fidelity. This robustness hypothesis predicts that at an optimal condition, Sry-α function is dispensable and Spt alone can support successful cellularization, whereas both are needed at sub-optimal conditions. To test this, we assayed for multinucleation phenotypes in sry-α−/− null mutants at an optimal temperature. We chose 18°C, the lowest temperature at which D. melanogaster thrives. We reasoned that the lower temperature would reduce the demand on the actin cytoskeleton, because F-actin is more stable at lower temperature [42]. As predicted, we found that multinucleation is suppressed in sry-α−/− mutants that are reared at 18°C (Figure 6A, 6B). At this temperature, the ratio of mononucleate cells to nuclei in sry-α−/− mutants is not significantly different than wild-type (Figure 6B). Nor did we detect changes in the dimensions of the cells that formed for sry-α−/− mutants at 18°C (data not shown). Thus, Spt activity is sufficient for cellularization to proceed at an optimal condition. However, multinucleation in sry-α−/− mutants was increasingly severe at higher temperatures (25–32°C; Figure 6A, 6B), showing that Spt is not enough to ensure reliable cellularization when conditions deviate from the optimal.
A second prediction of the robustness hypothesis is that reducing sry-α dosage will make cellularization more likely to fail when the embryo is challenged by perturbations [3], [4]. To test this, we reduced the Sry-α level using genetic heterozygosity, and looked for multinucleation at an extreme temperature of 32°C. This temperature marks an upper limit at which D. melanogaster embryos can survive, but developmental events are measurably impaired [5], [6]. We found that the occurrence of multinucleation at 32°C is significantly increased for sry-α heterozygotes (sry-α+/−; Figure 7A, 7B). Genotypes were confirmed by RNA FISH (Figure S6) [43]. So while Spt alone is sufficient for cellularization at an optimal temperature, Spt plus two copies of the sry-α gene are required when environmental conditions are extreme. This suggests that the expression of Sry-α serves as a robustness-promoting mechanism for cellularization.
A hallmark of robustness-promoting mechanisms is that they respond equally well to different kinds of perturbations (e.g. genetic, biochemical, or environmental) [44]. For example, shadow enhancers for the fly genes snail and shavenbaby promote robust gene expression at high temperature, and when mutations reduce input from their respective activation pathways [5], [6]. In addition, a systems-level analysis in yeast showed that genes that promote mutational robustness also promote high fitness in a wide range of stressful environments [44]. Thus, we also tested whether the Sry-α level ensures the fidelity of cellularization upon genetic perturbation. We checked for multinucleation in sry-α+/− heterozygotes that carry only a half maternal dose of profilin (½profilin). Profilin is an actin accessory protein that promotes actin polymerization [45]. Strikingly, we saw multinucleation in sry-α+/− heterozygotes in the ½profilin background even at 18°C (Figure 7C, 7D). We conclude that the fidelity of cellularization, in the face of both environmental and genetic perturbations, critically depends on the level of Sry-α.
Our data support a model wherein maternal Spt plus zygotic Sry-α work together to promote the robustness of cellularization. Our findings refute the idea that there is a clear passing off of developmental control from the maternal to the zygotic machinery at the MZT [25]–[27], [46]. For example, the maternal genome was previously thought to provide the basic cellular machinery for cellularization, while Sry-α and other zygotic players provided the instructions [26], [27]. More recently zygotic gene products have been shown to actively degrade maternal mRNAs, arguing that there may be a clean break from maternal to zygotic control [47]–[49]. In both cases, the zygotic contribution is largely viewed as being instructive. But our data speaks to a more collaborative interaction between the maternal and zygotic genomes: We show that cellularization can proceed with no input from zygotic Sry-α. Instead of controlling cellularization, we find that Sry-α actually adds to the activity of maternal Spt to make this morphogenetic event more reliable in the face of environmental and genetic perturbation. Sry-α is not taking over for the maternal product because it is only expressed in a pulse during the demanding event of cellularization, while maternal Spt persists far beyond cellularization. So, the relationship of Spt and Sry-α illustrates with exceptional clarity that the maternal and zygotic genomes also work together, with redundant activities, to make development robust. Certainly, this collaboration is likely to be broadly conserved (e.g. maternal plus zygotic contributions of Rac proteins in C. elegans and Cadherins in Xenopus support the progression of specific morphogenetic events [50], [51]).
Since maternal RNAs and proteins are loaded into oocytes and eggs long before they act in development (up to months) [25], their levels may not be very reliable [46]. Overlapping activities encoded by the zygotic genome could then buffer this variability and ensure successful progression through early embryogenesis. In our case, this relationship was revealed by the distinct expression profiles of paralogs provided from the maternal and zygotic genomes. But the same end is likely accomplished by expressing a single gene both maternally and zygotically. In fact, there is the recurrent observation that some proteins are expressed zygotically, while a significant maternal pool still persists (e.g. Drosophila β-Catenin) [46], [52]. In D. melanogaster, roughly 5% of the genome displays this expression profile, with a maternal contribution supplemented by zygotic expression at the MZT [47].
But why split the contribution between the maternal and zygotic genomes? For cellularization, why is the same level of robustness not achieved by simply expressing more maternal Spt? We can envision at least two possibilities: either Spt and Sry-α have some distinct functions, perhaps at different developmental stages; and/or high levels of Spt/Sry-α protein are harmful.
Sry-α and its expression at the MZT may be an adaptive genetic trait with its function, perhaps its sole function, serving to buffer cellularization against external and internal perturbations. According to our roll-call analysis, both spt and sry-α are present in the genomes of all Drosophilids, while other insects only encode spt (see Table S1). Some aspect of development specific to Drosophila may then stabilize the strategy of maternal Spt plus zygotic Sry-α. For example, cellularization in Drosophila builds tall columnar cells around thousands of nuclei, and so may be more demanding than cellularization in other insects where shorter cuboidal cells form around fewer nuclei [53]–[55]. Alternatively, Drosophilids share a fast rate of embryogenesis in comparison to many other insects [53], [54], [56], which could make their cellularization more difficult. Thus, the pulse of zygotic Sry-α may be advantageous for meeting the challenge of a very extreme cellularization event in Drosophila.
Our data suggests that zygotic Sry-α adds to the activity of maternal Spt to regulate actin, and so ensure the fidelity of cellularization. A significant future challenge will be understanding what specific actin-based activities promote robust morphogenesis. For example, F-actin is a critical determinant of tissue mechanical properties during development because it assembles with Myosin-2 and other crosslinkers, within single cells, to form a cortical network that (1) governs the rigidity and shape of the whole embryo [57]–[59]; and (2) generates the forces for and resistance to the cell shape changes that drive morphogenesis [60]–[64]. Thus, actin could promote robustness by buffering the mechanical properties of cells and tissues. In fact, Spt and Sry-α bind F-actin directly. Also, Spt and Sry-α are related to the F-actin crosslinker Vinculin, and they contain the conserved M-domain at the VH2, which in Vinculin dimerizes to support F-actin crosslinking [38]. So Spt and Sry-α could promote robust cellularization by crosslinking F-actin and modulating tissue mechanics. Alternatively, F-actin also controls the membrane remodeling that accompanies cell shape change and morphogenesis. Specific to Drosophila cellularization, F-actin antagonizes endocytosis to favor plasma membrane growth and furrow ingression [29], [33]. Consequently, F-actin could also promote robustness by controlling the membrane dynamics of morphogenesis. Since all morphogenesis depends on actin remodeling, both in embryos [65] and adults [66], we believe that robustness-promoting mechanisms that target actin are likely to be ubiquitous.
The sry-α−/− and sry-α+/− embryos were collected from Df(3R)X3F/TM3, Sb [26] crossed to OreR wild-type flies. The nulloX embryos were collected from C(1)DX, ywf [29], [30]. For sry-α−/− plus sptOE, first matαtub-GAL4 (II) was crossed with UASp-Spt-HA (III) (this paper) to make stock matαtub-Gal4; UASp-Spt-HA. Second, these flies were crossed with Df(3R)X3F/TM3, Sb, and finally embryos were collected from matαtub-GAL4/+; UASp-Spt-HA/Df(3R)X3F mothers crossed with their sibling matαtub-GAL4/+; UASp-Spt-HA/Df(3R)X3F fathers. For sry-α−/− plus sry-αrescue, Df(3R)X3F/TM3, Sb was crossed to sry-α genomic rescue stock (II) (gift of E. Wieschaus). For ½profilin, chic221 cn1/CyO; ry506 (Bloomington Stock Center #107932) was crossed with Df(3R)X3F/TM3, Sb; and embryos were collected from chic221 cn1/+; Df(3R)X3F/ry506 mothers crossed with their sibling fathers chic221 cn1/+; Df(3R)X3F/ry506. For RT-PCR and Western Blotting, OreR was used. For RNAi imaging, embryos were injected from stock Gap43-mCherry/CyO; Histone-GFP/TM3, Sb (parental stocks gifts of A. Martin and E. Wieschaus).
For UASp-Spt-HA flies, the coding sequence of D. melanogaster spt (CG8247) was fused with a C-terminal hemagglutinin (HA) tag, and cloned into pP(UASP) vector. Transgenesis and mapping followed standard methods (BestGene, Inc.). For anti-Spt antibody, the coding sequence of spt fused with a C-terminal histidine tag was cloned into pET vector, and recombinant protein was purified from E. coli. Antibodies were produced in guinea pigs according to standard methods (Panigen, Inc.). According to the publicly available data on FlyBase (flybase.org), spt is not expressed in adult heads. Thus, adult heads served as the negative control for antibody characterization (Figure S5).
Recombinant proteins GST-Sry-α (full length) and GST-SptΔ (amino acids 250–461) were produced in Sf9 cells (Baylor College of Medicine Baculovirus Core/Proteomics Shared Resource). The Spt truncate was used because full-length protein is insoluble in both bacterial and insect expression systems. Proteins were purified with Glutathione Sepharose (GE Healthcare) and dialyzed in F-buffer (20 mM Tris-HCl pH 7.5, 2 mM DTT, 2.5 mM MgCl2, 75 mM KCl and 10 mM NaCl). α-Actinin (Cytoskeleton, Inc.) was used as the positive control, and GST for the negative control was purified from Sf9 cells.
Rabbit muscle monomeric actin (G-actin) was extracted from Rabbit Muscle Acetone powder (Pel-Freez Biologicals). 20 mM G-actin was maintained in G-buffer (2 mM Tris-HCl pH 8.0, 0.2 mM ATP, 0.5 mM DTT, and 0.2 mM CaCl2) until polymerization in F-buffer by adding buffer and salt and incubating at room temperature for 1 hour. For the co-sedimentation assay, purified proteins were pre-cleared by centrifugation at 900,000 rpm for 30 minutes at 4°C to remove any precipitate. Pre-cleared proteins (∼1 µg) were incubated with F-actin for 30 min on ice and then centrifuged at 900,000 rpm for 30 min at 4°C. Supernatant was removed, and replaced by an equal volume of 3X sample buffer to resuspend pellets. Equal volumes of supernatants and pellets were separated on 5–10% SDS-PAGE gels and transferred to nitrocellulose. GST-Sry-α, GST-SptΔ, and GST were detected by 1∶10,000 mouse anti-GST (ab19256, Abcam). α-Actinin was detected by 1∶1000 anti-α-Actinin (A7811, Sigma-Aldrich). The goat secondary antibodies were HRP conjugates used at 1∶10,000 (Jackson ImmunoResearch).
Embryo collection cups were set up on apple juice plates according to published protocols [29], [30], [41]. For specific temperature incubations, collection cups were housed in air incubators of 18°C, 25°C, or 32°C within a range of ±1°C for 5, 4, or 3 hours, respectively. Plates were harvested and embryos fixed immediately.
For immunofluorescence, embryos were fixed in boiling salt buffer; or 4% paraformaldehyde, 0.1 M phosphate buffer (pH 7.4): heptane (1∶1) [29]. Antibody concentrations and fixation methods are listed in Table S2. Due to Myosin-2 antibody incompatibility with FISH, Septin (Peanut) was used as the furrow canal marker for experiments where heterozygosity was scored. For F-actin staining, embryos were fixed in 8% paraformaldehyde, 0.1 M phosphate buffer (pH 7.4): heptane (1∶1) and hand-peeled for staining with 5 U ml−1 Alexa 488-phalloidin (Invitrogen-Molecular Probes). For nuclear staining, Hoescht 33342 was used at either 0.25 µg ml−1 for standard immunofluorescence or 1.0 µg ml−1 for FISH (Invitrogen-Molecular Probes).
For FISH, 44 oligonucleotide probes, covering the sry-α coding sequence were synthesized (Biosearch Technologies), and then labeled with Alexa 488 [43]. Embryos were fixed in 4% paraformaldehyde, 0.1 M phosphate buffer (pH 7.4): heptane (1∶1) for hybridization with 50 µM Alexa 488-sry-α probe followed by immunofluorescence staining.
For fixed and living embryos, images were collected on a Zeiss LSM 710 confocal microscope with a 40X/1.2 numerical aperture water-immersion objective (Carl Zeiss, Inc.). Images were collected at a zoom of two, with resolution of 104 nm per pixel.
To segment images, the Image Processing Toolbox in MATLAB (MathWorks) was used. From raw images, the furrow canal network and nuclei masks were generated using a series of morphological operations, as follows. For furrow canal network masks: (1) A Gaussian filter was applied to the raw image, to retain only coarse features of the furrow canals. (2) An intensity threshold was selected manually and applied to generate a preliminary mask. (3) The mask was thinned iteratively until a 1-pixel width network was produced. For nuclei masks: (4) A Gaussian filter was applied to the raw image to smooth out noise, yet retain fine features of the image. A low intensity threshold was selected manually and applied to capture weak furrow canal links present in multinucleated cells. The resulting mask was closed to join disconnected furrow canal links, and thinned iteratively to capture all nuclei separations. Remaining disconnected furrow canal links were removed. (5) The preliminary furrow canal network mask resulting from step 2 above was thinned and added to the mask obtained in step 4. This mask was dilated and inverted to generate a preliminary nuclei mask. Holes inside the mask were filled. (6) Finally, a high threshold was applied to the result of step 1 above; the resulting mask was inverted and then multiplied by the result of step 5 to eliminate boundary artifacts produced by thinning operations.
For quantifications, levels of F-actin in furrow canals and numbers of Amphiphysin tubules were scored as previously reported [29]. The ratio of mononucleate cells to nuclei was generated by manually counting in two quadrants from a raw, single plane, surface view image (quadrant size = 2835 µm2) collected at the furrow canals; and the mean was calculated per embryo. The percent of embryos displaying multinucleation was counted manually using raw, single plane, surface view images collected at the furrow canals, where an entire embryo side was visible (≥1500 nuclei assayed per embryo).
For genotyping by FISH, maximum intensity projections were generated from image stacks, comprising ∼4 µm depth and ≥150 nuclei; and active sry-α transcription sites were manually counted. Embryos with a maximum of 0, 1, or 2 spots were scored as sry-α−/−, sry-α+/−, or sry-α+/+ (wild-type, WT), respectively. For genotyping by immunofluorescence, images were collected at the same microscope settings, and Sry-α signal was scored as either present or absent.
For RT-PCR, embryos were staged in halocarbon oil 27 under a dissecting microscope. Approximately 100 embryos per stage were homogenized in Trizol (Invitrogen Inc.), and total RNA was extracted in phenol: chloroform (1∶1). Total cDNA was made (SuperScript III First-Strand Synthesis System, Invitrogen), and sequences amplified by PCR. For developmental expression profiles, primers were: actin42a-F, actin42a-R, sry-α-F, sry-α-R, spt-F, and spt-R (for sequences see Table S3). Following RNAi, primers were: actin42a-F, actin42a-R, sry-α-F2, sry-α-R2, spt-F2, and spt-R2 (for sequences see Table S3). Samples were loaded on 1% agarose gels.
For Westerns, 1 hour collections of embryos were incubated at room temperature for 0, 1, 2, 3, 4, or 5 hours respectively. Per stage, 50–100 µl embryos were homogenized in 200 µl 0.05 M Tris pH 8.0, 0.15 M KCl, 0.05M EDTA, 0.5% NP-40, 1X protease inhibitor cocktail (Halt Protease Inhibitor Cocktail, Thermo Scientific). For antibody characterization, 20 adult heads were homogenized in the same buffer. Following a 15 minute spin at 15,000 rpm at 4°C, the cytoplasmic fraction was collected and quantified (Pierce BCA Protein Assay Kit, Thermo Scientific). Equal amounts of total protein were separated on 5–10% SDS-PAGE gels and transferred to nitrocellulose. Sry-α, Spt, HA and α-Tubulin were detected by 1∶5 mouse anti-Sry-α (1G10, Developmental Studies Hybridoma Bank), 1∶500 guinea pig anti-Spt (this paper), 1∶100 rat anti-HA (Roche), and 1∶1000 rat anti-α-Tubulin (T9026, Sigma-Aldrich), respectively. The goat secondary antibodies were HRP conjugates used at 1∶10,000 (Jackson ImmunoResearch).
Approximately 50 pl of sry-α and spt double-stranded RNA was prepared as previously described [41], with primers: sry-α-RNAi-F, sry-α-RNAi-R, spt-RNAi-F, and spt-RNAi-R (for sequences see Table S3), and injected into freshly laid embryos. Following incubation, mitotic cycle 13 embryos were mounted for imaging. RNAi controls were PBS buffer-injected embryos.
Protein sequences were retrieved via UniProt (uniprot.org), and alignments were generated using PROMALS3D (prodata.swmed.edu/promals3d) [37]. Phylogenetic trees were built with PhyML v3.0.1 (pbil.univ-lyon1.fr/software/seaview) [67], and branch supports were tested with the default aLRT SH-like option. In addition, for the small and large trees, bootstrap statistics were determined from 1000 and 500 iterations, respectively. Vinculin from the pre-metazoan M. brevicollis served as the outgroup [36]. For both alignments and trees, the same results were generated using either the entire Sry-α and Spt sequences, or the VH2 domain alone. For the roll call analysis, the D. melanogaster sequences for sry-α and spt were used as input for (1) a PSI-BLAST search of the refseq protein database on NCBI (blast.ncbi.nlm.nih.gov); and (2) a BLAST search of the UniProtKB database on UniProt. For those insects where only one paralog was present, we could not assign them as sry-α or spt based on sequence identity alone because the identity values were the same. Instead, we used the presence or absence of introns as the distinguishing factor: sry-α is intronless, whereas spt has introns in all Drosophilids. Thus, all insect proteins including introns were assigned as spt.
|
10.1371/journal.ppat.1000884 | Phylodynamic Reconstruction Reveals Norovirus GII.4 Epidemic Expansions and their Molecular Determinants | Noroviruses are the most common cause of viral gastroenteritis. An increase in the number of globally reported norovirus outbreaks was seen the past decade, especially for outbreaks caused by successive genogroup II genotype 4 (GII.4) variants. Whether this observed increase was due to an upswing in the number of infections, or to a surveillance artifact caused by heightened awareness and concomitant improved reporting, remained unclear. Therefore, we set out to study the population structure and changes thereof of GII.4 strains detected through systematic outbreak surveillance since the early 1990s. We collected 1383 partial polymerase and 194 full capsid GII.4 sequences. A Bayesian MCMC coalescent analysis revealed an increase in the number of GII.4 infections during the last decade. The GII.4 strains included in our analyses evolved at a rate of 4.3–9.0×10−3 mutations per site per year, and share a most recent common ancestor in the early 1980s. Determinants of adaptation in the capsid protein were studied using different maximum likelihood approaches to identify sites subject to diversifying or directional selection and sites that co-evolved. While a number of the computationally determined adaptively evolving sites were on the surface of the capsid and possible subject to immune selection, we also detected sites that were subject to constrained or compensatory evolution due to secondary RNA structures, relevant in virus-replication. We highlight codons that may prove useful in identifying emerging novel variants, and, using these, indicate that the novel 2008 variant is more likely to cause a future epidemic than the 2007 variant. While norovirus infections are generally mild and self-limiting, more severe outcomes of infection frequently occur in elderly and immunocompromized people, and no treatment is available. The observed pattern of continually emerging novel variants of GII.4, causing elevated numbers of infections, is therefore a cause for concern.
| Noroviruses, known as the viruses that cause the ‘stomach flu’ or as the ‘cruise ship virus’, cause sporadic cases and large outbreaks of gastrointestinal illness in humans. An increase in norovirus outbreaks was reported globally around 2002. Doubts remained as to whether this increase was real, or caused by improved detection-techniques and increased awareness. This study was performed to address this ambiguity, and to determine the possible virological causes for such changes. Using a population genetic approach, we studied sequences of epidemic norovirus strains collected through time and we indeed demonstrated expanding epidemic dynamics. Global epidemics were caused by subsequent variants of norovirus, observed in 2002, 2004 and 2006 and at a smaller scale in 1996, whereas no evidence for such epidemic evolutionary patterns occurring previous to these peaks. Based on the sequences analyzed the strains of the genotype under study here were shown to have circulated at least since the early 1980s, and likely earlier. We showed that not only surface exposed sites on the outside of the virus shell were under selective pressure, involved in avoiding host immune responses, but also codons that are apparently conserved for the purpose of virus replication.
| Noroviruses (NoV) are the most common cause of acute viral gastroenteritis [1], [2], with the numbers of reported outbreaks peaking characteristically between November and March in the northern hemisphere [3]. Illness is usually self-limiting and symptoms, comprising acute onset vomiting and watery diarrhea, subside within one to three days [4]. The relevance of studying NoV lies in their high prevalence in the population [5], and in the more severe and prolonged illness that is seen among elderly and immunosuppressed patients [6]–[8]. NoVs are highly infectious, due to the combination of an extremely low infectious dose (an estimated ID50 of less than 20 viral particles [9]), very high levels of shedding (around 108 but up to >1010 RNA copies per gram of stool) and prolonged shedding after clinical recovery [10], [11]. NoV outbreaks, which may affect hundreds of people and are notoriously difficult to control, are primarily associated with places where people are in close contact, for example hospitals and long-term care facilities.
NoVs are a genetically diverse group of positive sense single-stranded RNA viruses from the Caliciviridae family. Their 7.5 kb genome includes three open reading frames (ORFs). The first ORF encodes a polyprotein that is post-translationally processed to form the non-structural proteins, the second and third ORFs encode the major and minor structural proteins; VP1 or the capsid protein and VP2. The viral capsid is formed by 180 copies of the major capsid protein, and governs antigenicity, host-specificity and environmental stability.
NoVs are classified into five distinct genogroups, which are further subdivided into genotypes, based on their amino acid capsid sequence. Molecular epidemiological studies have shown that in recent years approximately 70% of NoV outbreaks among humans have been caused by one dominant genotype, GII.4 [12]–[17].
With continuous surveillance systems in place in some countries since the mid-1990s, it has become apparent that the number of reported NoV outbreaks, and especially those caused by GII.4 strains has risen since the appearance of the 2002 variant of GII.4 [12], [15], [18]–[21]. Since then, genetically distinct GII.4 variants have emerged, and spread rapidly across the world causing epidemic waves of NoV illness [20], [22]. To date, three variants named after the year when they were first detected have been identified in populations across the world (the 2002, 2004 and 2006b variants [22]). The emergence of each of these three variants was followed by ‘hot’ NoV winters with sharply increased numbers of reported outbreaks. Older strains belonging to the lineage designated 1996 were also detected around the world, although surveillance was limited at that time.
The pattern of continuous lineage turnover, referred to as epochal evolution [23], with emerging new variants replacing previously predominant circulating ones, is strongly reminiscent of what is observed in the molecular epidemiology of Influenza A virus (IAV). The evolutionary interaction between IAV and the human immune response results in antigenic drift, illustrated by the characteristic ladder-like tree shapes for hemagglutinin and neuraminidase surface proteins [24]. Long-term partial immunity to the virus induces sharp fitness differences among strains and drives rapid amino acid replacement at key antigenic sites, pinpointed by in vitro and in silico analyses [25], [26]. Whereas antigenic data can be readily generated for IAV, allowing the comparative mapping of antigenic and genetic evolution [27], research of NoV antigenic properties has been hampered by the lack of a simple cell culture model [28]. However, recent publications indicate that the genetic differences between NoV genotypes, and also between variants of the GII.4 genotype, translate into distinct antigenic types, although molecular determinants remain largely unclear [23], [29]. Thus, individuals may be repeatedly infected by strains belonging to different genotypes, and also, because immunity against NoV infection is short-lived at best [30]–[32], possibly repeatedly by strains of the same genotype. As a result, the impact of immune responses on NoV epidemiology remains poorly understood and phylodynamic and molecular adaptation studies may provide some key insights.
In this study, we aimed to provide a rigorous measurement of NoV GII.4 diversity through time, and we investigated viral population expansions in relationship to the increased numbers of infections reported in recent years. Evolutionary and population dynamics of GII.4 NoVs were estimated by a Bayesian coalescent approach, using two different datasets of sequences from strains with known detection dates, between 1987 and 2008. One set of sequences contained full capsid sequences, the other short partial polymerase sequences, which had been obtained for standard-procedure genotyping in NoV surveillance in Europe [13] (http://www.noronet.nl/fbve/) and from the global NoV surveillance network Noronet [22] (http://www.noronet.nl/).
We also tested whether these dynamics differed from neutral expectations, so whether and how they were shaped by selective pressure, and we attempted to further elucidate the molecular determinants of NoV evolutionary and epidemiological dynamics using in silico techniques. To identify the molecular characteristics of NoV GII.4 strain replacement, we investigated both directional and diversifying selection and elucidated capsid protein positions showing evidence for co-evolutionary dynamics acting between sites.
To examine the extent to which recombination has shaped NoV evolution, we analyzed an alignment of 20 GII.4 sequences, two for each variant, spanning the genome from region A (a gene region in ORF1 that is commonly used for genotyping purposes, see Materials and Methods, and [22] and [33]) up to the 3′end of the capsid. Using GARD (Genetic Algorithm for Recombination Detection), VisRD (Visual Recombination Detection) and RDP3 (Recombination detection program), significant phylogenetic variability was identified in this genome region, which could be attributed to a recombination event for the 2003Asia variant, a GII.4 variant previously identified as a recombinant lineage, mainly detected in Asia, and rarely in Europe, Oceania or the Americas [22]. The crossover point lay in the ORF1/2 overlap, a position previously identified as a recombination hotspot in NoV [34], [35]. Further analyses below were based on polymerase and capsid gene sequences that do not include this breakpoint, and no recombination could be detected in those individual data sets using the Phi test [36].
To test whether GII.4 evolution deviated from selective neutrality, we applied a genealogical neutrality test that involved Bayesian coalescent inference, a tree-based summary statistic (DF), and posterior predictive simulation [37]. Using a constant population size demographic prior, the capsid data set (194 sequences, 1623 nt) showed significantly more negative DF than expected (P<0.01), suggesting a selective process that generated a significant excess of mutations on terminal branches. The same was true using an exponential growth prior, but a Bayes factor test did not support an exponential growth scenario (ln BF constant versus exponential growth = −2.47). A Bayes factor comparison also favored a constant population size model over an exponential growth model for the matched polymerase data set (172 sequences, 247 nt)(ln BF = −2.23). In this case, we did not observe a significant difference in DF (P = 0.15). Because the polymerase sequences were considerably shorter, they provided less information to evaluate branch length properties. To counteract the loss of power we increased the number of sequences, and indeed, an analysis of the complete polymerase sequences dataset (1383 sequences, 247 nt) rejected the model of neutral evolution (P = 0.001). As previously introduced [37], this neutrality test relies on relative restricted demographic models governed by a limited number of parameters to capture large-scale demographic trends. To investigate the sensitivity to demographic detail, we extended the posterior predictive simulation procedure to accommodate highly parametric demographic models, which result in a more complex picture of norovirus dynamics (see below). Using a Bayesian skyline plot (BSP) model as demographic function [38], similar conclusions could be drawn from the neutrality test: significantly more negative DF values than expected under neutrality were observed for the capsid data set (P = 0.019) and the complete polymerase data set (P = 0.018), whereas the matched polymerase lacked power in rejecting neutrality (P = 0.029).
The demographic inference using the BSP model is summarized in Figures 1A and 1B, which essentially plot Neτ as a function of time. Ne τ can be considered a measure of relative genetic diversity that, in turn, reflects the number of effective infections established by the virus (see also the Materials and Methods section). Uncertainty in the estimated parameters was evaluated using 95% Highest Probability Density (HPD) intervals. The Maximum Clade Credibility (MCC) trees from the same Bayesian analyses (Figures 1C, D) summarize the NoV evolutionary histories, and the stepwise emergence of the subsequent variants on a time scale. For comparison, surveillance data of reported NoV outbreaks with confirmed GII.4 variant type were imposed on the BSPs.
The changing patterns of NoV genetic diversity generally revealed seasonal dynamics, albeit with markedly varying resolution among the two datasets. The BSP for the polymerase dataset (Figure 1A) showed peaks for Neτ that coincided with the epidemic peaks observed in norovirus surveillance systems in the northern hemisphere winters 2002–03, 2004–05 and 2006–07. The BSP obtained from the capsid dataset (Figure 1B) showed a pattern that was more difficult to reconcile with epidemiological observations. Values for Neτ were highest in the years 1997–1999, and the emergence of the 2002 variant, which had a strong impact in the population according to surveillance data, did not coincide with a pronounced upsurge in the BSP. Comparison of the BSPs obtained for both genes, illustrated that unraveling seasonal population dynamics with associated population bottlenecks for viruses like NoV, may require a sufficiently high sampling density. In fact, reducing the partial polymerase dataset to a similar number of sequences drastically diminished the resolution of the BSP analysis (Figure S1). In particular the 2004–05 and 2006–07 epidemics were not well reflected in the BSP derived from this sub-set of polymerase sequences that matched the capsid dataset in size, both genetically and temporally. The 2002–03 epidemic, following the replacement of the 1996 variant by the 2002 variant, wás however clearly noticeable in the matched polymerase set, whereas it was not in the capsid data set. Considering the associated MCC trees (Figures 1C, 1D), it is conceivable that following the relatively long build-up of genetic variation during the circulation of the 1996-variant, its replacement by the 2002 variant signified a massive and sudden loss of diversity; a population bottleneck. The 2002 variant split into two distinct subclusters for the polymerase dataset. These lineages arose almost immediately after the emergence of the 2002 variant, and individually coalesced to a Most Recent Common Ancestor (MRCA) shortly before their diversification. The capsid 2002 variant cluster also grouped in two sublineages, but they coalesced more gradually to their MRCA.
Comparison of the variant dynamics in the MCC trees to their respective BSPs suggested that variant replacement was not always absolute across subsequent epidemic seasons. To investigate this in more detail, we performed the coalescent analyses on partial polymerase datasets for the individual major GII.4 variants separately (Figure 2). Whereas the pattern of rapid emergence, followed by an (epidemic) peak and later peaks of diminishing size observed for the 2002, 2004 and 2006a variants were very similar, the patterns obtained for 1996 and 2006b were quite different. The 1996 variant, that was detected in the population during a relatively long period, but at low reporting frequencies after the initial epidemic (winter of 1995–1996, in the northern hemisphere) (Figure 1A/B), showed an increasing trend in the Neτ values persisting long after this first peak. The 2006b strain showed a less defined pattern, with multiple, smaller peaks.
The demographic component is part of a full Bayesian model that enables the inference of time-scaled evolutionary histories and rates of molecular evolution from temporally-spaced sequence data. Rates of nucleotide substitution and the MRCA's of the included GII.4 sequences were listed in Table 1. The substitution rates found for the less densely sampled datasets, namely the complete capsid sequences (5.33×10−3 substitutions per site per year) and the matched polymerases set (4.32×10−3 substitutions per site per year) are lower than the rate found for the large partial polymerase dataset (8.98×10−3 substitutions per site per year). The estimated MRCA for these GII.4 strains lies in the first half of the 1980s (Table 1).
Because the genealogical test rejected selective neutrality for the capsid gene, we attempted to identify the molecular determinants of this selective process through two different approaches that were not previously applied on NoV data, namely DEPS [38] and co-evolutionary analysis of amino acids [39], and complemented this with novel extensions of previously performed codon substitution model analyses [23], [40]. The partial polymerase sequences under analysis in this study are very short and have therefore not been included in these analyses.
In order to apply a codon model based on a general bivariate discrete distribution (GBDD) of dN and dS [41], we employed a small sample AIC, which suggested that six rate classes (D = 6) provided the best fit to the capsid data. The proportion of sites within these classes and corresponding dN and dS estimates are represented by Figure 3A and 3B. The model included one rate class describing positive selection (dN ( = 0.77)>dS ( = 0.00)), with an estimated 0.93% of sites occupying this class. An empirical Bayes approach identified sites 6, 9, 15, 47 and 534 (0.93%) to be under diversifying positive selection (Table 2). Three of the sites (6, 9 and 534) were confirmed by a site-by-site Fixed Effects Likelihood (FEL) analysis at p<0.05, while the remaining two (15 and 47) were borderline significant (p = 0.06 and p = 0.10). The rates of false positives for FEL analyses at p = 0.05 was approximately 0.04 and 0.08 at p = 0.1, based on dataset-matched neutral simulations, suggesting that the putatively selected sites were not due to elevated rates of false positives at given nominal significance values.
To uncover population level selection processes, FEL analysis may be more appropriately applied to internal branches (iFEL) [42]. That this approach suited our data was also suggested by our genealogical tests, which identified an excess of slightly deleterious mutations on terminal branches, indicative of within host evolution. The use of iFEL enabled us to avoid this effect and revealed 8 codons under positive selection at the population level including 6, 9, 47, 352, 372, 395, 407, 534, with p≤0.05.
Codon models are powerful tools to detect an unusually high rate of nonsynonymous replacement, which generally occurs under a scenario of diversifying selection. However selection of episodic nature, e.g. directional selection or frequency-dependent selection is more difficult to detect and involves the question of which residues are being selected for or against [38]. A directional evolution in protein sequences analysis (DEPS) of NoV capsid sequences revealed elevated substitution rates towards 4 residues: V, S, A, T. Four sites were identified to be involved in this directional evolution; amino acids 9 (with inferred amino acid substitution pattern: N→T/S→N), 294 ((V→)A→S/P→A→T ), 333 (L→M/V/L→M→V), and 395 (-→T→A) (Figure S2).
In folded proteins amino acids are not arranged linearly; many functionally interact, making their evolution dependant on that of others. Various types of interactions exist, and interacting sites are not necessarily direct neighbors in either the protein sequence or in the 3D protein structure. We used Bayesian graphical models (BGM) to detect co-evolving sites. The sites identified, are shown as a network in Figure 4, and sites for which co-evolution was detected but seemed less supported are shown in Figures S3A and S3B. Two values for the posterior probabilities are given, obtained from the analyses allowing for either one or two co-dependencies. Sites 231 and 209, and 238 and 504, which co-evolved as two coupled sets (Figures S3A and S3B), were not involved in recent variant transitions. Therefore we conclude that they were not under selective pressure that governed variant replacement dynamics.
All sites for which molecular adaptation was detected are listed presented in Table 3. We marked relevant sites located on top of the capsid dimer in Figure 5.
Our study identified codon 6 to be under positive, diversifying selection and codon 9 to be under diversifying as well as directional selection; others performing dN/dS analyses detected positive selection at these sites as well [23], [40]. The variation in codon 6 is due to AAT (Asn) to AGT (Ser) changes. The signal for codon 9 is solely attributable to substitutions at the second codon position of this site; AAC (Asn), ACC (Thr) and AGC (Ser). The N-terminal region of ORF2 was otherwise highly conserved. Because contrary to the other amino acids that were identified to be under selective pressure, the part of the protein encoded by these two amino acids is located on the inside of the virus capsid structure, and not surface exposed, we investigated the potential RNA secondary structure encoded by this region. In silico replacement of nucleotides at position 17 (codon 6) did not lead to secondary structure changes (not shown). Secondary structure predictions of the RNA encoding the 5′end of ORF2 were performed with all four possible nucleotides modeled at position 26 (Figure 6). The 4 nucleotides upstream of the ATG, generally thought to form the boundary of the subgenomic RNA [43] were included in the predictions. The presence of A, C, or G generated similar structures, when however a T (U) was modeled, extra pairing possibilities arose, lengthening the stem of the first stem-loop structure, and thus shortening the stretch of free nucleotides, available for ribosome binding, from 11 to 9.
Following the emergence of the genetically and antigenically novel GII.4 2002 variant in 2002, a sharp rise in the number of reported NoV outbreaks was recorded in several surveillance systems throughout the world [15], [18], [21]. Earlier, in 1995–96, a similar rise had been reported resulting from the emergence of another cluster of GII.4 - the 1996 variant [44]. The 1996 and 2002 variants have by now been succeeded by successive, highly prevalent GII.4 variants. Some uncertainty remained, however, as to whether the increase in the number of reported outbreaks was the result of a true rise in the numbers of infections, or of improved detection techniques combined with better reporting due to heightened awareness. In order to resolve this ambiguity in NoV epidemic history we applied Bayesian statistical phylodynamic techniques (Bayesian Skyline Plots, or BSPs) to alignments of a large number of GII.4 sequences. We showed that these techniques confirm the epidemic behavior of the GII.4 variants during the past decade, and that the increase in the number of NoV infections suggested from global surveillance data coincides with the rise of the GII.4 variant. Additionally we applied in silico techniques to identify sites under various types of selective pressure and to unravel epistatic interactions in the capsid protein.
The first rise in the BSP estimates of Neτ (a measure of effective population size) in the polymerase dataset was seen just before January 1996, around the time the first global GII.4 epidemic was noted [45] (Figure 1). Interestingly, rather than appearing as a defined epidemic peak lasting one winter season, Neτ increased continuously through January 2000. Surveillance data from around the world identified relatively few GII.4 outbreaks among NoV positive outbreaks in the period between 1997 and spring 2002 [12], [22]; instead, a relatively high diversity of other, non-GII.4 genotypes was detected in this period. The high Neτ observed for this period is congruent with the long branch-lengths in the MCC tree during this period. Given the prolonged circulation time of the 1996 variant compared to the later variants, it seems likely that the BSP in this instance is a better representative of the relatively high genetic diversity built up in an extended period of co-circulation of the 1996 variant with other genotypes, rather than of the number of infections with this particular variant. The three most recent GII.4 epidemics caused by emerging GII.4 variants were clearly visible in the polymerase BSP. The 2002 epidemic was preceded by a sharp decline in Neτ, indicative of a purifying selection event against the previously dominating variant, in which the diversity that had been gradually built up by the long circulation of the 1996 variant collapsed. Next, coinciding with the off-seasonal peak observed in public health surveillance systems in the northern hemisphere, Neτ in the BSP rose sharply. After a brief and insignificant decline another peak occurred in the winter season of 2002–03. During the epidemic caused by the 2004 variant Neτ peaked slightly less high and less long than during the 2002 epidemic peak. During the 2006–07 epidemic, with two distinct GII.4 variants circulating, Neτ was the highest (Figure 2), and after this winter Neτ values stayed high, corresponding with continued circulation of the 2006b variant, as also observed in surveillance.
Altogether, from the first half of the 1990s to present time, two major changes were observed. First, epidemic waves have arisen that could not be detected in earlier times, and second, the number of infections has gone up. The baseline before 1996 corresponds to the trough levels in between contemporary epidemics. Hence, the increased amount of GII.4 outbreaks observed in outbreak surveillance seems to reflect an actual increase of GII.4 infections in the population.
These conclusions are undoubtedly impacted by the comparatively few available sequences from earlier years, which are, unfortunately, not widely available, and as shown by the recent publication by Bok et al., not easily attainable [40]. A search in archival stool samples revealed that during the years 1974–1981 and 1987–1991 GII.4 was not the most prevalent NoV genotype in hospitalized children with gastroenteritis, but GII.3 was. Our sampling, albeit at lower frequency during this period, would probably have detected possible unidentified epidemic surges had they occurred. Altogether, although surveillance data and demographic estimates are very different types of information, their dynamics match remarkably well. The surveillance data is presented on a linear scale and reflects the reported outbreaks of NoV-gastroenteritis, which is probably an incomplete description of NoV circulation, as many cases of NoV illness remain unreported. The BSPs are presented on a log-scale, which makes for less sharp peaks than the peaks observed in the surveillance data graphs.
The analysis of each major GII.4 variant separately (Figure 2) indicated that Neτ values, a measure for relative genetic diversity, and a proxy for the number of effective infections, of each variant reached approximately similar proportions. The 2002, 2004 and 2006a variants each had one epidemic season, but the turnover was relatively slow compared to e.g. influenza [46], resulting in repeated seasons (of decreasing magnitude) of illness caused by the same variant. For example, one main peak for Neτ was observed for the 2004 variant during the 2004–05 winter, but a smaller peak followed during the 2005–06 winter. This slow turnover may very well have been caused by the fact that only incomplete immunity is mounted against NoV after infection or alternatively that the pool of susceptibles is not depleted within one season. Different BSPs were obtained for GII.4 variants 1996 and 2006b, that persisted longer than one season. Interestingly Neτ for the 1996 variant increased just before the emergence of the 2002 variant. The 2006b variant showed no defined epidemic peak, but a series of smaller peaks that did not coincide with annual winter-peaks. Two sublineages of the 2006b variant, distinguishable by up to 5 amino acid differences in the full capsid sequence [47] circulated in the population. Of these only S368G has been recognized as a polymorphism relevant for antigenic properties. Interestingly, while only two 2006b strains with this mutation are present in our capsid dataset, later full capsid sequencing revealed more strains of this sublineage, and the polymerase dataset included more sequences from this cluster. Thus, the 2006b variant may have persisted in the population by changing its antigenic properties.
While Bayesian coalescent analyses of the large partial polymerase dataset reflected seasonal epidemic dynamics, the analysis of the longer capsid sequences offered little detail about the phylodynamic patterns of NoV. To investigate the nature of this difference, we performed an analysis of polymerase sequences matching the capsid sequences genotypically as well as temporally, which showed that a lack of phylodynamic detail can generally be attributed to a lower sampling density (supplemental materials, Figure S1). This may not be so surprising as it was previously thought that coalescent analyses would not be so effective at capturing cyclical population dynamics [48]. Only recently, a comprehensive analysis of a large H3N2 influenza virus dataset (1302 taxa) was able to uncover seasonal dynamics from genetic data [46]. Interestingly, this data set was similar in size compared to the large dataset of partial NoV polymerases (1383 taxa) presented here, although in the influenza study full gene segments were analyzed. The population bottlenecks in the NoV GII.4 population history are to a large extent comparable to those seen for influenza and constitute repetitive large scale losses of genetic diversity. We note that viruses with seasonal dynamics do not necessarily have to exhibit such dynamics in genetic diversity. Short infections with strong cross-immunity, as seen for measles virus, allow many strains to co-circulate with frequencies contingent on neutral epidemiological processes [24]. For such viruses, seasonal epidemics may arise from repeated exhaustion of susceptible host populations [49]. Therefore, our phylodynamic analysis predicts that there may only be partial subsequent cross-immunity against GII.4 variants. It is important to note that sampling size impacts the resolution of phylodynamic inference, but the actual sampling scheme does not dictate a pattern of fluctuating population size. Rambaut et al. [46] performed simulations using various demographic scenarios but with a sampling scheme used to obtain influenza genetic data from seasonal epidemics. In all cases, the simulated demographic history was accurately recovered. A sparse sampling prior to 2000 also makes it difficult to unequivocally conclude an increase in GII.4 infections. However, we note that the value for Neτ before the first documented epidemic resulting of the emergence of a new genetic variant (the 1996-variant) in the BSP is lower than the estimates between subsequent epidemic peaks. This suggests an increase in the number of NoV GII.4 infections, which is further reinforced by a recent study of archival stool samples from the Children's Hospital, Washington, DC (1974 to 1991) (Bok et al., 2009). Although this study identified GII.4 strains in the early seventies, this was not the predominant genotype before 1991 (Bok et al., 2009). An increase of the GII.4 variant therefore seems to provide a plausible explanation for the coincident increase in the number of norovirus infections.
The estimated substitution rates (9×10−3substitutions per site per year for the partial polymerase sequences and 5.3×10−3 substitutions per site per year for the capsid sequences) corresponded to the values recently reported for NoV GII.4 [40] and were well within the range of what is commonly found for RNA viruses, e.g. between 3.5×10−3 and 8.5×10−4 for HMPV complete genomes [50] and for influenza A virus the highest rate, reported for the HA gene, was 5.72×10−3 substitutions per site per year [46]. A lower rate was observed for the polymerase subset matched to the capsid data set. Since the TMRCA estimates were consistent between these two polymerase data sets, the lower rate may be explained by differences in rate variation among sites, in particular for the proportion of invariant sites, which is sensitive to the number of taxa in the data set [51]. Dating the MRCA for these strains back to the early 1980's does not mean that the GII.4 lineage arose only then, but rather suggests that the strains that circulated during the past two decades share a common ancestor at that time. It seems not unlikely that the GII.4 lineage was less diverse before the 1980's, not comprising different variants as during the past decades. Alternatively, if multiple GII.4 variants did exist before the MRCA of the current GII.4 variants, the occurrence of a population bottleneck may have left progeny virus of only one variant. The strains detected in the 1970s reported recently seem to confirm that multiple GII.4 variants existed before the Camberwell cluster arose [40]. Analyzing data of multiple NoV genotypes will provide a more detailed insight into the branching times of these different genotype clusters, and also in this case, including older sequences will be more elucidating.
Ideally we would have used full genome sequences. However, for (GII.4) NoV these are only sparsely available. Instead, we showed that Bayesian coalescent demographic analyses of a large dataset containing very short sequences offered important and reliable insights into GII.4 variant dynamics. For the GII.4 NoV datasets presented here, the densely sampled but short polymerase sequences provided data that better defined the epidemic history than fewer longer capsid sequences. Perhaps additional to the limited size of the capsid dataset, strong selective pressure on the capsid protein confounded analysis of the capsid gene. We were also aware of the possibility that our capsid sequences dataset may have been biased in sampling. Sequencing of full capsid genes is not standard practice; the viruses of which sequences were available have all previously been selected by various researchers as ‘interesting enough to sequence’. Thus, relatively many sequences of strains belonging to the 1996 variant are present in the dataset, especially compared to strains belonging to the younger variants, 2004, 2006a and 2006b.
Our genealogical test clearly rejected neutral evolution for the NoV capsid dataset. Although P-values were somewhat higher using a Bayesian skyline plot model in the posterior predictive simulation compared to more restricted demographic functions, we arrived at the same conclusions for all analyses. This demonstrates that the neutrality test is not overly sensitive to the demographic detail in the analysis. Nevertheless, through the advances made here, we demonstrate that this test can now be performed under any complex demographic scenario, a generalization that may further promote its use. To investigate the selective forces in more detail, we fitted different evolutionary models to identify sites in the capsid under selective pressure. We examined an extension of previously performed dN/dS analyses [23], [40] to detect positively selected sites involved in population level selection, avoiding the effect of within host evolution. Eight positively selected codons were identified at p<0.05 with the iFEL approach. Of these sites, four (352, 372, 395 and 407) are located in the protruding regions of the protein. These sites have also been identified previously as ‘informative’ sites (at least two shared an identical amino acid mutation in the alignment) [20]. Amino acid 395, that was also detected as directionally evolving, is located in a surface exposed loop of the capsid protein, that is part of a variable site of carbohydrate interaction (amino acids 393-394-395) that has been identified by a number of studies as a locus for ligand binding and specificity [23], [52], [53]. Codon 394, located in this same domain, is part of an intricate network of co-evolving positions, also containing codons 297, 368 and 372. Amino acid 297 is part of a site identified by Allen et al, (296-297-298) [52], predicted as another of two host ligand binding sites. This particular site was not identified by Cao et al. [53] who performed co-crystallization assays with P-particle dimers and A and B-trisaccharides, the host ligands of NoV. It is structurally close to amino acids 368 and 372, on top of the protein, flanking the ligand binding pockets, and to 294, under directional evolution, which is located on the outside of the 296–298 loop, relative to the binding pocket.
Lindesmith et al. also identified sites under positive selection, using fewer sequences. They used three different methods, single likelihood ancestor counting (SLAC), fixed effects likelihood (FEL) and random effects likelihood (REL), under the Tamura-Nei model of evolution. More codons under positive selection were identified in this study, but less strict nominal significance values were used. Bok et al. [40] applied SLAC analyses for detection of positive selection and found six amino acids under positive selection. We chose to identify sites under selective pressure for internal branches in a phylogeny using iFEL because external branches are prone to deleterious mutational load. Such mutations are expected to be young and more likely fall on the external branches of a population-level phylogeny [54], where they can confound the identification of positively selected sites. To avoid this adverse effect we focused on internal branches only, on which advantageous mutations are more likely to fall.
Codon 333 was previously identified as an informative site [20] and our analyses found it to be under directional selection. It is located in the hydrophobic part of the P-dimer interface, just below the carbohydrate binding site described by Tan et al. [55], facing its counterpart in the other protomer of the same dimer (distance 4 Å) [53]. Changes in this amino acid are not likely to be involved in antigenic change but more likely structural compensation for other mutations.
Codons 6 and especially 9 have consistently been identified as sites under strong selective pressure, and involved in defining the distinction between variants, not only in this study but in others as well [23], [56]. Given their positions at the N-terminus of the protein, inside the shell, they seemed unlikely to have been under selective pressure through antibody recognition. This notion led us to investigate the RNA encoding the 5′end of ORF2. The nucleotide sequence upstream of codon 9 (nucleotide 26) is strongly conserved among all NoV genotypes, and a highly similar sequence is found at the 5′end of ORF1 in NoVs. Mutations have rarely been detected here, apart from at nucleotide 26, and at nucleotide 17 (codon 6). We propose that secondary structure predictions of these RNA regions provide an explanation for this pattern (Figure 5). Highly conserved stem-loop structures create the circumstances necessary for translation initiation of ORF1 and ORF2. Nucleotides A, C and G at position 26 all result in almost identical structures, in which 7 nucleotides from the start of ORF2, or 11 when including the 4 nucleotides upstream from the AUG, are left free. When theoretically a U/T is inserted here the predicted structure changes, resulting in a diminished length of the free nucleotide strand of 5 nucleotides. We are unaware of sequences with nucleotide T at position 26 present either in our or the public databases (data not shown), leading us to believe that a length of at least 7 unpaired nucleotides counting from the first AUG is necessary for efficient RNA translation from the subgenomic RNA. Thus, while normal capsids could be formed from strains with (silent) mutations in this area, the replication process of the virus may be disrupted by altering the secondary RNA structures. This theory is further supported by the observation that no other synonymous or non-synonymous mutations are found in the 9th codon, e.g. at the third nucleotide position, nor at other nucleotides in this genomic area, apart from the previously mentioned nucleotide 17, in an RNA loop. Tentative analyses of other NoV genotypes demonstrated that the 5′ region is equally conserved within the different genotypes and yielded similar secondary RNA structures, which allow for point mutations in the loops of the structures (data not shown). To further substantiate this hypothesis site-directed mutagenesis studies are required, which go beyond the scope of this study.
All of the substitutions described above were associated with at least one variant transition; they appear at branches that give rise to new variants (Figure 5 and Figures S2, S3A and S3B). This indicates that these mutations include the molecular determinants of cluster replacement. Previously identified antigenic sites [23], [52] did not enable distinction between all the different GII.4 variants (e.g. considering amino acids 296–298 and 393–395, the 2006b and 2007 variants that are clearly phylogenetically distinct, share identical amino acids).
When amino acids 6, 9, 294, 333, 352, 368, 372, 407 and 534, identified here as under positive, directional or co-evolutionary pressure, are added to these six amino acids, all currently identified GII.4 variants (excluding the Bristol and Camberwell strains, that circulated before the 1996 variant) are separated by at least two amino acid differences (Figure S4). Thus, specific analysis of these sites will aid early recognition of novel variants in the future. The two most recent distinct GII.4 variants, that have both been detected throughout the world in both 2008 and 2009, albeit at low prevalence, the 2007- and 2008-variants, are identical to the still dominant 2006b variant in amino acids 296–298, that were identified by Allen et al, and 2007 is also identical in site 393–395, whereas the 2008 variant has two substitutions on those sites (2006b: STT, 2008: D/NTA). Thus, considering the sites listed above, the 2008 variant would have the best chance of becoming the next dominant strain, whereas, when considering the full capsid sequence, the 2008 variant is more similar to the 2006b variant than is the 2007 variant.
Using Bayesian phylodynamic techniques we showed that since 2002 the number of GII.4 infections has experienced expansion dynamics. Additionally we further substantiated the evidence for signature sites for variant transition, which may aid in the early recognition of potential new epidemic variants, although we stress that examining pre-defined amino acids does not enable certain identification of GII.4 variants, for which full capsid sequences should be determined. We showed that it is important to select the genomic region to analyze by phylodynamic, coalescent methods with care, and our different datasets illustrated that for the phylodynamic analysis of pathogens undergoing repeated selective bottlenecks a considerable sampling density through time is required.
We compiled two different NoV GII.4 datasets. First, partial polymerase gene sequences with known detection month and year were collected. These sequences, encoding a short genomic region commonly referred to as Region A, have been collected for genotyping purposes, as an essential part of the ongoing surveillance practice in institutions around the world [22], [33]. This dataset includes sequences from participants of the Foodborne Viruses in Europe (www.fbve.nl) and of the Noronet (www.noronet.nl) networks, the contributing institutions of which are listed in the acknowledgements. Sequences of sufficient length (i.e. covering at least the final region) were included, generating a dataset of 1383 taxa, 247 nt in length. Strains originated from systematic surveillance collections, and form the best representative reflection of the circulating strains currently available.
Second, complete capsid sequences of GII.4 NoV strains with known sampling date were collected. This resulted in a dataset of 194 taxa, 1623 nt long. To allow comparison of results from capsid based versus polymerase based analysis, a set of 172 partial polymerase gene sequences matching the sequences in the capsids dataset (identical variant typing and similar detection dates) were selected from the total polymerase dataset. The 2003Asia variant was excluded from this mirror-dataset, since it was identified as recombinant and ORF1 does not belong to the GII.4 genotype.
Details on the nature of the strains comprised by the two generated datasets are provided in the Supporting Materials Table S1. The distribution of the sampling dates of the included sequences is depicted in Figure S5.
For recombination analyses, a set comprising 20 sequences, two of each GII.4 variant, spanning the genome region between Region A, in ORF1, and the complete capsid sequence was collected.
Sequences were aligned using the Clustal W algorithm implemented in Bioedit (version 7.0.9.0) and edited where necessary. Sequence alignments can be obtained from the authors on request.
Recombination within the genomic area under study invalidates the use of phylogenetic approaches. Therefore, we checked for possible recombination signal by analyzing 20 sequences (two for each variant) spanning Region A through the complete capsid protein (2404 nt). Different evolutionary histories across this genome region were inferred using the genetic algorithm for recombination detection (GARD) [57] and specific recombinants were identified using a modified VisRD algorithm [58] and using RDP3 [59]. In addition, we used the Phi test, shown to perform well under strong population growth and to be able to distinguish recurrent mutations from recombination events, to identify recombination signal in the NoV alignments [36].
Evolutionary dynamics were estimated using a Bayesian Markov chain Monte Carlo (MCMC) approach implemented in BEAST (BEAST version 1.4.7 [60]). BEAST MCMC analysis estimates marginal posterior distributions for every parameter in a full probabilistic model comprising the timed evolutionary history, based on the incorporation of sampling times in a molecular clock model, the substitution process and demographic history. We used the GTR+I+Γ4 model of substitution and the uncorrelated lognormal relaxed clock model to accommodate variation in substitution rates among different branches [61].
To test selective neutrality of GII.4 molecular evolution, we adopted the genealogical framework presented by Drummond and Suchard [37]. This involves the full model-based Bayesian analysis to obtain a posterior distribution of trees, genealogical summary statistics, and posterior predictive simulation to detect departures from the neutral expectations for these statistics. We employed the genealogical Fu and Li statistic (DF), which compares the length of terminal branches to the total length of the coalescent genealogy. Strongly negative values for this statistic indicate terminal branch lengths being larger than expected, which reflects an excess of slightly deleterious mutations on these branches. This statistic has proven to be most sensitive in uncovering non-neutral evolution, and has for example rejected neutrality for human IAV hemagglutinin genes [37]. Posterior predictive simulation is performed according to the same demographic model as used in to obtain the posterior tree distribution. To evaluate the impact of large demographic trends, we compared analyses using both constant and exponential growth population size priors. To further investigate the impact of demographic detail on the neutrality test we extended the simulation procedure to highly parametric demographic models, including piecewise constant demographic functions that define a Bayesian skyline plot model. We validated the simulation procedure by comparing the reconstructed Bayesian skyline plot from the trees generated by posterior predictive simulation with the Bayesian skyline plot inferred from the sequence data, which yielded consistent results.
To reconstruct the NoV GII.4 demographic history in more detail, we employed the Bayesian skyline plot (BSP) model, which generates piecewise constant population size trajectories [62]. In this coalescent setting, demography is measured as the product of effective population size (Ne) and generation time (τ), Ne τ, through time. To obtain a detailed measurement of NoV GII.4 diversity through time, given the large dataset, we specified 40 groups in the piecewise constant population size function. All chains were run sufficiently long to achieve stationarity after burn-in, as checked using TRACER (http://tree.bio.ed.ac.uk/software/tracer/).
Additionally, the polymerase dataset was split up into separate subsets, each comprising all available sequences from a major GII.4 variant, and these were analyzed individually using the same model and settings as was used for the whole polymerase dataset.
To examine how different sampling densities through time can impact our demographic estimates, we performed an additional analysis of a subset of the polymerase data, containing sequences selected to best mirror the sequences in the capsid dataset, using the same specifications as described above.
Secondary structure predictions of the 5′end of the ORF2 encoding RNA were generated using the web based RNA Fold Webserver (http://rna.tbi.univie.ac.at/cgi-bin/RNAfold.cgi) [69].
|
10.1371/journal.pntd.0005662 | Incidence and mortality due to snakebite in the Americas | Better knowledge of the epidemiological characteristics of snakebites could help to take measures to improve their management. The incidence and mortality of snakebites in the Americas are most often estimated from medical and scientific literature, which generally lack precision and representativeness.
Authors used the notifications of snakebites treated in health centers collected by the Ministries of Health of the American countries to estimate their incidence and mortality. Data were obtained from official reports available on-line at government sites, including those of the Ministry of Health in each country and was sustained by recent literature obtained from PubMed. The average annual incidence is about 57,500 snake bites (6.2 per 100,000 population) and mortality is close to 370 deaths (0.04 per 100,000 population), that is, between one third and half of the previous estimates. The incidence of snakebites is influenced by the abundance of snakes, which is related to (i) climate and altitude, (ii) specific preferences of the snake for environments suitable for their development, and (iii) human population density. Recent literature allowed to notice that the severity of the bites depends mainly on (i) the snake responsible for the bite (species and size) and (ii) accessibility of health care, including availability of antivenoms.
The main limitation of this study could be the reliability and accuracy of the notifications by national health services. However, the data seemed consistent considering the similarity of the incidences on each side of national boundaries while the sources are distinct. However, snakebite incidence could be underestimated due to the use of traditional medicine by the patients who escaped the reporting of cases. However, gathered data corresponded to the actual use of the health facilities, and therefore to the actual demand for antivenoms, which should make it possible to improve their management.
| A better knowledge of snakebites incidence and mortality might improve their management. However, they are difficult to estimate, in particular because most of them are based on extrapolations from scientific and medical publications that are not representative of the epidemiological situation. This study, based on data available on-line at government sites in the Americas—reflecting notifications from health services—sustained by recent publications to provide useful information on snakebites treated in health centers of American countries. On average, nearly 60,000 snake bites are managed every year in health facilities in the Americas and approximately 370 deaths are reported officially. The development of snake populations results from environmental conditions favorable to their feeding and camouflage. Moreover, the activities of human—notably agricultural—explain encounters with the snakes. The literature underlines that the severity of the envenomation depend on the species responsible for the bite and quality of the management of the patient. Without excluding an underestimation of snakebite incidence, due to the frequent use of traditional medicine, this study should enable health authorities to better analyze the epidemiological situation of snakebites, including their frequency, distribution and severity, in order to improve the management of the envenomation.
| Snakebite is an important public health issue in the Americas, particularly in inter-tropical America [1; 2]. A better understanding of the epidemiological burden of snakebite, i.e. incidence, geographical distribution, population at risk, bite circumstances and severity, would improve their management [3] and should be used to urge World Health Organization (WHO) to include definitely snakebites in the list of neglected tropical diseases (NTD) and to convince international agencies and foundations of funding. It would also help the antivenom manufacturers to produce the necessary quantity, and the Health Authorities to supply the health centers according to the declared incidence and the geographical distribution of the envenomations.
However, most data available for the Americas are fragmentary and poorly representative, mainly because they come from the literature, making them incomplete and biased. This is particularly true in Central and South American countries. In recent years, international workshop dedicated to the improvement of antivenoms pointed out that “renewed efforts were required on national and regional basis to improve the epidemiological surveillance system in order to gather a more precise picture of the impact of this health problem” [4] and recommended “to improve the information systems on the epidemiology of snakebite envenomations in the region, that is essential for the design of effective distribution policies and training programs” [5]. As a consequence, most Latin American countries introduced mandatory notification of snakebites during the 2000s.
The bites by opistoglyphic (rear fanged) snakes and those of the families lacking fangs delivering venom (Boidae, Aniilidae in particular) being weakly toxic [6], represent a low demand for health services, although the incidence is far from trivial [7]. The snakes belonging to the Scolecophidia suborder (Typhlopidae and Leptotyphlopidae) are definitely non-toxic and unable to bite. As a consequence, two snake families share responsibility for snake envenomations in the Americas: the Viperidae (including half a dozen genera, the most frequent being Crotalus, Bothrops and Agkistrodon) and the Elapidae of which Micrurus is the main genus [8]. The bites of the latter represent less than 1% of the envenomations [9–13].
The symptoms caused by viper bite are mainly hemorrhagic and cytotoxic, the latter sometimes resulting in limb amputation or permanent disability [14; 15]. Some species of Crotalus may also produce neurotoxic symptoms similar to envenomation by Elapidae [16], and sometimes associated with acute renal failure [17]. Unlike the neurotoxins of rattlesnake venoms that act on presynaptic receptors (β-neurotoxins), the α-neurotoxins of Elapidae venoms bind to postsynaptic cholinergic receptors [13]. In both cases, paralysis of the cranial nerves can occur, inducing in some cases a potentially fatal respiratory arrest in the absence of specific (antivenom) and/or symptomatic treatment (artificial ventilation).
The aim of this work was to assess the epidemiological burden of snakebite, including the incidence, mortality, population at risk and main explanatory characteristics of their frequency and severity: season, environment, altitude, density of human population, management, etc., in order to provide recent and useful data to improve the management of snakebites in the Americas.
A bibliographic search was performed by querying MedLine (PubMed last access 06/11/2016) using the keywords "America AND snake * AND [envenom * OR antiven *]". From a total of 4,514 references, 187 concerned the epidemiology and/or management of snakebites in the Americas.
Furthermore, websites regarding i) the epidemiology of snakebites (using the words “health surveillance”, “surveillance bulletin”, “epidemiology surveillance”, “snakebite envenomation”, “snakebite death”), ii) population demography (using the words “population demography”) and iii) administrative and environmental geography (using the word “map”) were identified using the Google search engine for each of the countries of America and using the official language of each country (English, Spanish, Portuguese, French and Dutch). Access to these websites was made between September 2010 and December 2016. The list of the websites and the last access date to each are mentioned in Table 1. However, a few websites were closed during this period and sometimes replaced by new ones, the use of which was often restricted by a password.
All the data were transferred and analyzed using Excel software. The trend curves and R2, the coefficient of determination that is the square of the coefficient of correlation indicates the extent to which the dependent variable is predictable, were calculated through Excel. The comparisons were made using parametric tests (t-test, χ2 and Pearson correlation) or non-parametric (Mann-Whitney), depending on the distribution of studied variables and number of cases/groups. The significance level was equal to 0.05 and the means were expressed using a 95% IC. Statistical analyzes were performed using the BiostatTGV online software (http://marne.u707.jussieu.fr/biostatgv/).
Topographic, physical and political maps were taken from the World Atlas of Wikimedia (https://commons.wikimedia.org/wiki/Atlas_of_the_world) and drawn on the basis of the data obtained in this study.
The average incidence is about 57,500 snakebites a year (6.34 per 100,000 population), resulting in almost 370 deaths (0.037 per 100,000 population), with a case fatality rate below 0.6% (Table 2). However, there are wide variations across countries and within each of them.
The data are detailed for each country according to the websites mentioned in Table 1, eventually temperate by recent epidemiological or clinical publications.
Information is available online since 2007.
According to Ministry of Health records, there was an average of 700 envenomations (1.63 per 100,000 inhabitants) and 5 deaths per year (0.012 per 100,000 inhabitants) during the 2007–2014 period. There was a steady decrease in annual incidence (Fig 1A). The incidence showed a decreasing gradient from north to south (Fig 2), which corresponded to the climatic trend between the Chaco province which climate is subtropical, and the Patagonia province more rigorous on the one hand, and the Andean climate of eastern Argentina, on the other hand. Two provinces presented a higher incidence than the others: Santiago del Estero in the north with a low population density (7 inhabitants per km2) and Misiones in the north-east with a higher density (35 inhabitants per km2) but predominantly agricultural. The seasonal distribution of envenomation showed a summer incidence five to six times higher than the winter one (Fig 1B).
These results corroborated those by Dolab et al. [18] obtained from a questionnaire survey conducted at health facilities. These authors have shown the strong geographical heterogeneity of the incidence which can reach 150 envenimations per 100 000 inhabitants in certain places. They confirmed the low case fatality rate (0.04% according to the survey). Bothrops were responsible for 96.6% of the bites, Crotalus 2.8% and Micrurus 0.6%. Population at risk consisted of young men bitten during agricultural activities. Most envenomations (90%) were treated within the first four hours by an antivenom.
No information on snakebites has been obtained for Belize.
However, on the basis of existing data from neighboring countries showing similar environments, the annual number of bites can be estimated at 35 (10 per 100,000 population) and deaths at 1 every 2 to 4 years (0,1 per 100 000 inhabitants).
The information is available online from 1996 but the notification was interrupted between 2001 and 2009. It returned to availability from 2010. The presentation of the data has been standardized, in particular as regards the classification of age groups. By 2015, the information has been supplemented by the addition of the gender of the patients. However, mortality from envenomation is still not provided.
In their study on snakebites in Bolivia, Chippaux and Postigo [19] reported a national incidence of 8 bites per 100,000 population per year with a case fatality rate of 0.42 per 100,000 population. They extrapolated mortality from household survey data, which lacks precision and reliability but was consistent with the mortality observed in neighboring countries. Updating data available up to 2015 confirmed the impact over this period with over 900 annual bites (9.1 per 100,000 population). The number of deaths is still not reported but has been estimated at around 40 per year [19]. The main results of these authors, in particular the geographical distribution and the distribution by age group, were confirmed by the notifications during the years 2010–2015.
The annual incidence increased significantly between 2010 and 2015 (Fig 3A). The distribution of the specific incidence showed a steady growth according to age (Fig 3B). The sex ratio (M/F) was 1.81. The geographical distribution was very heterogeneous. The incidence is very low (less than 1 snakebites per 100,000 inhabitants) in the high mountain region, notably the Altiplano (departments of Potossi, Oruro and most of that of La Paz where altitude exceeds 3,500 m asl). The lowland or steppe departments, such as the Chaco region (Departments of Tajira, Santa Cruz, Chuquisaca, Cochabamba and Beni) have an incidence of between 5 and 50 per 100,000 inhabitants. Finally, the incidence exceeds 50 bites per 100,000 inhabitants in the Department of Pando in the Bolivian Amazon [19]. The seasonal distribution (Fig 3C) showed a clear difference between the Chaco province (medium-altitude steppe) where the incidence is highest in the dry season, and Amazonia (low-lying primary forest) where bites occur mainly during the season rains. Finally, if the relationship between population density and incidence was not shown, Chippaux and Postigo [19] observed a significant inverse correlation (P < 1.6·10−4) between incidence and altitude.
The notification of snakebites is performed for a long time in Brazil but the results are online only since 2001. According to Chippaux [20] from the data reported by the health facilities and available online on the site of SINAN which is the main Database on causes of morbidity and mortality available online since 2001 [21], the average number of snakebites was about 27,200 per year (15 per 100,000 population) with more than 115 deaths (0.06 100,000 inhabitants) during the period 2001–2012.
The geographical distribution showed a clear predominance in northern Brazil, especially in the Amazon (Fig 4). The seasonal distribution of bites was more pronounced in the summer, particularly in the southern regions (Fig 5). The incidence by age group varied greatly from region to region. It was higher among young people in the Amazon and in people over the age of 40 on the inland plateau [20]. Bothrops species were responsible for most of the bites everywhere in Brazil. Bites by Crotalus durissus are more frequent in the eastern and central savannas. The bites by Lachesis sp. are mostly observed in the Amazonian region. Those by Micrurus sp. are rare. Finally, there was a strong inverse correlation between incidence and population density [20].
The population at risk was made up of male farmers. Risk factors were more or less directly related to the agriculture and rural housing of the victims [22; 23]. Bochner and Struchiner [24] showed that these characteristics have been constant since the first epidemiological studies carried out by Vital Brazil in the early 20th century.
Incidence and mortality increased discreetly and seemed to follow demographic trends [20].
Snakebites appeared to be very rare in Canada because of a climate unfavorable to the establishment of snake populations, and a highly mechanized agricultural activity. The presence of Sistrurus catenatus is attested in southern Ontario, the most populated region of the State, and Crotalus oreganus occurs in British Columbia. Crotalus horridus disappeared from Oregon since 1941 and from Quebec more recently [25]. Rumors of his return to southeastern Canada, including Quebec, have not been validated by the Recovery Commission for the Ontario Rattlesnake [26].
According to Dubinsky [27], there were about sixty snakebites reported in Ontario each year. There would have been 2 deaths between 1900 and 1960 [28] and since the 60s none has been reported in the literature. There were no figures for British Columbia and snakebites are considered very rare.
In total, it can be assumed that snakebites are fewer than 100 annually and no death was reported in Canada since a long time. Snakebites are distributed in the two southern states (Ontario and British Columbia) close to the border of United States of America where rattlesnakes are still encountered. However, some snakebites recorded could be illegitimate bites inflicted when manipulating a snake in the field or in captivity.
There are neither Bothrops, nor Crotalus, nor Micrurus in Chile. Snakebites by opistoglyphic snakes were reported but not considered as public health issue [29].
Notifications have been available online since 2009. The number of reported snakebites was approximately 4,150 per year (8.5 per 100,000 population), resulting in about 35 deaths (0.08 per 100,000 population) between 2009 and 2014.
The incidence of snakebites increased significantly during the period (Fig 6A) without clear explanation. Maybe, the case report system–or political situation–improved enough to obtain more reliable data. The geographical distribution was heterogeneous. Incidence was relatively high in the whole of the country, especially in the Amazonian departments in the south (Fig 7) and much lower in central Colombia, both mountainous and urban. There was no correlation between population density and incidence. The seasonal distribution was constant throughout the year (Fig 6B).
The notification has been available online since 2005 with some shortcomings—or delays in data capture—after 2012.
From 2005 to 2012, an average of nearly 700 snakebites (15 per 100,000 inhabitants) and 7 deaths (0.15 per 100,000 inhabitants) were reported.
The incidence of snakebites was significantly higher in the eastern provinces. In the center of the country, it was lower, especially in the province of San José, which is the most densely populated and mountainous (Fig 8). Annual snakebites ranged 500–1,000 without any particular trend [30]. The sex ratio was 1.7 (M/F) and increased in adult male whereas it decreased in women over 15 years (Fig 9A). The seasonal incidence was relatively stable during the year, however, with marked variability in the rainy season from May to November when the majority of snakebites occurred (Fig 9B).
These results were in agreement with those from the literature. The highest mortality is observed in the provinces of Puntarenas in the south and Limon in the east, linked to the abundance of Bothrops asper [31; 32]. Based on the notification of cases and environmental information, Hansson et al. [33] were able to model high risk zones of bites by Bothrops asper, and to recommend a targeted supply of antivenoms.
Access to full epidemiological data for years prior to 2013 was limited [34]. From 2013, the weekly notification was available online but showed many shortcomings.
According to González-Andrade and Chippaux [34], there were nearly 1,500 snakebites (9.8 per 100,000 inhabitants) resulting in about 10 deaths (0.06 per 100,000 inhabitants) each year.
The highest incidence occurred in the Amazonian provinces (Oriente province), with 37% of the envenomations and an average annual incidence of 100 envenomations per 100,000 inhabitants (Fig 10). The majority of snakebites (58%) were in the coastal region (Costa province) with an average incidence of 12 bites per 100,000 inhabitants. In highland provinces in the center of the country (Sierra province), incidence was about 5 bites per 100,000 population (5% of the snakebites). During the rainy season, from January to April, the incidence of snake bites is twice as high as in the dry season. The incidence of snakebites is twice as high after the age of 10 and remains stable from teenagers to elderly. The very young children below 5 are ten times less involved than adults.
Although notification of snakebites has been mandatory since 2010, data are not accessible. On the other hand, they are subject to periodic reports put online. We used that of 2013 which compiled the data from 2010 to 2012.
About 300 annual snakebites (5 per 100,000 inhabitants) were irregularly distributed during the year. Based on the results of neighboring states, the annual number of deaths can be estimated at 3 (0.05 per 100,000 inhabitants). The six months of the rainy season (May to October) accounted for nearly 65% of the envenomations (Fig 11A). The population at risk was mainly composed of young men. Patients aged 10 to 30 constituted 51% of the bites, while this age group represented less than 40% of the population. In addition, the sex ratio (M/F) was 1.5. During this period, no death was reported. The geographic distribution of incidence was heterogeneous, i.e. lower on the coast and in the center of the country (Fig 12), a probable consequence of the local population density, which is the highest of the Americas (Fig 11B).
There was no recent data concerning this small French department. According to the literature, mostly from surveys dating back to the 1980s, the annual incidence of envenomation exceeded 25 cases per 100,000 inhabitants with relatively high mortality [35–37].
Data were available online since 2001 with some gaps, notably in 2005.
With almost 900 snakebites on average each year (2001–2010), the distribution of the incidence was very heterogeneous (Fig 13). Mortality was not documented. It was estimated on the basis of neighboring country mortality at about 10 deaths per year (0.06 per 100,000 population).
There was no notification of snakebites in Guyana. However, a study of cases of envenomation treated at the Georgetown Public Hospital Corporation (GPHC) in 2014 provided an estimate of the burden of envenomation for Guyana as a whole. However, data for the Amazon region, which is sparsely populated but with high snakebite risk, was highly under-estimated, partly because it was likely that few patients visit the health facilities and, on the other hand because the evacuation possibilities on Georgetown are almost nonexistent.
According to Bux [38], there would be more than 200 snakebites each year in Guyana, an incidence greater than 25 bites per 100,000 inhabitants. The number of deaths was not specified, but Langston [39] mentioned a high number of deaths. The press reported 3 deaths in Georgetown between 2011 and 2014, which was probably underestimated since it did not take into account deaths in provincial health facilities.
More than 80 snakebites were treated each year at the Georgetown Reference Hospital during the 2010–2012 period. However, the geographical distribution was biased due to the lack of reliable data for the South (Amazonian region) of the country (Fig 14). The age-specific incidence calculated on the basis of hospital data showed a constant increase of snakebite incidence until the age of 30–40 years and then a steady decline up to 60 years.
Notification of snakebites has been mandatory since 2009 but online display was interrupted at the end of 2013.
A little more than 650 snakebites occurred annually on average (10 per 100,000 inhabitants). The number of deaths was not reported but was estimated at 7 per year (0.08 per 100,000 population) based on observations in neighboring countries.
Snakebites were mostly distributed to the north and east of the country (Fig 15), regions with the lowest altitude. The number of snakebites is relatively stable throughout the year with a slight increase in incidence during the rainy season from May to October.
Notification of snakebites was not mandatory in Martinique and records were not available online.
According to the literature [40;41;42], about 15 envenomations are treated in hospital every year (about 5 per 100,000 inhabitants). The average number of deaths was 4 every 20 years (0.05 per 100,000 population) during the 1990–2010 period. Only one venomous species (Bothrops lanceolatus) is present on the island [8]. The geographical distribution of the bites covered the whole of the island, but mainly involved small agricultural communes (Fig 16). However, no obvious link was observed between snakebite incidence and agricultural work in the two main types of plantations of Martinique (bananas and sugarcane).
Venomous animal attacks was reported since 1996 but snakebites were separated and available online only since 2003.
The annual number of bites averaged 4,000 (3.3 per 100,000 inhabitants) with steady growth between 2003 and 2015 (Fig 17A). The number of deaths was below fifty per year (0.035 per 100,000 inhabitants). As showed by Frayre-Torres et al. [43], the mortality rate decreased from 0.25 per 100,000 population in the 1970s to 0.05 during the 2000s. The lowering continued after the 2010 and is now less than 0.04 per 100 000 (Fig 17B).In addition, mortality was higher in the South than in the North of Mexico and increased significantly after the age of 40, whereas it appeared to be stable before. Case fatality rate was higher among males than females (P <0.028). The geographical distribution was relatively homogeneous (Fig 18) with a decreasing trend from the north, where the mean incidence was close to 2 per 100,000 inhabitants, towards the center (average incidence 7 per 100,000 inhabitants) and the South (incidence greater than 9 bites per 100,000 inhabitants). The sex ratio (M/F) was 1.97. The seasonal distribution showed a marked summer increase in snakebites (Fig 18).
Notification of snakebites is not available online. The epidemiological data were based on the work by Hansson et al. [44] the source of whom was the Ministry of Health.
According to these authors, there were about 650 snakebites each year (56 per 100,000 inhabitants) and 7 deaths (0.6 per 100,000 inhabitants). The geographical distribution was very heterogeneous, with a higher incidence in the south of the country, largely dependent on altitude, land use and health supply [44].
Notification of snakebites in Panama was not available online.
According to the Ministry of Health, the average annual incidence could be 1,900 snake bites (55 bites per 100,000 inhabitants). Valderrama et al. [45] mentioned about fifteen deaths per year (0.5 deaths per 100,000 inhabitants). The incidence was highest in the provinces of Darién, Coclé, Los Santos (three provinces in the center of the country) and Veraguas in the east, although in the latter the data were much underestimated. The work by Barahona de Mosca (2003, quoted by Valderrama et al. [45]) showed that people aged 20 to 44 were the most affected (44%), followed by teenagers aged 10–19 (23%), and children 0–9 (18%). In all age groups, males were most often bitten. Highest incidence occurred during the rainy season (from May to November).
Notification of snakebites has been mandatory since 2008 but was only truly functional from 2009.
Nearly 250 snakebites were reported annually (3.5 per 100,000 population) during the period 2004–2015. Snakebites decreased regularly between 2009 and 2013, and then increased dramatically in 2014 and 2015. However, the general trend of incidence is decreasing (R2 = 0.7319) suggesting that the annual variations are random and risk is reducing. The average number of deaths was 5 per year (0.08 per 100,000 inhabitants).
The seasonal incidence is relatively constant throughout the year with a slight increase during the rainy season (December to April). The incidence was higher in northern and eastern Paraguay (Fig 19).
Notification of snakebites has been available online since 2000.
On average, 2,150 snakebites occurred per year in Peru (7.2 per 100,000 population), resulting in about 10 deaths (0.043 per 100,000 population) during the years 2000–2015. The increase in incidence was significant. However, after a steady increase until 2011, the incidence tends to stabilize or even to decrease slightly in recent years (R2 = 0.739). The highest incidence was observed in the Amazon region, while the incidence in the coastal region and the south of the country was low (Fig 20). The seasonal incidence is constant for most of the year with a net decrease in the middle of the dry season (mainly from June to September).
There was no information about Saint Lucia. However, the epidemiological situation should be comparable to that of Martinique, which corresponded to about ten bites per year (6 per 100,000 inhabitants) and one death every 5 to 10 years (0.1 per 100,000 inhabitants). Bothrops caribbaeus, a species close to B. lanceolatus, is endemic to the island [8; 46].
Notification of snakebites was not mandatory in Suriname and no information on envenomation has been found. Based on the situation in French Guiana, the annual number of snakebites can be estimated at 135 (25 per 100,000 inhabitants) and the number of deaths at 5 deaths (0.9 per 100,000 inhabitants).
Notification was not mandatory in the island of Trinidad for which there was no information on snakebites.
Based on the data collected in coastal Venezuela and Guyana, it can be expected 130 snakebites (10 per 100 000 inhabitants) and 1 to 2 deaths (0.1 per 100 000 inhabitants) each year.
Four poisonous species occur in Trinidad: Micrurus lemniscatus and M. circinalis, both Elapids, and Bothrops atrox and Lachesis muta that are vipers. M. circinalis and M. fulvius are present in some Bocas islands. There is no Elapidae or Viperidae in Tobago [8].
The notification of snakebites in the US was old but hardly available online. Several sources were used and the data were regularly reported in the literature [47–56]. These data were based on notifications from separate systems but were consistent and highly convergent.
Between the late 1950s and early 2000s, incidence decreased by half (3.6 versus 1.7 per 100,000 population) as a result of both the reduction in the number of bites (6,680 in 1959 versus 4,735 in 2005) and the increase in population (185 million versus 285 million). The reduction in incidence concerned most of the States, particularly in the southern and eastern US (Fig 21). However, using the National Electronic Injury Surveillance System, Langley et al. [56] estimated the number of snakebites (including from non-venomous snakes) to be close to 9,200 on average per year over the period 2001–2010. The number of bites for which the species was identified as venomous would be more than 2,800 per year. Furthermore, Morgan et al. [57] reported 97 health deaths from 1979 to 1998, i.e. 4.85 on average per year (0.002 per 100,000 population).
The population at risk was predominantly composed of people whose age is between 10 and 50 years. However, the age-specific incidence showed a peak in teenagers (incidence higher than 5 bites per 100,000 young people aged 10–14 years) and then a steady decrease in adults to about 2 bites per 100,000 Subjects over 65 years of age. The sex ratio (M/F) was 2.7. Most bites occurred from late spring to fall [53].
However, the information provided by the various databases did not detail whether the bites were accidental or illegitimate, the latter probably more frequent in USA, and not seasonal.
Notification of snakebites was mandatory but data were not available online. However, the Ministry of Health published a summary report on snakebites between 1986 and 2001 and a second on the cases of 2010 and 2011. Despite the lack of information between 2002 and 2009, the incidence was likely to be stable.
There are nearly 80 snakebites annually (2.4 per 100,000 population) and 2 deaths (0.033 per 100,000 population). The geographical distribution showed a very high incidence in the eastern part of the country, high in the west and low in the south, especially in the Montevideo region (Fig 22). The age-specific incidence was the highest in young subjects between 15 and 30 years of age. The sex ratio was highly imbalanced in favor of man (M/F = 4.9). The seasonal incidence showed a marked increase in the spring-summer period (October to April) with a peak in March (average cases twice higher than those of other summer months).
The reporting of snakebite incidence and mortality has been mandatory since 1995 and has been available online since 1996 and 1995 respectively [58].
From 1995–96 to 2012, the average number of snakebites and deaths was 5,700 (20 per 100,000 population) and 32 (0.1 per 100,000 population) a year, respectively. Incidence increased from 1996 to 2006 (R2 = 0.7194) and then drastically decreased until 2011 (R2 = 0.9576, the last available year. The overall trend is slightly decreasing from 1996 to 2011 (R2 = 0.1507). A possible explanation could be deterioration in the collection of data after 2010 but it is not excluded that changes in economical activities induced a lower snakebite risk.
The geographical distribution was relatively homogeneous (Fig 23). There is a correlation between the mean incidence of snake bites and population density (R2 = 0.6568). Interestingly, the incidence was likely to be underestimated–compared to data from other countries–in some states of the Amazon region, which could be due to either low performances of case reporting system or peculiar treatment seeking behavior by patients, both linked to poor health care offer.
Mortality was relatively constant over time [59]. However, the relative risk of death as a function of age was roughly constant from childhood to adulthood up to 40 years (between 0.05 and 0.09 per 100,000 subjects of each age group) and rose in older people to exceed 0.5 per 100,000 population above 60 years of age.
Every year, near 60,000 snakebites (6 per 100,000 inhabitants) are managed by the health services of the Americas. Despite the lack of mortality data in a few countries, most of which are small and poorly populous, the total number of deaths can be estimated at 370 per year (0.04 per 100,000 inhabitants), based on the data from the neighboring countries and risk factors described below.
The previous epidemiological estimates, based mainly on medical and scientific literature, mentioned greater numbers of snakebites: about 115,000 [84,110–140,981] with 2,000 deaths [652–3,466] in the study by Kasturiratne et al. [2] and even 150,000 snakebites of which 5,000 deaths in Chippaux's one [1]. The number of bites did not decreased in the last twenty years (see below), in contrary of deaths. These figures were therefore overestimated, which can be explained by the highly biased epidemiological source of information. Indeed, most authors who publish epidemiological or clinical studies on snakebites report facts upon regions with high incidence—or severity—of envenomation that are often poorly representative [60].
Nevertheless, the general incidence is much lower than in Asia or Africa [1; 2; 61], excluding for particular regions such as the Amazon. However, mortality remains moderate, except in enclosed or poorly equipped areas.
Most of the data collected in this study comes from the Ministries of Health of the concerned countries.
Until now, epidemiological surveys were needed to obtain information that was most often limited geographically according to the constraints and choices of the investigators. Sometimes methodological biases, particularly in site selection, led to approximations or significant errors in the estimation of the incidence or severity of envenomations [62]. For the past decade, mandatory reporting of snakebites resulted in better epidemiological data in most countries of the Americas.
Mandatory reporting of cases allows covering a country as a whole rather than a few sites chosen by the investigators, leading to poorly representative figures. However, data gaps and limitations are still observed resulting from a poor surveillance system. On the one hand, it is expected that over time the data collection will improve and on the other hand the standardization of the questionnaires will make it possible to have more robust, reliable and complete information. For example, useful, often missing data, particularly severity, treatment (brand and dose of antivenom) and clinical outcomes (mortality, sequelae) need to be collected, which is not currently the case in most situations. However, in some countries (Brazil, United States), these data are available, showing that such a goal is feasible.
It is rarely stated whether the notification of snakebites included asymptomatic bites, which is probably the case in most countries. Asymptomatic snakebites may result either from a bite by a non-venomous snake or a venomous one that did not inject venom (dry bite). According to the countries and authors, asymptomatic snakebites represent between 10 and 40%, about one third of which are dry bites [7; 63; 64].
As a consequence, the comparison with the recent literature has been very useful for, a) confirming (or supplementing) the data from other sources and, b) providing additional information, in particular on the clinical severity of envenomations, details on circumstances of the bite or implementation of the treatment.
It was emphasized that the notification was not very precise and reliable, at least variable from one country to another. However, the reporting system improves over the time and, of course, provides a minimal—conservative—incidence of snakebites seen by healthcare institutions from which it can be inferred treatment needs, especially antivenoms. The increase in incidence observed in some countries (Bolivia, Brazil, Colombia, Mexico, Peru, Venezuela) can be attributed to an improvement in data collection, particularly in the early years of its implementation. The stabilization or reversal of the upward trend confirms this. However, environmental (e.g. reduction of snake population) or demographic (population migration to urban centers with low snakebite risk (see below)) causes should not be underestimated. It is notable, for example, that the incidence is often similar on both sides of a border between two neighbor countries—despite likely differences in data collection efficiency -, reflecting a constant figure regarding both risk and population reaction to the snakebite. Actually, administrative policies are different on each side of the border, but populations are often the same on the both sides… It is known, for example, that many patients prefer to use alternative medicine rather than a modern treatment provided by health center. This occurrence is poorly addressed in Latin America, but it probably plays a significant role in underestimating the incidence and possibly severity (mortality) of envenomations. However, some inconsistencies can be explained either by different environmental conditions affecting the risk factors mentioned below, or by significant differences in the quality of the notifications. The report still suffers from inadequacies, resulting in underestimations of snakebite incidence and mortality in some regions of Latin America [44].
The geographical distribution of the incidence was heterogeneous: it was higher in the intertropical region and in developing countries. The incidence depends mainly on environmental and anthropic factors that are detailed below.
The number of deaths appeared to be more difficult to determine due to the lack of notification in several countries. However, these countries are generally sparsely populated regions, which limit the impact on the total result. We proposed here a reasonable estimate for each of these countries at the risk of a trivial error.
Basically, the incidence results from the encounter between a man and a snake. It is therefore legitimate to consider the activities and the presence of the first as well as the behaviors of the latter. It is difficult to explain what affects snakebite incidence because of the complexity of possible causes and their interactions, such as the biology of animal populations composed of many species or the demographics of human populations that are dependent on many social, economic, environmental factors. The coefficient of determination R2 indicates the proportion of the variance in the dependent variable that is predictable from the independent variable, i.e. it gives some information about the goodness of fit of a model. The closer R2 is to 1, the better the data match the model, but this does not mean the model is relevant.
Incidence tends to grow mechanically as a function of demography although there is a partial offset related to a decrease due to anthropization of the environment which reduces snake populations and/or snake-man contacts. In addition, the proximity of human populations to the natural environment explains a greater frequency of encounters with snakes. As a consequence, snakebites occur usually in rural areas during agricultural activities, especially in developing countries where farming is an important and weakly mechanized economic activity.
Population density was sometimes inversely correlated with the incidence of bites, as in Brazil [20], suggesting that a high human presence limited the development of snake populations. However, other reasons may locally explain the inverse correlation, e.g. when the human population remains large while snakes do not encounter favorable conditions for their development. For instance, the altitude and roughness of the climate appeared to have a negative impact on snake populations as shown in Bolivia or El Salvador, and Canada or Argentina, respectively.
Isolated areas are the most affected, mainly due to lack of good roads linking urban centers and activities of the population performed in precarious conditions (forestry, subsistence agriculture and hunting, among others). These occurrences increase both the likelihood of encounters with snakes and the difficulty of receiving timely medical help. As a consequence, scarcity of health centers is a factor that indirectly influences the incidence of snakebites and directly (and significantly) affects the clinical outcomes of envenomations [33; 44; 65].
The abundance of snakes, especially species that inhabit cultivated or settled areas and sometimes even reproduce there, varies according to climatic (heat and humidity) and environmental (vegetation and landscape) factors that determine food supply, both qualitative and quantitative, and camouflage opportunities [66]. While some species established in natural environments, such as the Amazon rainforest, e.g. Bothriopsis taeniata, are absent or rare in anthropogenic areas, others come near to human settlements and may even grow there [67], at least to some extent. Some species of Crotalus, e.g. C. viridis or C. oreganus in the USA [68; 69], or Bothrops, as Bothrops asper in Costa Rica [36], are attracted to anthropogenic areas where they find their food.
Ecological niche modeling (ENM) allows, using appropriate algorithms, to predict the geographic distribution of a species from climatic and environmental data. Yañez-Arenas et al. [70] used the ENM to assess the potential distributions of several species of rattlesnakes in Veracruz and to associate them with a prediction of abundance estimated by the distance from the niche centroid (DNC). These authors found a significant inverse relationship between the snakebites and DNCs of two common vipers (Crotalus simus and Bothrops asper), partially explaining the variation in the incidence of snakebites. Moreover, the DNCs of the two vipers, combined with the marginalization of human populations, accounted for 3/4 of the variation in incidence. Thus, several factors, environmental, socio-economic and sanitary, contribute to explain the incidence of snakebites.
Populations at risk were very similar in most countries. While children and teenagers constituted an important part of the population, sometimes the majority in developing countries, they were not the mostly bitten. Population at risk was predominantly composed of young men between the ages of 15 and 45, living in rural areas and bitten during agricultural activities. This may explain why bites occur most often during hot (summer) and wet (rainy season) periods, usually at harvest time.
The severity of the envenomation, in particular mortality, is related to the species, but also the size, of the snake responsible for the bite, which determine the composition of the venom and the quantity injected respectively [14; 15; 71]. This explains why some snakebites are asymptomatic, when the snake is not venomous, or when it does not inject its venom [6; 7; 63; 64]. It is more difficult to explain some of the factors identified by Jorge et al. [71] as the season or time of day. This may be due to a particular distribution of species within stands, depending on time and space according to their ecological tropisms.
Age of the patient appeared to be a risk factor, especially at both ends of life, in children and elderly persons–a priori more vulnerable [72]. However, as we have seen above, children are not the most exposed.
In addition, the mortality and incidence of complications–most notably the sequelae–depend on the management of snakebites, i.e. the health care system as a whole (number and distribution of health facilities, equipment, access to antivenoms and adequacy of therapeutic protocols, skill of health personnel, etc.). For example, the significant decline in mortality in many countries–particularly in Costa Rica [30–32], Ecuador [34], Mexico [43] and Venezuela [59] while the number of snakebites in these countries remained stable or even increased–can be attributed to better management of snakebites, notably through the improvement of primary health care and access to medical services, including availability of antivenoms.
However, other factors may also affect the mortality and severity of envenomations, such as the availability of health centers and treatment, which may be very irregular, particularly in remote areas where activities of the indigenous population are often very close to nature. The delay in treatment may thus compromise the clinical course of envenomation. Nevertheless, the treatment seeking behavior is complex and many patients, particularly in remote areas, still use traditional medicine. The latter should be associated with modern medicine in order to define relevant recommendations that do not put them into competition but optimize the therapeutic approaches to avoid complications and disabling sequelae as is still often the case.
This study summarized the burden and epidemiological characteristics of snakebites in the American continent. The incidence and severity of envenomation appeared to be lower than previously assessed, although many risk factors have been already known and studied. This work showed the importance of mandatory reporting of snakebites to improve their management, provided that health authorities endorse, analyze and exploit the data.
It therefore seems necessary to continue this effort, improve the case reporting system and take the measures that can be inferred from the obtained analysis of the available information.
|
10.1371/journal.pbio.2000689 | The enteric nervous system promotes intestinal health by constraining microbiota composition | Sustaining a balanced intestinal microbial community is critical for maintaining intestinal health and preventing chronic inflammation. The gut is a highly dynamic environment, subject to periodic waves of peristaltic activity. We hypothesized that this dynamic environment is a prerequisite for a balanced microbial community and that the enteric nervous system (ENS), a chief regulator of physiological processes within the gut, profoundly influences gut microbiota composition. We found that zebrafish lacking an ENS due to a mutation in the Hirschsprung disease gene, sox10, develop microbiota-dependent inflammation that is transmissible between hosts. Profiling microbial communities across a spectrum of inflammatory phenotypes revealed that increased levels of inflammation were linked to an overabundance of pro-inflammatory bacterial lineages and a lack of anti-inflammatory bacterial lineages. Moreover, either administering a representative anti-inflammatory strain or restoring ENS function corrected the pathology. Thus, we demonstrate that the ENS modulates gut microbiota community membership to maintain intestinal health.
| Intestinal health depends on maintaining a balanced microbial community within the highly dynamic environment of the intestine. Every few minutes, this environment is rocked by peristaltic waves of muscular contraction and relaxation through a process regulated by the enteric nervous system (ENS). We hypothesized that normal, healthy intestinal microbial communities are adapted to this dynamic environment, and that their composition would become perturbed without a functional ENS. To test this idea, we used a model organism, the zebrafish, with a genetic mutation that prevents formation of the ENS. We found that some mutant individuals without an ENS develop high levels of inflammation, whereas other mutants have normal intestines. We profiled the intestinal bacteria of inflamed and healthy mutants and found that the intestines of inflamed individuals have an overabundance of pro-inflammatory bacterial lineages, lack anti-inflammatory bacterial lineages, and are able to transmit inflammation to individuals with a normally functioning ENS. Conversely, we were able to prevent inflammation in the ENS mutants by either administering a representative anti-inflammatory bacterial strain or restoring ENS function. From these experiments, we conclude that the ENS modulates intestinal microbiota community membership to maintain intestinal health.
| The intestinal tract serves to harvest nutrients and energy, protect against harmful toxins and pathogens, and clear out waste. These functions can be modulated by both the enteric nervous system (ENS) and the trillions of symbiotic bacteria that reside within the gut [1–3]. Importantly, the influence of microbiota on intestinal functions and health depends on the constituent microbes. Alterations in microbial composition from those observed in “healthy” subjects are often defined as “dysbiotic,” which refers to communities that become perturbed in their composition such that they acquire pathogenic properties [4–6]. Given that the composition of the microbiota is critical for host health, it is significant that the intestinal microbial community is generally stable despite the highly dynamic internal environment of the intestinal tract [7,8], which experiences disruptions such as influxes of ingested matter, host secretion and epithelial cell turnover, and coordinated outward flow of material. How microbial community stability is achieved amid these constant perturbations is unknown. Hosts with impaired intestinal motility can develop dysbiosis and intestinal pathology [9,10], which suggests a profound role for the ENS in constraining microbiota composition. Here, we explore how the ENS shapes the ecology of the intestine, and we address key questions about the assembly of dysbiotic microbial communities, their functional properties, and strategies for their treatment—three aspects of dysbiosis that have been challenging to address from observational studies in humans. Our analysis reveals how, without ENS constraint, imbalances in pro- and anti-inflammatory members of the microbiota can drive intestinal pathology.
The most severe example of ENS dysfunction in humans is Hirschsprung disease (HSCR), an enteric neuropathy that results from a failure of neural crest–derived cells to form the distal ENS [3]. Approximately 30% of HSCR patients develop a severe form of intestinal dysbiosis, known as Hirschsprung-associated enterocolitis (HAEC) [9–11], which is distinguished by diarrhea, distension, fever, and, in extreme cases, sepsis and death [12]. Studies suggest that the etiology of HAEC has a microbial component, as both pathogenic bacteria [13] and alterations in commensal communities [9,10] have been linked to HAEC. Interestingly, patients with a broad range of human diseases, such as inflammatory bowel disease (IBD), cystic fibrosis [14], diabetes [15], malnutrition [16], and myotonic muscular dystrophy [17,18], also experience debilitating gastrointestinal (GI) symptoms. Although cause and effect are difficult to determine, these diseases are associated with both small intestinal bacterial overgrowth, a clinical syndrome often seen with impaired intestinal motility, and an altered microbiota, suggesting that impaired ENS function could be a driver of dysbiosis.
To explore how the ENS may prevent dysbiosis by constraining microbial populations, we turned to a zebrafish model of HSCR. Multiple well-described zebrafish lines carry mutations in HSCR loci [19–22]. The most extreme ENS loss is seen in mutants homozygous for a null mutation in the HSCR gene sox10 [23,24]; these mutants entirely lack an ENS [24]. The mutant allele t3 (sox10t3) homozygotes have diminished rhythmic peristaltic activity [21], making this an ideal model for dissecting the role of the ENS in host–microbe interactions. Zebrafish are well suited for examining ENS contributions to microbiota composition because we can monitor ENS development, absolute bacterial abundance, and disease phenotypes, such as neutrophil accumulation, across the entire intestine of individual larvae. Thus, we can assess system-level functional readouts that describe properties of the associated microbiota. Furthermore, the high fecundity and ease of working with zebrafish provide us with large sample sizes to increase the power of our experiments such that we can monitor how natural microbiota variation at the species level drives phenotypic variation.
In this study, we demonstrate that the ENS constrains the abundance and composition of the microbiota. We find that loss of the ENS in sox10t3 mutants results in assembly of a dysbiotic community leading to a microbe-driven intestinal inflammation that varies among individuals and resembles HAEC. Microbiota profiling across the spectrum of inflammatory states revealed that extreme intestinal inflammation is linked to an outgrowth of pro-inflammatory bacterial lineages and a reduction of anti-inflammatory bacterial lineages. Moreover, administering representative anti-inflammatory bacterial strains or transplanting wild-type (WT) ENS precursors to restore a WT ENS corrects the pathology in sox10t3 mutant hosts. Our analysis reveals that ENS function is a key feature of intestinal health that constrains the composition of the resident microbiota and prevents overgrowth of bacterial lineages that can drive disease.
The complete loss of ENS in sox10t3 mutants (S1 Fig) results in defective intestinal motility [21]. Given the connection between altered intestinal motility and small intestinal bacterial overgrowth, we hypothesized that functional consequences of these mutants would include changes to intestinal ecology and alterations in resident microbial populations. To visualize the abundance and distribution of bacteria along the length of the intestine, we used fluorescent in situ hybridization (FISH). In sox10t3 mutants, we noted large populations of bacteria throughout the intestine, with marked accumulations of bacteria at the esophageal-intestinal junction (Fig 1A and 1B), a location not typically heavily colonized with bacteria. We also quantified the total number of colony-forming units (CFU) per intestine and found that sox10t3 mutants had a significantly higher bacterial load (Fig 1C). These results suggest that sox10t3 mutants experience bacterial overgrowth, which is consistent with defective intestinal transit. Defective intestinal transit has been observed in mutants in another allele, sox10m241, which have intestinal peristalsis but do not clear ingested fluorescent beads as well as WTs [25]. To demonstrate delayed intestinal transit in sox10t3 mutants, we adapted a previous single color assay [26] into a two-color intestinal transit assay (S1 Fig). The delayed transit and impaired clearance we observed in sox10 mutants likely contribute to bacterial overgrowth within their intestines. For the work described in this manuscript, we use sox10t3 mutants, hereafter referred to as sox10 mutant or sox10-.
We next asked whether the bacterial overgrowth phenotype in sox10- resulted in signs of intestinal inflammation. Thus, we quantified intestinal neutrophil populations, a marker of inflammation, in cohoused WTs and sox10 mutants by staining for the neutrophil-specific enzyme myeloid peroxidase. At 6 d post fertilization (dpf), intestinal neutrophil accumulation in sox10 mutants was significantly increased compared to WTs (Fig 1D and 1E). Notably, sox10 mutants exhibited a much greater variation in intestinal neutrophil accumulation (0–18; n = 30) compared to WT siblings (0–7; n = 31); some sox10 mutants had intestinal neutrophil levels similar to WTs, whereas others had significantly elevated neutrophil populations. Intestinal neutrophil accumulation under homeostatic conditions in WT fish requires the pro-inflammatory tumor necrosis factor (TNF) pathway [27,28]. The increased neutrophil response in sox10 mutants also depends on this pathway, as inhibiting expression of the TNF receptor using an antisense morpholino [27,28] abolished the increased neutrophil response (Fig 1D and 1E). Another indicator of intestinal pathology is epithelial cell proliferation. At 6 dpf, sox10 mutants had markedly increased intestinal cell proliferation relative to cohoused WT animals. Unlike the normal intestinal epithelial cell proliferation response to microbiota, which is TNF independent [29], we found that elevated cell proliferation in the sox10 mutant intestine was TNF dependent (Fig 1F), suggesting that this was an inflammation-dependent pathological response.
To determine whether the intestinal microbiota of sox10- hosts is necessary to induce the increased intestinal neutrophil response, we derived sox10 mutants and their WT siblings germ free (GF). We found that GF sox10 mutants have a low neutrophil population, indistinguishable from their WT siblings (Fig 2A). To determine if the microbial community established in sox10 mutants is sufficient to induce inflammation, we performed an experiment in which we transferred microbiota from sox10 mutants into WTs. As donors, we used microbial communities from conventionally raised (CV) WT, sox10 mutant, or WT intestinal alkaline phosphatase morpholino (iap MO)-injected larvae. iap MO-injected fish are hypersensitive to lipopolysaccharide and thus develop elevated intestinal inflammation without evidence of dysbiosis [27]. These fish serve as control for the possibility that nonbacterial factors such as host pro-inflammatory cytokines rather than microbial derived factors cause transmissible intestinal inflammation (Fig 2B) [30]. At 6 dpf, for each separate group (WT, sox10-, and iap MO), we dissected, pooled, and homogenized the donor intestines. As a negative control, we included transplantation from homogenized intestines of GF fish. The homogenate from each group was inoculated into flasks housing GF 4 dpf WT fish (Fig 2C). We found that inoculation with microbes from sox10 mutants was sufficient to induce elevated intestinal inflammation in WTs as compared to inocula from GF, CV WT, or CV iap MO fish, none of which induced intestinal inflammation (Fig 2D). To test whether the capacity of sox10 mutant microbiota to induce elevated neutrophils was due to increased bacterial load, we transplanted 5× CV WT microbes, which corresponded to the bacterial load of sox10 mutant transplants. This larger inoculum did not induce more intestinal inflammation (S2 Fig), which indicates that the microbial community assembled in sox10- hosts is functionally distinct from WT microbiota and is sufficient to induce inflammation in fish with a normal, functional ENS.
sox10 mutants exhibit a wide range of intestinal neutrophil populations (Figs 1D and 2A) as well as variation in bacterial load (Fig 1B). Therefore, we asked whether intestinal neutrophil abundance corresponded to increased bacterial abundance. We used transgenic sox10 mutant hosts expressing green fluorescent protein (GFP) under control of the neutrophil-specific mpx promoter to quantify both neutrophil population and intestinal bacterial load in individual fish (Fig 3A). When we compared sox10 mutants that fell in the bottom half of neutrophil response (“sox10- low”) or in the top half of neutrophil response (“sox10- high”) to WTs, we found that all sox10 mutants, regardless of neutrophil level, carried significantly higher bacterial loads than WTs (Fig 3B). Thus, impaired intestinal clearance (S1 Fig) leads to an increased bacterial load; however, the bacterial overgrowth per se in sox10- does not drive an increased intestinal neutrophil response. We further characterized the pro-inflammatory signature of the sox10- high- and low-neutrophil subsets by monitoring expression of a panel of immune genes in the intestine (Fig 3C). These results aligned with our observations of the neutrophil population, as the sox10- high-neutrophil subset had elevated levels of mpx, saa, and tnfα expression compared to WT and the sox10- low-neutrophil subset (Fig 3C); however, the increase in saa transcription was the only one to reach statistical significance. Consistent with the significantly elevated intestinal neutrophil response in these samples, saa is known to mediate intestinal neutrophil behavior stimulated by microbes [31]. Collectively, our results suggest that a pro-inflammatory compositional change occurs in the microbial community of a subset of sox10 mutants.
To address the possibility of a pro-inflammatory compositional change in the sox10- microbiota, we profiled microbial communities by performing 16S rRNA gene sequencing on intestinal communities isolated from cohoused WT and sox10 mutant individuals. We collected samples across three independent experiments. To uncover differences in microbiota composition that explain the variable severity of neutrophil accumulation in sox10 mutants, we collected intestinal neutrophil response data for the same individuals from which we isolated microbial DNA and grouped samples as “WT,” “sox10- low” (intestinal neutrophil response 0–8), or “sox10- high” (intestinal neutrophil response of greater than or equal to 22); these groups include the top 26% and the bottom 29%, respectively (Fig 4A). By standard metrics of community variability (non-metric multidimensional scaling of Canberra distances, richness, Faith’s Phylogenetic Diversity, unweighted UniFrac), these three groups were not significantly different (S3 Fig), which indicates that these communities are largely made up of the same microbes, and community differences driving neutrophil differences are perhaps due to changes in minor members [28].
We next asked whether the relative abundance of any bacterial operational taxonomic units (OTUs) correlated with intestinal neutrophil number across all individuals surveyed in the study. Of 129 OTUs present in at least 20 individuals, we found a small subset whose percent abundance was associated with neutrophil number, as measured by Spearman’s correlations. Strikingly, these neutrophil-associated OTUs were tightly clustered in only two genera found in this population of fish intestines (Fig 4B). The Escherichia/Shigella genus (hereafter referred to as Escherichia) had ten OTUs that negatively correlated with neutrophil abundance, although only two had borderline significance after false discovery rate correction. All OTUs of the Vibrio genus had significant positive correlations with neutrophil abundance (Fig 4B). Examination of OTU abundances revealed not only that the two most abundant genera were Vibrio and Escherichia (S3 Fig) but also that they were significantly decreased and increased, respectively, in the “sox10- high” group relative to the WT and “sox10- low” groups (Fig 4C and 4D).
The observation of pro-inflammatory activity associated with Vibrio is consistent with our previous analysis of a zebrafish-derived Vibrio strain (ZWU0020) [32] that is phylogenetically closely related to the Vibrio OTUs in the current experiment (Fig 4E). Previously, we showed that Vibrio strain ZWU0020 (hereafter referred to as Vibrio Z20) promotes intestinal neutrophil accumulation in a concentration-dependent manner in gnotobiotic zebrafish [28]. Similarly, in the current study, the log10(relative abundance) of Vibrio was significantly positively correlated with neutrophil number (Fig 5A), and we also observed that the log10(relative abundance) of Escherichia was negatively correlated with intestinal neutrophil accumulation, although the amount of variation explained was low (Fig 5A). This relationship mirrors the relationship we previously observed in simple microbial communities in gnotobiotic zebrafish between the abundance of Shewanella strain ZOR0012 (hereafter referred to as Shewanella Z12) and a proportional decrease in neutrophil number [28]. Of note, two OTUs from the Shewanella genus included in our analysis in this study did not have a significant correlation with neutrophil number. Combining the loss of Escherichia and the gain of Vibrio does not increase the amount of variation in neutrophil number explained by the gain of Vibrio alone (Table 1, Fig 5B). We used Akaike’s Information Criterion (AIC) [33] to test the relative quality of each of these models; the model that accounts only for Vibrio reports the lowest AIC value, which identifies Vibrio as the best microbial predictor of intestinal neutrophil number variability (Table 1). These analyses suggest that a balance of the Vibrio and Escherichia lineages may be important for maintaining intestinal homeostasis, with Vibrio abundance being a major determinant of intestinal inflammation.
To confirm the functional contribution of Vibrio to the increased neutrophil responses in sox10-, we first added Vibrio Z20 to CV sox10 mutants at 4 dpf and assayed neutrophil numbers at 6 dpf. Exogenously added Vibrio Z20 induced a significant increase in neutrophil accumulation over the number seen in CV sox10 mutants (Fig 5C). Furthermore, the absolute abundance of Vibrio Z20 colonizing these fish was positively correlated with intestinal neutrophil number, similar to observations made in the 16S rRNA data set (Fig 5D). We noted that the extent of the increase in neutrophil accumulation upon addition of Vibrio depended on the level of intestinal neutrophils present in control hosts (S4 Fig), which we think reflects fluctuations in bacterial community composition in CV zebrafish between experiments and a limited ability to change the neutrophil-inducing capacity of an intestinal microbiota already dominated by Vibrio strains. We furthermore observed the inflammation-inducing capacity of Vibrio Z20 in monoassociation by adding Vibrio Z20 to GF sox10 mutants. In these conditions, Vibrio Z20 was still sufficient to induce high intestinal neutrophil influx (Fig 5C). In monoassociation the range of Vibrio colonization was too narrow (S4 Fig) to explore a correlative relationship. These experiments support our hypothesis, based on the microbiota profiling of these fish, that an overabundance of Vibrio species causes a dysbiotic and pro-inflammatory microbiota.
We hypothesized that the dysbiotic state of the sox10 mutant intestine could be corrected by balancing the pro-inflammatory activity of Vibrio species with the addition of anti-inflammatory isolates, such as Escherichia species or Shewanella Z12 [28]. Consistent with this prediction, we found that addition of Escherichia coli HS, a commensal Escherichia strain isolated from a healthy human adult [34] that is closely related to Escherichia OTUs in CV fish (S4 Fig), can colonize the zebrafish intestine (S4 Fig), reducing neutrophil numbers in CV sox10 mutants (Fig 6A) and maintaining GF levels of intestinal neutrophil accumulation in monoassociation (Fig 6A). Moreover, the absolute abundance of colonizing E. coli HS in CV sox10 mutants displayed a similar negative correlation with intestinal neutrophils (Fig 6B), as observed with the sequenced OTUs (Fig 5A). Shewanella Z12, another species that displays a negative correlation between abundance and intestinal neutrophil accumulation [28], also reduced intestinal neutrophils in sox10 mutants, which suggests this relationship may be a hallmark of anti-inflammatory bacterial strains (S4 Fig). Thus, sox10- dysbiosis can be corrected by adding anti-inflammatory bacteria to the community. Shewanella Z12 uses an unidentified secreted factor present in cell-free supernatant (CFS) to mediate its anti-inflammatory activity [28] (S4 Fig). However, E. coli HS CFS was insufficient to reduce sox10 mutant intestinal inflammation (S4 Fig), suggesting that these species use two distinct mechanisms to control the host innate immune response.
As an alternative to manipulating the microbiota directly, we postulated that correcting the underlying deficit in the ENS would also alleviate the inflammation in sox10 mutants. To test this hypothesis, we performed a rescue experiment in which vagal neural crest cells, the ENS precursors, were transplanted from WT donors into sox10 mutant hosts. Our previous studies showed a correlation between the number of ENS neurons and gut motility [21]; thus, after the transplant, we assayed for formation of a normal-appearing ENS along with intestinal neutrophil accumulation. Following transplantation, the sox10 mutants that developed a normal-appearing ENS extending along the entire length of the intestine had WT levels of intestinal neutrophils (Fig 6C and 6D), demonstrating that the ENS is sufficient to prevent intestinal inflammation. Together, our results demonstrate that the ENS contributes to intestinal health by maintaining a balanced gut microbiota, revealing a previously unappreciated role for the ENS in host–microbe interactions.
Many intestinal diseases, such as IBD, and many diseases with intestinal symptoms, such as cystic fibrosis [14], are associated with compositional changes in the intestinal microbiota, implying dysbiosis. However, it has been extremely challenging to establish whether such alterations are indeed dysbiotic and the underlying driver of disease. Overcoming this challenge will be an important step toward identifying new therapeutic targets and strategies for ameliorating dysbiosis-associated disease. To connect alterations in microbial communities with host pathology, three questions are crucial to address: (1) How do dysbiotic microbial communities assemble? (2) Which member(s) of dysbiotic communities contribute to disease? (3) How can dysbiosis-related disease be mitigated?
Microbiota are assembled through fundamental ecological processes, including dispersal, local diversification, ecological drift, and environmental selection [35]. We have previously shown that a portion of early larval zebrafish intestinal communities follow a neutral pattern of assembly [36]. This observation suggests that features of the gut environment constrain which microbes colonize and persist in the gut environment. We hypothesized that the ENS, which controls motility and aspects of intestinal homeostasis [3], may also directly or indirectly serve as a significant constraint on intestinal microbial community assembly, such that loss of the ENS constitutes a major ecological shift. Consistent with this hypothesis, we show that zebrafish lacking an ENS have an altered intestinal microbiota and deficits in clearing food from the gut, suggesting gut motility is a mechanism by which the ENS influences microbiota composition. This is further supported by the recent finding that GI transit time is one of the largest predictors of microbiota composition [37]. Moreover, intestinal motility profoundly influences the spatial organization of bacterial populations and has been found to promote competitive exclusion within resident communities [38]. This suggests that abnormal GI transit patterns can significantly reshape ecological interactions within the gut. The ENS also contributes to epithelial barrier function and secretion; however, whether and how these functions are altered in the sox10 mutant has not yet been described. Therefore, observed alterations to the microbial community may be the result of changes to any (or all) of these functions (Fig 7). Of all ENS mutants, the sox10 mutant has the most overlapping characteristics with the human disease HSCR; however, given that the ENS is interconnected with many other organ systems, our work reveals the need to investigate other model systems of ENS dysfunction. Currently, no other available mutant both entirely eliminates the ENS as seen in sox10 mutants and retains normal craniofacial structures [21]. For example, another severe mutant, ret, has a few residual ENS neurons and also exhibits severe craniofacial defects that may impair bacterial colonization [39]. A zebrafish sox10 cell ablation model exists [40] but requires treatment with the antibiotic metronidazole, which would alter the microbiota and confound our experiments. For future experiments, developing a new line in which it is possible to specifically ablate enteric neurons at specified developmental stages will be essential.
Mounting evidence suggests that ENS defects of HSCR patients, as well as those of HSCR animal models, are not restricted to the aganglionic region of the intestine but rather extend to more proximal intestinal regions; thus, these defects are poised to precipitate dysbiosis associated with HAEC [9,41,42]. Patients with chronic IBD can also have functional and structural abnormalities in the ENS that disrupt motility [43,44]. Although these pathologies are generally thought to be secondary to inflammation [45], our data raise the possibility that, regardless of the origin of ENS defects, they have the potential to disrupt the microbial community and thus contribute to a feedback loop that prevents a healthy microbiota from establishing after an inflammatory episode. Such a cycle could explain the variable outcomes of treating IBD patients with probiotics [46]. However, in healthy hosts, this feedback loop may instead enforce stability and homeostasis within the system.
The intimate connection between human health and microbiota suggests that health is an effect of services provided by the microbial ecosystem [35], and thus to manage health through the microbiota, we need to identify the taxa that provide specific ecosystem services. One strategy to identify bacterial species that specifically influence host health or disease phenotypes is to define a dose–response relationship, or correlation, between a phenotype of interest and a microbial isolate. For example, we used gnotobiotic methods to identify Vibrio Z20 as pro-inflammatory in zebrafish by its positive correlation between abundance and intestinal neutrophil number [28], defining a dose–response relationship for this isolate. In this study, we have now expanded this approach to complex communities and discovered that Vibrio species fulfill this pro-inflammatory role in the highly inflamed sox10 mutant gut—finding a positive relationship between relative abundance of Vibrio and number of intestinal neutrophils. Correlations between OTUs and host phenotypes have been important in the identification of “indicator” species of interest in chronic obstructive pulmonary disease [47], ulcerative colitis [48], and asthma [49]. Perez-Losada and colleagues [49] expanded this concept by comparing the host and bacterial transcriptomes of asthmatics and healthy controls. They revealed positive correlations between both bacterial phyla (Proteobacteria) and functions (adhesion) with the pro-inflammatory cytokine IL1A [49]. Similarly, a recent large-scale study used correlations to identify microbial drivers of cytokine expression in healthy humans [50]. These studies highlight the potential for using correlation to identify bacterial species, or properties of bacterial species, that have functional consequences for the host in health and disease.
The zebrafish is an especially good model for this type of analysis because we can manipulate host genetics and the environment to control microbial variability across samples. For human studies, the heterogeneity in microbial communities among subjects may be a limiting factor in performing this type of analysis. Furthermore, zebrafish microbial communities are less complex than those of humans, which allows us to probe the data at a higher resolution, with less data reduction [48], and to analyze host–microbe interactions at the OTU level. Our analysis at this resolution revealed conserved functions at the genera level. The phylogenetic conservation of certain bacterial traits suggests that interactions between zebrafish and their resident microbiota serve as a model for identifying bacterial lineages that influence phenotypes across many host species. For example, like the pro-inflammatory Vibrio identified in sox10 mutants, some Vibrios, such as Vibrio parahaemolyticus, induce inflammatory gastroenteritis in humans [51]. We also identified the Escherichia genus as anti-inflammatory in the zebrafish, and some Escherichia are used as probiotics in the treatment of inflammatory intestinal disorders like ulcerative colitis [52]. These data suggest that characterizing species correlated with host phenotypes in model organisms may help to identify individual members of complex communities that contribute to disease phenotypes.
Viewing host–microbiota interactions as an ecological system allowed us to identify two system components, the ENS and key bacterial species, which greatly influence ecosystem function, as measured by host intestinal inflammation. With this information, we can ask whether manipulation of these components provides us with control over ecosystem function. For example, an expanded population of Vibrio lineages combined with a decreased population of Escherichia lineages in sox10 mutants induces increased neutrophil influx. We manipulated this component by introducing the anti-inflammatory Escherichia or Shewanella Z12 [28] and thus ameliorated the disease phenotype. Notably, the most consistent microbial signature of IBD patients is the loss of an anti-inflammatory species, Faecalibacterium prausnitzii [53], the colonization level of which decreases in a step-wise manner from healthy subjects, to patients in remission, to patients with active colitis, to patients with infective colitis [54]. Furthermore, administration of F. prausnitzii reduces disease severity in mice with chemically induced colitis [55]. These results highlight the important immunomodulatory role played by specific bacterial species within the intestinal microbiota and the need to identify these species to devise therapies for reestablishing control of the intestinal environment and ameliorating dysbiosis.
Treatment of dysbiosis-associated diseases with probiotics is likely to require continual probiotic administration if there is an underlying disease mechanism leading to its depletion. A more fruitful approach would include a treatment for the underlying ecological perturbation along with the introduction of probiotic strains. For example, restoring ENS function via transplantation or drug administration are possible ways to treat ENS dysfunction. We demonstrated that transplantation of WT ENS precursors into sox10 mutant hosts restored a normal-appearing ENS and rescued the inflammatory gut phenotype. We think that the normal inflammatory response indicates restored ENS function. However, in future experiments, determining the functional capacity of the transplanted ENS to restore motility, secretion, and epithelial barrier function will help elucidate which specific ENS functions contribute to the constraint on the intestinal microbiota. Recently, there has been significant success in establishing a functional ENS in mouse by transplantation of induced pluripotent stem cells [56], which, together with our results, suggests that this strategy could contribute to a successful cure of disease in cases of HAEC or IBD.
Here, we have utilized the zebrafish as a powerful model to examine the complex relationship between the ENS, the immune system, and the microbiota. We demonstrated the critical role played by the ENS in shaping the ecology of the intestine by constraining the functional properties of the resident microbiota. Our analysis reveals how, without this constraint, imbalances in pro- and anti-inflammatory members of the microbiota can drive intestinal pathology. The imbalances we discovered could not be described by large changes in phylum level abundances or the acquisition of a single pathogenic lineage but rather by subtle differences in the abundances of key commensal species that have the potential to either protect against or promote inflammation. We note that this discovery reveals the reciprocal relationship between the microbes and the ENS, as ENS activity and development can be altered by microbiota; in fact, individual bacterial species can have distinct effects on ENS function [57]. Furthermore, immune cell responses influence ENS function both under healthy conditions [58] and in inflammatory states [59]. Therefore, intestinal homeostasis depends on a complex tri-directional conversation that occurs between the microbiota, the ENS, and the immune system, with proper functioning of each branch depending on signals from the other two branches. Uncovering new therapeutic strategies for chronic intestinal diseases will require a profound understanding not only of each branch of this system but the multifaceted interactions that connect them and how alterations made to one system ripple out to affect the function of the other two branches. Developing scalable and tractable model systems, such as the zebrafish, in which we can monitor all three branches of this system will be critical for addressing these complex questions.
All zebrafish experiments were done in accordance with protocols approved by the University of Oregon Institutional Animal Care and Use Committee (protocol numbers 15–15, 14-14RR, and 15-83A8) and conducted following standard protocols as described in [60].
CV-raised WT (AB x Tu strain), heterozygote sox10t3- (referred to as sox10-) [24], and Tg(BACmpx:GFP)i114 (referred to as mpx:GFP) [61] fish were maintained as described [60]. Homozygous sox10 mutants were obtained by mating heterozygotes and identified by lack of pigmentation [24]. The sox10t3 line was used for neutrophil experiments unless otherwise indicated. The mpx:GFP line [61] was crossed with sox10+/- adults to create a line that when in-crossed resulted in offspring that were sox10-/- and Tg(BACmpx:GFP)i114 (referred to as sox10, mpx:GFP). No defects were observed in heterozygous siblings, which have pigment, develop normally, and survive to adulthood, and thus they are grouped with homozygous WTs [21,23,24,38,62]. For all experiments, WT siblings and homozygous sox10 mutants were cohoused.
See S1 Text.
Splice-blocking MOs (Gene Tools, Corvallis, OR) were injected into embryos at the one cell stage. For knockdown of TNF, the tr1v1/tr1v2 MOs (1.2 moles and 6 moles, respectively) were used as previously described [27,28]. For knockdown of intestinal alkaline phosphatase, the iape212 MO (3 pmoles) was used as previously described [27].
Zebrafish larvae were fixed in 4% paraformaldehyde (PFA) overnight. Whole larvae were stained with Myeloperoxidase kit (Sigma) following the manufacturer’s protocol and processed and analyzed as previously described [27]. For analysis of neutrophils in mpx:GFP fish, GFP+ cells in the intestine were quantified as previously described [28]. For proliferation, larvae were immersed in 100 μg/ml EdU (A10044, Invitrogen) for 16 h prior to PFA fixation. Subsequent processing and analysis were done as previously described [29]. See also Histology and neutrophil analysis in S1 Text.
At 6 dpf, larvae were humanely killed with Tricaine (Western Chemical, Inc., Ferndale, WA), mounted in 4% methylcellulose (Fisher, Fair Lawn, NJ), and their intestines were dissected using sterile technique. Dissected zebrafish intestines were placed in 100-μl sterile EM, homogenized, diluted, and cultured on tryptic soy agar plates (TSA; BD, Sparks MD). After incubation at 32°C for 48 h for conventionally colonized fish or for 24 h for inoculated fish, colonies were counted.
Zebrafish embryos were derived GF as previously described [63]. All manipulations to the GF flasks were performed under a class II A/B3 biological safety cabinet. Zebrafish inoculated with donor microbial populations were generated by inoculating flasks with 4 dpf GF zebrafish with 104 CFU/mL of donor microbes (1×). Donor microbes were collected by dissecting CV zebrafish and based on colonization data (Fig 1B) each fish was assumed to carry 105 total CFU/gut; a total of 25 dissected guts were pooled and homogenized to create the donor microbes. We inoculated CV fish with live Vibrio (106 bacterial cells/ml), E. coli HS (107 bacterial cells/ml), and Shewanella Z12 (106 bacterial cells/ml) as previously described [28]. For monoassociations, each strain was inoculated at 106 bacterial cells/ml. We isolated and concentrated CFS as previously described [28]. Flasks were kept at 28°C until analysis of myeloperoxidase positive cells on 6 dpf.
RNA isolation and cDNA preparation were performed as previously reported [27] except either five (for saa, mpx, and il1b primers) or 18 (for mmp9 and tnfα primers) dissected intestines were pooled. RNA was harvested by homogenizing and extracting with Trizol reagent (Invitrogen). Contaminating genomic DNA was eliminated using the Turbo DNA-free kit (Ambion) per manufacturer’s instructions. The RNA (100 ng for saa, mpx, and il1b primers; 320 ng for mmp9 and tnfα primers) was used as templates for generating cDNA with Superscript III Reverse Transcriptase and random primers (Invitrogen) following manufacturer’s instructions. The cDNA was measured in a qPCR reaction with SYBR Fast qPCR master mix (Kapa Biosystems). Assays were performed in triplicate using ABI StepOne Plue RealTime. Data were normalized to elfa and analyzed using ΔΔCt analysis. Sequences and annealing temperatures are presented in S1 Table.
Dissected intestines were placed in 2-mL screw cap tubes with 0.1 mm zirconia silica beads and 200-μL sterile lysis buffer (20 mM Tris-Cl; 2 mM EDTA; 2.5-mL 20% Tx-100) and frozen in liquid nitrogen. DNA was extracted using Qiamp DNA micro Kit (Qiagen) as detailed in S1 Text. The microbial communities of each sample were characterized by an Illumina HiSeq 2500 Rapid Run (San Diego, CA) sequencing the 16S rRNA gene amplicon by the University of Oregon Genomics and Cell Characterization Facility. The read length was paired-end 150 nucleotide, targeting the V4 region (primers listed in S2 Table). The 16S rRNA gene Illumina reads were clustered using USEARCH 8.1.1803 [64]. The final OTU table was rarefied to a depth of 100,000 (see S2 Data for metadata, S3 Data for OTU taxonomy, and S4 Data for OTU table). Measures of community diversity and similarity (OTU richness, phylogenetic distances, unweighted UniFrac) were calculated in R using vegan, picante, and GUniFrac (See S2 Code). Correlations were calculated in R, and false discovery rate was adjusted using the Benjamini & Hochberg correction in p.adjust (See S1 Code).
WT donor embryos were labeled by injection of 5% tetramethylrhodamine dextran (3000 MW) at the 1–2 cell stage and reared until the next manipulation in filter-sterilized EM. Embryos at the 12–14 somite stage were mounted in agar, a small hole dissected in the skin, and cells transplanted as previously described [65] and detailed in the S1 Text.
Statistical analysis was performed using Prism (Graphpad software). Statistical significance was defined as p < 0.05. Data whose distributions were bounded by 0 were log transformed + 1 prior to statistical analysis. For correlations in Figs 5 and 6, log transformations of neutrophil number and percent OTU were performed so the data met the assumptions of normality and homoscedasticity for linear regression. We note that the relationships and result of multiple linear regression were the same if the data were not log transformed. Throughout, box plots represent the median and interquartile range; whiskers represent the 5–95 percentile. Data for all figures are available in S1 Data.
|
10.1371/journal.pntd.0000522 | A Novel Animal Model of Borrelia recurrentis Louse-Borne Relapsing Fever Borreliosis Using Immunodeficient Mice | Louse-borne relapsing fever (LBRF) borreliosis is caused by Borrelia recurrentis, and it is a deadly although treatable disease that is endemic in the Horn of Africa but has epidemic potential. Research on LBRF has been severely hampered because successful infection with B. recurrentis has been achieved only in primates (i.e., not in other laboratory or domestic animals). Here, we present the first non-primate animal model of LBRF, using SCID (-B, -T cells) and SCID BEIGE (-B, -T, -NK cells) immunocompromised mice. These animals were infected with B. recurrentis A11 or A17, or with B. duttonii 1120K3 as controls. B. recurrentis caused a relatively mild but persistent infection in SCID and SCID BEIGE mice, but did not proliferate in NUDE (-T) and BALB/c (wild-type) mice. B. duttonii was infectious but not lethal in all animals. These findings demonstrate that the immune response can limit relapsing fever even in the absence of humoral defense mechanisms. To study the significance of phagocytic cells in this context, we induced systemic depletion of such cells in the experimental mice by injecting them with clodronate liposomes, which resulted in uncontrolled B. duttonii growth and a one-hundred-fold increase in B. recurrentis titers in blood. This observation highlights the role of macrophages and other phagocytes in controlling relapsing fever infection. B. recurrentis evolved from B. duttonii to become a primate-specific pathogen that has lost the ability to infect immunocompetent rodents, probably through genetic degeneration. Here, we describe a novel animal model of B. recurrentis based on B- and T-cell-deficient mice, which we believe will be very valuable in future research on LBRF. Our study also reveals the importance of B-cells and phagocytes in controlling relapsing fever infection.
| Research on Borrelia recurrentis, the agent of louse-borne relapsing fever (LBRF), has been hampered by the lack of a feasible non-primate animal model. By using immunocompromised SCID mice deficient in B- and T-cells, we were able to establish a stable, persistent B. recurrentis infection with low spirochetemia. Furthermore, systemic depletion of phagocytes by use of clodronate liposomes increased the numbers of bacteria in blood, which demonstrates the importance of both the humoral response and phagocytosis in controlling relapsing fever infection. Lice are favored by the conditions related to the unfortunate turmoil and refugee camps prevailing in the Horn of Africa, and hence LBRF is more important now than it has been for several decades. The newly published genome sequence of B. recurrentis and techniques to genetically manipulate RF borreliae will be instrumental in understanding its complex biology. We therefore believe that our novel animal model will be a great asset that can facilitate future studies of the infection biology of B. recurrentis.
| Bacteria of the genus Borrelia are spirochetes that cause either Lyme disease or relapsing fever (RF). Borrelia species are transferred from animals to humans by tick bites, with the single exception of B. recurrentis, which is transmitted between humans by the body louse Pediculus humanus humanus. The louse-borne disease is not transmitted by the bite per se, but rather through contamination of abraded skin by feces or coelomic fluid released from lice that are crushed by scratching. P. humanus humanus is strictly a human-specific parasite that lives on the body and in the clothing of its host, and B. recurrentis has been found only in lice and humans [1],[2].
In most cases, louse-borne relapsing fever (LBRF) presents with a sudden onset of fever (typically 38.7–41°C) and chills. The first fever period lasts on average 5–7 days and is accompanied by malaise, nausea, general aches, and enlargement of the spleen and liver. Compared to tick-borne RF, LBRF usually involves fewer relapses, but it results in far greater mortality, which can be as high as 40% if left untreated but as low as 1%–5% when antibiotic therapy is given. The Jarish-Herxheimer reaction and delayed onset of antimicrobial therapy are associated with an elevated mortality risk [1],[2],[3].
LBRF previously occurred worldwide in massive epidemics, the latest of which were seen during the two world wars. Today, the only endemic area is in the highlands of Ethiopia, and sporadic outbreaks have been observed in Sudan, where the disease is associated with natural disasters, famine, and refugee camps [1],[2],[4],[5],[6]. Although B. recurrentis is currently found only in the Horn of Africa, it can become established wherever there are human body lice and thus it has a high potential to cause global epidemics [7], especially today due to the massive political turmoil in the Horn of Africa. A population of lice can increase by 11% a day, which gives a clue as to just how rapidly an outbreak can spread in places like refugee camps [1],[8]. Therefore, LBRF may be even more important now than it has been for decades.
Despite increasing epidemic potential and important scientific progress such as the recent genome sequencing of B. recurrentis [9], research is hampered by the lack of feasible animal models. In the early 20th century, scientists attempted to infect commonly used laboratory animals as well as several species of domestic and wild animals, but only experiments using primates seemed to have been successful [10]. Moreover, B. recurrentis was not isolated in vitro until 1994, when Cutler and coworkers [11] managed to grow a few strains in BSK broth. That event paved the way for microbiological and biochemical investigations, but the lack of an animal model has continued to hinder performance of all types of studies focused on host-pathogen interactions, as well as other experiments that require an in vivo model system.
We used immunodeficient mouse strains to develop an animal model of B. recurrentis infection that is more practical in all aspects compared to a primate-based model. SCID mice carry the Prkdcscid mutation, which results in severe combined immunodeficiency due to a defect in V(D)J recombination; this condition impairs the animals' ability to generate B- and T-cell antigen receptors, thus leading to very low numbers of functional lymphocytes [12]. NUDE mice lack a thymus, and they have impaired T-cell function due to the Foxn1nu mutation [13], whereas the cells of their innate immune system (monocytes, macrophages, natural killer cells and neutrophils) and their complement system remain functional. Mice with the BEIGE mutation (Lystbg) have defective natural killer (NK) cells [14]. Hence, SCID BEIGE mice lack B-, T-, and NK-cells.
Clodronate liposomes have been used extensively in infection models to study the pathobiological effect of systemic depletion of phagocytic cells (e.g., macrophages) and the immunological importance of such cells in combating infection [15],[16],[17],[18]. These liposomes are artificially prepared lipid vesicles that encapsulate clodronate, and they can be injected intravenously to attack phagocytic cells that are present in or in contact with the blood (e.g., macrophages in the spleen and liver). The clodronate is ingested by and accumulated within phagocytic cells, and after an intracellular threshold concentration of the drug is exceeded, the cells are irreversibly damaged and die by apoptosis, as described elsewhere [19]. Still, it is important to remember that clodronate does not eliminate all phagocytic cells, and that new phagocytes will continuously appear as soon as the liposomes are consumed; in other words, the phagocytic activity cannot be totally inhibited. Free clodronate (i.e. released from dead macrophages) has a very short half-life and is quickly removed from the circulation by the renal system, furthermore it can not easily pass through cell membranes and thus can not affect non-phagocytic cells [20].
By using animals deficient in various immune cells and inducing systemic depletion of phagocytes, much can be learned about the immune defense against RF borreliosis and host-pathogen interactions. In the current investigation, we found that SCID mice could support growth of B. recurrentis, which led to a low-grade, persistent disease. Moreover, employing clodronate liposomes to deplete phagocytic cells resulted in a one-hundred-fold rise in the spirochete titers. Thus, in this paper we present the first non-primate animal model for studies of LBRF. We also characterize B. recurrentis infection and compare it with the closely related species B. duttonii, which is virulent in wild-type mice and has been studied extensively in mouse models [21],[22].
The two B. recurrentis strains A11 and A17 were isolated in Ethiopia and were kindly provided by Sally Cutler, University of East London. Isolation and characterization of these strains has been described by other investigators [11]. B. duttonii 1120K3 was kindly provided by Guy Baranton, Institute Pasteur. All bacteria were cultured at 37°C in BSK-II medium supplemented with 10% (v/v) rabbit serum and 1.4% (w/v) gelatin, as described elsewhere [23]. The bacteria used in experiments had less than 10 passages through animals or in vitro in our lab.
To create and validate a B. recurrentis non-primate animal model, we injected 1×106 B. duttonii, B. recurrentis A11, or B. recurrentis A17 subcutaneously (s.c.) into four 6-week-old male mice of each of the following strains (all from Taconic, Denmark): BALB/c (BALB/cAnNTac), BALB/c NUDE (C.Cg/AnNTac-Foxn1nu NE9), SCID (C.B-Igh-1b/IcrTac-Prkdcscid), and SCID BEIGE (C.B-Igh-1b/GbmsTac-Prkdcscid-Lystbg N7). The animals were kept in a filter cabinet and given food and water ad libitum, with all maintenance performed according to Swedish animal welfare guidelines. Tail blood was collected daily, and bacteria in the samples were counted by phase contrast microscopy during the first 20 days of infection. SCID and SCID BEIGE mice infected with B. recurrentis were kept until day 150 post infection (p.i.), and spirochetemia was quantified weekly by microscopy. All animal experiments were approved in advance by the Laboratory Animal Ethics Committee of Umeå University.
Starting one day before infection and subsequently every fifth day, an intravenous (i.v.) injection of 100 µl clodronate liposomes in phosphate-buffered saline (PBS) was administered in the tail of SCID BEIGE and BALB/c mice to deplete phagocytic cells present in the blood and organs (e.g., spleen and liver); this was done as described elsewhere [19]. Since the phagocytes might also have ingested liposomes that contained PBS instead of clodronate, which would have reduced the phagocytic efficiency, we gave mice i.v. injections of 100 µl PBS as negative controls. The animals were infected with either B. recurrentis A17 or B. duttonii 1120K3 five days after the first clodronate injection.
Spirochetemia (non-Gaussian) was analyzed by the Mann-Whitney U-test, and the results are presented as medians with 25th and 75th percentile bars to illustrate variance. Difference in spleen weight was assessed by Student's t-test.
We conducted tests to determine whether any of the commercially available immunodeficient mouse strains can support growth of B. recurrentis. Initially, the two B. recurrentis strains A11 and A17, and B. duttonii strain 1120K3 were inoculated into immunocompetent wild-type BALB/c mice, NUDE mice lacking T-cells, SCID mice lacking B- and T-cells, and SCID BEIGE mice lacking B-, T-, and NK-cells. As expected, only B. duttonii established detectable infection in the BALB/c mice (Figure 1A), which concurred with the results of similar previous experiments [22],[24]. The total B. duttonii spirochetemia did not differ significantly between NUDE mice and BALB/c mice (Figure 1A–B), and neither the A11 nor the A17 B. recurrentis strain caused detectable spirochetemia in BALB/c or NUDE mice, indicating that T-cells are of minor importance for RF immune defense. However, both the B. recurrentis strains did establish infection in SCID and SCID BEIGE mice, although the spirochetemia was about 200-fold lower than that caused by B. duttonii infection (Figures 1 C–D and 2). Due to the lack of antibodies, B. recurrentis spirochetemia did not display a relapsing pattern. Instead, spirochete titers remained fairly constant over time, and there were only minor fluctuations between the mice, which were probably caused by individual, biological variations in the host-pathogen interactions (Figure 2). Both the A11 and A17 bacteria were persistent and remained at fairly low levels in the blood (about 2×105/ml) at least until day 150 post infection. The A11 strain caused a significantly (p<0.01) milder infection, with spirochetemia about half the magnitude of that caused by the A17 strain in both SCID and SCID BEIGE mice (Figure 2). The animals behaved normally and showed no outer sign of disease, and their weight pattern corresponded to what was seen in uninfected animals (data not shown). To ascertain whether the passage through SCID BEIGE mice rendered the bacteria capable of causing spirochetemia in wild-type mice, we inoculated four BALB/c mice with spirochetes obtained from two A11-infected and two A17-infected animals. None of those four mice developed detectable spirochetemia. The B. recurrentis A17 infection induced significant (p<0.01) splenomegaly in SCID BEIGE mice. However, the spleens were much smaller in B-cell- and T-cell-deficient mice than in wild-type mice (Figure 3), which implies activation and multiplication of splenic immune cells. Surprisingly, B. duttonii did not cause a lethal infection in the B-cell-deficient animals, despite the indispensable role of B-cells suggested by the antibody-mediated clearance of antigenic variants in immunocompetent models (Figure 1C, D). Although both B. recurrentis and B. duttonii spirochetemia were higher in B-cell-deficient mice (p = 0.02), which verifies significance of the B-cells, the disease was kept under control by B-cell-independent mechanisms (Figures 1C, D and 2).
Notably, we also found that clodronate-treated (i.e., phagocyte-depleted) mice infected with B. duttonii were unable to restrict the bacterial infection. These animals developed very high spirochetemia, which killed all the SCID BEIGE mice before day 8 p.i. (Figure 4B). The pattern was similar in BALB/c mice, although one single individ survived and managed to control the infection (Figure 4A). B. recurrentis A17 also caused substantial spirochetemia in clodronate-treated SCID BEIGE mice (Figure 4C), although the level that was reached was about 10 times higher than the spirochetemia induced by B. duttonii in untreated wild-type mice (Figure 4A and 4C). However, B. recurrentis was unable to establish infection in phagocyte-depleted BALB/c mice, which underlines the significance of B-cells and possibly also T-cells.
Several attempts have been made to establish B. recurrentis infection in various animals, but, until now, only primate models have been successful [10], which has severely hampered research on LBRF. In the present study, we used mice deficient in T- and B-cells and induced phagocyte depletion by administering clodronate liposomes to create the first non-primate animal model of LBRF infection. The results of our experiments show that B. recurrentis could infect both SCID and SCID BEIGE mice. The two B. recurrentis strains A11 and A17 caused moderate, persistent infection in B- and T-cell-deficient SCID mice and B-, T- and NK-cell-deficient SCID BEIGE mice (Figure 2), but did not induce detectable spirochetemia in either wild-type BALB/c animals or NUDE mice lacking mature T-cells. The A11 and A17 strains we used were isolated from different LBRF patients in Ethiopia [11] and had an estimated history of ∼20 BSK passages. Since RF spirochetes do not lose plasmids in vitro (which is the case in Lyme Borrelia spirochetes), and they maintain infectivity even after several passages in vitro [9],[25], the two strains we chose to use were probably well representative of B. recurrentis.
Despite the ability of RF spirochetes to evade the host humoral response through antigenic variation, antibodies are definitely an important part of immune defense against this disease [26],[27]. This is also reflected by the present findings showing higher B. duttonii spirochetemia and establishment of B. recurrentis infection in the B-cell-deficient mice. Moreover, the interaction between the humoral response and B. recurrentis seems to be somewhat different than noted for other RF-inducing bacteria, since B. recurrentis generally causes 0–4 relapses, whereas other African species cause 3–9, as reviewed by other researchers [1],[2]. T-cells are apparently less important, since we observed that B. duttonii spirochetemia was equally high in wild-type BALB/c animals and the athymic, T-cell-deficient NUDE mice (Figure 1A, B). Similar results have been reported in experiments on both the RF agent B. turicatae and Lyme borreliosis [28],[29]. Furthermore, an investigation of RAG2−/− and RAG2/IL-10−/− mice infected with B. turicatae has suggested that NK-cells play an important role [30]. In contrast, we found that SCID mice with the BEIGE mutation, which causes NK-cell deficiency, showed the same levels of B. recurrentis or B. duttonii spirochetemia as seen in SCID mice (Figures 1C, D and 2). On the other hand, phagocyte depletion by use of clodronate liposomes had a dramatic effect on both spirochetemia and disease. B. duttonii infection was uncontrollable in all animals but one, and even B. recurrentis reached spirochetemia of over 5×108/ml. The apparent “relapses” in clodronate-treated animals were not due to antigenic variation, but instead to partial host recovery from the phagocytic depletion that was repeated every fifth day (Figure 4A, C). Even between the two peaks in B. recurrentis-infected mice, median spirochetemia never receded below 8×106/ml, which is a fairly high level (Figure 4C). These results clearly illustrate that it is important for phagocytic cells to be able to filter and to some extent also control spirochetes in the blood, even in the absence of B- and T-cells. Furthermore, experiments in vitro have indicated that B. recurrentis avoids complement opsonization and phagocytosis by binding complement regulators such as factor H and by degrading C3b from its surface in order to utilize “hijacked” host plasmin [31],[32]. However, this remains to be convincingly demonstrated in vivo. Despite all of these isolated findings, it is essential to bear in mind that the immune system is a tightly connected web of cells and signal molecules, and that removal of one cell type might have other downstream effects on overall immunity.
The recent genome analysis performed by Lescot et al. [9] revealed that B. recurrentis evolved from B. duttonii through extensive genetic decay, probably involving loss of the genes mutS and recA, which are important for DNA repair. Both of these species are considered to be more or less specific to humans, and it has been debated whether this is due to a restricted host infectivity range, or simply to the fact that both the B. duttonii tick vector Ornithodorus moubata moubata and the B. recurrentis louse vector P. humanus humanus strongly prefer humans as a source of food. B. duttonii establishes RF in wild-type laboratory mice and rats that is similar to the human disease, and this spirochete has recently been found in chickens and pigs raised in proximity to tick-infested human dwellings in Tanzania [33], indicating a wider potential host range than was previously believed. In short, B. duttonii can infect different species of animals, but it is ecologically very restricted by the host range of its vector. It might be assumed that B. recurrentis should also be able to infect non-primate animal species, perhaps even more severely than does B. duttonii, since in humans it is basically a more pathogenic strain of B. duttonii. Interestingly, this greater virulence seems to originate from the loss of a gene or genes rather than the gain of novel genes [9], as has also been described in other louse-borne infections, such as Rickettsia prowazekii [34]. B. recurrentis has lost its ability to survive in hosts other than humans, and the way this spirochete is transmitted (i.e., not by the louse bite but by fecal or hemolymph contamination of abraded skin) is a somewhat unsophisticated strategy that suggests a short evolutionary adaptation. It is tempting to speculate that the lost capacity of B. recurrentis to infect non-primate animals may also be a result of genomic decay. B. duttonii probably possesses mechanisms for evading the immune defense of host animals, which have disappeared in B. recurrentis during its rapid evolution simply because they are not needed since humans are the only host. Inasmuch as SCID mice are defective in B- and T-cell production, they readily accept foreign cells and tissues without rejection. Future experiments by our group should be aimed at creating humanized SCID-hu mice that generate human immune cells or carry stem-cell-producing human xenotransplants to pinpoint the factors that restrict LBRF to humans. Such strategies have been successfully applied in other human-specific infectious diseases, as described by other investigators [35]. For instance, SCID-hu mice with human B-cells may constitute a better model of human B. recurrentis infection that can facilitate studies of aspects such as antigenic variation that seems to be responsible for a difference causing fewer relapses compared to tick-borne RF [2].
B. recurrentis was grown in vitro for the first time in 1994 [11], which opened the door for microbiological and immunological studies conducted in vitro [32],[36]. In addition, the complete genomes of B. recurrentis and B. duttonii were recently sequenced and published for the first time [9], and the first site-specific genetic manipulation of an RF agent was performed last year when the variable tick protein was knocked out and reconstituted in B. hermsii [37]. All of these achievements will definitely encourage further development of genetic tools and facilitate molecular biological investigations of RF. Here, we have described the first non-primate animal model of LBRF, which we believe is the fourth cornerstone needed to bring B. recurrentis research into the 21st century.
|
10.1371/journal.pcbi.1005062 | Learning Reward Uncertainty in the Basal Ganglia | Learning the reliability of different sources of rewards is critical for making optimal choices. However, despite the existence of detailed theory describing how the expected reward is learned in the basal ganglia, it is not known how reward uncertainty is estimated in these circuits. This paper presents a class of models that encode both the mean reward and the spread of the rewards, the former in the difference between the synaptic weights of D1 and D2 neurons, and the latter in their sum. In the models, the tendency to seek (or avoid) options with variable reward can be controlled by increasing (or decreasing) the tonic level of dopamine. The models are consistent with the physiology of and synaptic plasticity in the basal ganglia, they explain the effects of dopaminergic manipulations on choices involving risks, and they make multiple experimental predictions.
| To maximize their chances for survival, animals need to base their decisions not only on the average consequences of chosen actions, but also on the variability of the rewards resulting from these actions. For example, when an animal’s food reserves are depleted, it should prefer to forage in an area where food is guaranteed over an area where the amount of food is higher on average but variable, thus avoiding the risk of starvation. To implement such policies, animals need to be able to learn about variability of rewards resulting from taking different actions. This paper proposes how such learning may be implemented in a circuit of subcortical nuclei called the basal ganglia. It also suggests how the information about reward uncertainty can be used during decision making, so that animals can make choices that not only maximize expected rewards but also minimize risks.
| In situations where actions are associated with rewards, knowledge of the reliability of rewards for alternative choices is critical for selecting the optimal action. Normative models have suggested that optimal foraging requires adaptively switching between risk aversion and risk seeking depending on the circumstance [1, 2]. Indeed, experimental data suggest that humans and animals tend to seek or avoid choice options with reward uncertainty in different situations [1, 3]. To implement such policies, animals and humans need to have estimates of the reward variability associated with different sources, as well as the ability to control how this variability should influence their choices. In addition, knowledge of the reliability of reward feedback is important for learning about the mean reward, as it sets the optimal learning rate. Indeed, in high uncertainty situations, a single new data point should not influence the animal’s previously held estimate as strongly as it would in situations where the uncertainty associated with the data point is fairly low [4]. Furthermore, the estimate of reliability of rewards is helpful in optimizing the exploration-exploitation trade-off [5], because when an animal wishes to find which action yields the highest average reward, it takes more samples to get an accurate estimate of the mean reward for actions with more variable rewards. Hence, such actions should be preferably explored.
One of the key regions of the brain underlying action selection is the basal ganglia (BG). The BG is thought to be involved in learning the expected values of rewards that are associated with given actions and in selecting the actions associated with the highest expected values while inhibiting the others. The learning process in BG is facilitated by neurons releasing dopamine (DA), which encode the reward prediction error, defined as the difference between reward obtained and expected [6, 7]. This signal allows BG to update its estimates of reward accordingly [8, 9].
The pathologies that affect the function of BG influence how it learns or makes decisions in situations involving uncertainty. For instance, a subset of patients with Parkinson’s disease, who suffer from selective death of dopaminergic neurons in the substantia nigra in the midbrain, are impaired in a task involving choices between options with different spreads of their respective reward distributions [10]. When they are on medication (DA agonist), these patients exhibit a well-reported phenomenon of obsessive gambling, in which the patients seem to exhibit a change in their subjective values of risk and reward [11]. This change can be reversed by taking the patients off medication [12]. Additionally, manipulating the levels of dopamine in humans and animals adjusts their decision making under risk [13].
These pieces of evidence suggest that uncertainty is encoded in BG (but one has to note that although BG is the main target of dopaminergic projections, DA neurons also innervate cortex, so some of the effects mentioned above may also have cortical contribution). While computational models have been developed to explain how BG can estimate the expected reward [8, 9, 14, 15], it is still unclear how the reliability of the reward can be estimated in BG, given its anatomical and physiological properties.
Here we show that there exists a class of models consistent with the physiology of BG that can at once learn both the expected reward from a given action and the reliability of the reward, i.e., the spread of its probability distribution. We then show how the models can use learned information about reward uncertainty in decision making, and how the models can account for the effect of dopaminergic medications on decision making in tasks involving risk.
In the next section (“Models”), we review previously proposed models of reinforcement learning in BG, on which our models are built. The new models that can learn reward uncertainty are presented in Section “Results”. Readers familiar with the actor-critic model [16] and Opponent Actor Learning model (OpAL) [15] can skip directly to “Results”.
The models of reinforcement learning in BG have been developed in two frameworks: a simpler framework considering only an “actor” and a more complex “actor-critic framework.” We review both of these frameworks, as both can be extended to learning reward uncertainty.
This framework assumes that BG estimates average rewards for selecting different actions. Let Q i ( t ) denote an estimate of expected reward for selecting the action i on trial t. Let us assume that after selecting the action, a reward r(t) is provided, which comes from a distribution with mean μi and standard deviation σi.
We start by considering an abstract Rescorla-Wagner rule [17] for estimating the expected reward for a given action. According to this rule, after receiving a reward, the expected reward is updated in the following way:
Q i ( t + 1 ) = Q i ( t ) + α r ( t ) - Q i ( t ) (1)
According to the above equation, the change in the estimate of the expected reward is proportional to the reward prediction error (r ( t ) - Q i ( t )), scaled by the learning rate constant α, where 0 < α < 1. It is intuitive to see why this rule works: If r(t) is underestimated, our estimate Qi will increase (i.e., Q i ( t + 1 ) > Q i ( t )). If r(n) is overestimated, Qi will decrease, and if r(t) is estimated perfectly (Q i ( t ) = r ( t )), then Qi will remain the same. In addition, the amount we increment by will be scaled by the magnitude of the prediction error ( r ( t ) - Q i ( t ) ), so that we learn more quickly when we have a lot of learning to do than when our estimate is quite close to the true mean already. Also note that having α < 1 ensures that the new data point updates our estimate but does not completely replace it (as would be the case if α were in fact equal to 1), an implicit acknowledgement of the existence of uncertainty in the reward and noise in the system.
The actor-critic model [16] includes two components: an actor that learns tendencies to select particular actions and a critic that learns an overall value of the current context or state. In the actor-critic model, the value V of being in this state is learned by the critic according to the standard Rescorla-Wagner rule [17] (cf. Eq 1):
V ( t + 1 ) = V ( t ) + α r ( t ) - V ( t ) (2)
Note that V(t) is updated regardless of which action i is selected, so V(t) is not an estimate of expected reward associated with a particular action, but rather an average reward in the current state.
In the standard actor-critic model, after choosing action i, the tendency to choose it, which we denote by Qi, is learned by the actor using the following update rule:
Q i ( t + 1 ) = Q i ( t ) + α r ( t ) - V ( t ) (3)
According to the above equation, the tendency to choose action i is also modified proportionally to the reward prediction error, i.e., it is increased if the action resulted in a higher reward than expected by the critic and decreased if the reward was below expectation.
The actor-critic model naturally maps on the matrix-patch organization of the striatum [18]. Such mapping assumes that V(t) is encoded in the synapses between cortical neurons selective for the current context and striatal patch neurons, as shown in Fig 1. The patch neurons directly inhibit dopaminergic neurons [19], so that if the dopaminergic neurons also receive input encoding reward, then their activity may encode r(t) − V(t). The actor part of the model is mapped on matrix neurons [18] that send projections to the output nuclei, which in turn project to areas controlling movement, so they can affect which movement is selected. Finally, the dopaminergic neurons modulate plasticity of the synapses of both patch neurons and matrix neurons. It is worth adding that some studies map actor and critic on dorsal and ventral striatum respectively [20], but this mapping is related to the matrix-patch mapping, as the patch neurons are more common in ventral than dorsal striatum [21].
A recent model called Opponent Actor Learning (OpAL) [15] takes into account the fact that the matrix neurons can be subdivided into two groups, which express D1 and D2 DA receptors, respectively. These project through different nuclei of BG, as shown in Fig 1 [22] and have opposite effects on movement initiation [23, 24]. In particular, D1 neurons project through the “direct” pathway to the output nuclei, and their activity facilitates movements [25] because they inhibit the output nuclei and thus release thalamus from inhibition. By contrast, D2 neurons project through the “indirect” pathway, and their activity inhibits movement [25].
The OpAL model describes learning about the tendencies to choose or inhibit actions i in a given state, which we will denote by G i ( t ) (for Go) and N i ( t ) (for NoGo), respectively. The OpAL model proposes that these tendencies are encoded in the strengths of synaptic connections between the cortical neurons associated with that state and the striatal D1 or D2 neurons selective for action i, respectively [15], as illustrated in Fig 1. In the OpAL model, after selecting action i the synaptic weights are modified according to:
G i ( t + 1 ) = G i ( t ) + α G i ( t ) r ( t ) - V ( t ) (4) N i ( t + 1 ) = N i ( t ) - α N i ( t ) r ( t ) - V ( t ) (5)
Thus if the reward prediction error is positive, the tendency to select the action is increased, while the tendency to inhibit it is weakened, and vice versa. Additionally, in the OpAL model, the reward prediction error is scaled by G i ( t ) and N i ( t ), which prevents G i ( t ) and N i ( t ) from becoming negative. For example, if G i ( t ) becomes close to 0, the changes in its value also tend to 0.
The OpAL model additionally proposes how the probabilities of actions depend on the weights in Go and NoGo pathways, through a generalized version of the softmax rule [26, 27]:
P i ( t ) = exp a G i ( t ) - b N i ( t ) ∑ k exp a G k ( t ) - b N k ( t ) (6)
In the above equation, normalization by the denominator ensures that the P i ( t ) add up to 1 across all possible actions. Parameters a and b control how deterministic the choice is: when a = b = 0, all actions have equal probability, while with higher a and b, the influence of the learned tendencies on choice increases. The relative value of parameters a and b describes to what extent the neurons in the Go and NoGo pathways contribute to choice (when a = b, both pathways contribute equally; otherwise, one pathway dominates). The rationale for introducing two parameters a and b is that the activity levels of the striatal D1 and D2 neurons are modulated in opposite directions by levels of DA; hence, they can differentially contribute to activity in the output nuclei [15] (see Fig 1).
We first describe the conceptually simpler actor-only model, which will allow for a clearer explanation of the essential mechanisms of learning reward uncertainty. Then, we show how the model can explain the effect of dopaminergic stimulation on choice in tasks involving selection between safe and risky options, Subsequently, we present generalizations of the model, and compare it with the OpAL model.
In the models including only the actor, learning about the reward distribution of an individual action is independent of learning about the distribution of another. Thus for simplicity of notation, while introducing the model we will consider just a single context and a single action, and denote the corresponding synaptic weights of D1 and D2 neurons on trial t by G(t) and N(t), respectively. Furthermore, we will denote the mean and standard deviation of reward distribution by μr and σr.
The model employing the original Rescorla-Wagner rule (Eq 1) keeps track of an abstract variable Q(t) that describes the overall tendency to select action i, but in BG this tendency is encoded in the synaptic weights of D1 and D2 neurons, G(t) and N(t). So let us relate these variables by:
Q ( t ) = G ( t ) - N ( t ) (7)
The update rules for the weights in the Actor learning Uncertainty (AU) model have the following form:
G ( t + 1 ) = G ( t ) + α r ( t ) - Q ( t ) + - β G ( t ) (8) N ( t + 1 ) = N ( t ) + α r ( t ) - Q ( t ) - - β N ( t ) (9)
In the equations above, the prediction errors are transformed through threshold-linear functions |x|+ and |x|− which are equal to |x| if x is positive or negative respectively, and 0 otherwise. In other words, |x|+ = max(x, 0), and |x|− = max(−x, 0). Thus if the prediction error is positive, then so is the corresponding term in Eq (8), and G increases, while if the prediction error is negative, then the corresponding term in Eq (9) is positive, and N increases. Furthermore, the decay terms (last terms in Eqs (8) and (9)) are scaled by a separate constant 0 < β < 1.
As we will explain below, the AU model encodes the estimate of mean reward μr in G(t) − N(t), while the estimate of reward spread σr in G(t) + N(t). Before giving a proof for this property, let us first provide an intuition. The AU model encodes the mean reward in G(t) − N(t) due to its similarity with the Rescorla-Wagner rule. In particular, when the reward is higher than expected, G tends to increase, while when the reward is lower than expected, N tends to increase, so in both cases G(t) − N(t) tends to move towards the value of the reward.
To gain some intuition for how the model can encode reward uncertainty in G(t) + N(t), it is useful to consider the changes in the weights in two different cases: when the rewards are deterministic, i.e., of the same magnitude each time the action is selected, and when they are stochastic. In the case of deterministic rewards, on initial trials, reward prediction error will be positive, hence only G will increase but not N, as illustrated in the top left panel of Fig 2. By contrast, in the case of stochastic rewards, on some trials the reward prediction error will be negative. Hence, N will also increase, as illustrated in the top right panel of Fig 2.
Finally, the decay terms in the above equations serve to ensure the convergence of the synaptic weights, as in their absence, the update rules would only allow G and N to either increase or stay the same upon every iteration, but never decrease.
Let us now show that the AU model can learn expected reward. By subtracting Eq (9) from Eq (8) we obtain:
Q ( t + 1 ) = Q ( t ) + α r ( t ) - Q ( t ) - β Q ( t ) (10)
The threshold-linear functions disappear when Eqs (8) and (9) are subtracted, because if the prediction error is positive, the corresponding terms in Eqs (8) and (9) are equal to the prediction error and 0 respectively, so when subtracted give the prediction error. Conversely, if the prediction error is negative, the corresponding terms in Eqs (8) and (9) are equal to 0 and the negative of the prediction error, so when subtracted they also give the prediction error. Comparing Eqs (10) and (1), we note that this update rule is similar to the standard Rescorla-Wagner rule, with an added decay term.
For a fixed value of α, the variable Q never converges when σr > 0, but constantly fluctuates. Nevertheless, it is useful to consider a value around which it fluctuates. After sufficiently long learning, the expected change in Q will be zero. In other words, for large enough t,
E Q ( t + 1 ) - Q ( t ) = 0 (11)
The value of Q(t) at which Eq (11) holds is referred to as the stochastic fixed point, and we will denote it by Q i *. By combining Eq (10) with Eq (11), we obtain:
E α r - Q * - β Q * = 0 (12)
Rearranging the terms in the above equation, we see that Q at the stochastic fixed point is equal to:
Q * = α α + β E [ r ] (13)
Although in the AU model Q* is not equal to the expected reward, it is proportional to it, with a proportionality constant that is equal across all actions. Thus, choosing an action with the highest Q* is equivalent to choosing an action with the highest expected reward.
We now show that the AU model learns reward uncertainty. In order to do so, we will analyze how the sum of the synaptic weights evolves. Thus, let us define:
S ( t ) = G ( t ) + N ( t ) (14)
By adding Eq (9) to Eq (8):
S ( t + 1 ) = S ( t ) + α r ( t ) - Q ( t ) - β S ( t ) (15)
From the above equation we see that at the stochastic fixed point:
S*=αβE [| r−Q*|]=αβE [| r−αα+βμr |] (16)
The above equation implies that when Q* = μr, the sum of G and N is equal to the deviation of the reward from the mean. In S1 Text we illustrate that, when Q* = μr, then S* is directly proportional to the standard deviation or variance of rewards (depending on the shape of the reward distribution). When Q* ≠ μr, S* is not exactly proportional to the deviation of the rewards from the mean. To see more clearly when it approximates the deviation, let us rewrite the above equation as:
S * = α β E ( r - μ r ) + 1 α β + 1 μ r (17)
From the equation above, we see that S* becomes proportional to the deviation of rewards when the second term inside the expected value is dominated by the first. This can occur in two cases. First, since the magnitude of the first term increases with σr, while that of the second term is proportional to μr, then S is close to an estimate of the deviation of rewards when σr is relatively high with respect to μr.
Fig 2 shows simulations of the model for different reward mean and standard deviations of rewards and illustrates changes in synaptic weights as learning progresses. The simulations shown in different rows correspond to mean reward being positive, equal to 0, and negative, respectively. Note that the difference between G and N always approaches a value proportional to the expected reward. The simulations shown in different columns correspond to progressively higher standard deviation of reward. When μr = 0, the value that G and N approach increases linearly with σr. By contrast, when μr is higher, the encoding of reward uncertainty is less precise. For example, in the top row of Fig 2 we observe that the values of synaptic weights change very little as σr increases from 0 to 2. The increase in weights is slightly higher as σr increases from 2 to 4. Nevertheless, Fig 2 shows that increasing reward uncertainty still results in higher values of both G and N. Note that in each row, the larger the reward uncertainty, the larger G and N.
Second, the second term in Eq (17) decreases with the ratio of parameters β α. Thus the lower β is relative to α, the closer S* is to a linear function of the deviation of rewards. This property is illustrated in Fig 3, which plots S as a function of the standard deviation of rewards for different values of β. It is evident in the figure that, on average, S is a monotonic function of σr. Hence, it is worth noting that although S is an estimate of reward uncertainty, it is possible for the neural system to obtain a closer estimate by learning the function mentioned above and thus decode the estimate of reward deviation from S (i.e., correct the biases of S in estimating σr). However, this function has a flat region for low σr, so that the model’s estimate of the reward deviation will not be precise in that range of σr. For example, one can observe in Fig 3 that when β = α the value of S ≈ 0.5 arises for a wide range of σr, so knowing that S = 0.5 we cannot accurately tell the value of σr. The size of the region where σr is not well estimated can be reduced by decreasing β relative to α. Nevertheless, Fig 3 illustrates that there is a trade-off: Lower β α results in a higher magnitude of weights, and thus higher metabolic cost, and lower β also slows learning (see [28] for details).
Let us now consider how the mean and spread of a reward distribution, learned by the model described above, can be used by BG in action selection. In the model the tendency to choose or avoid risky options is controlled by the tonic level of DA. Before giving mathematical justification for this property, let us first provide an intuition for it.
Fig 4 illustrates states of a network choosing between two options, one safe and the other risky, represented by neurons shown in blue and orange, respectively. In the figure, the strength of cortico-striatal connections is denoted by the thickness of the arrows. Thus both options are associated with positive mean reward (as the connections Gi are thicker than Ni), but the orange option has higher estimated spread of rewards (as the orange connections are thicker than the blue ones). DA is known to activate the D1 or Go neurons and inhibit D2 or NoGo neurons, which is represented in Fig 4 by green arrows and lines ending with circles. The top panel illustrates a situation when the tonic DA level is high. In this case the NoGo neurons are suppressed (indicated by bleak color) and the choice is driven by the activity of the Go neurons. Thus with high DA, the more risky, orange option is more likely to be chosen, as G2 > G1. By contrast, with low levels of DA, the Go neurons are inhibited (bottom panel of Fig 4), and the choice is driven by NoGo neurons. Thus with low DA, the risky option is inhibited (as N2 > N1), and the model is more likely to select the safe option.
The above example illustrates that the model has the tendency to choose more risky options when the level of DA is high, and safer options otherwise. Let us now show this property formally. The choice rule of Eq (6) can be rewritten to make the effect of the mean and deviation of reward visible. To do so, we first write Gi and Ni in terms of Qi and Si (defined in Eqs (7) and (14)):
G i ( t ) = 1 2 S i ( t ) + Q i ( t ) (18) N i ( t ) = 1 2 S i ( t ) - Q i ( t ) (19)
Substituting the above into Eq (6) we obtain:
P i ( t ) = exp 1 2 U i ∑ k exp 1 2 U k , where U i = ( a + b ) Q i ( t ) - ( b - a ) S i ( t ) (20)
In the choice rule above, the probability of choice depends on a utility function Ui that is a linear combination of mean reward and the deviation of reward (cf. [29, 30]). By increasing b relative to a in the above choice rule, one can explicitly control how choice probability is affected by the deviation of rewards. In particular, when b > a, the uncertainty of rewards reduces the probability of selecting the corresponding action, resulting in risk aversion. By contrast, setting b < a increases the probability of choosing actions with uncertain rewards, resulting in risk seeking.
Recall that parameters a and b describe in the OpAL model [15] to what extent D1 and D2 neurons contribute to determining choice. Since high levels of DA activate the direct pathway and suppress the indirect pathway, increasing the tonic level of DA will correspond in the model to increasing a and decreasing b, which according to the analysis above would result in more risk-seeking behavior. Thus such modulation provides a mean by which an organism can control whether the action selection should be risk-averse or risk-seeking. The above analysis explains why a tendency for gambling in Parkinson’s patients [12, 31] may arise from increasing the level of DA by medications or from deep brain stimulation of subthalamic nucleus (which would also weaken the indirect pathway so would correspond to lowering b).
The presented model accounts for the effect of pharmacological manipulations affecting dopaminergic receptors on risk aversion in reinforcement learning tasks. In a particularly comprehensive study [32], rats were trained to choose between 2 levers: pressing one of them resulted in certain delivery of a single food pellet, while pressing another could result either in delivery of 4 pellets or none. The probability of receiving the large reward after the selection of the risky lever was varied across conditions. After the rats were well-trained in the task, they were injected with different drugs, and changes in the fraction of risky choices made were measured. An overall increased tendency to choose the risky option was observed either after injection of D1 agonist or D2 agonist, as shown in Fig 5. Furthermore, the injection of D1 antagonist or D2 antagonist decreased the tendency to choose the more risky option [32].
The fraction of risky choices made in simulations by the AU model is shown by curves in Fig 5. In the simulations, the parameters controlling learning were fixed to standard values (α = β = 0.1), and only the parameters controlling choice (a and b) were fit to the data. Parameters a and b were fit separately to the data in each panel of Fig 5, as each panel was obtained from a different group of rats. While fitting the model to the data from D1 receptor manipulations, it was assumed that a differed between control and drug conditions, while b did not change. Thus three parameters were fit: acontrol, adrug, and b. We did not enforce any relationship between acontrol and adrug, but as we will explain below, the estimated parameters followed the relationship expected from the known effects of drugs. Analogously, while fitting the model to the data from D2 receptor manipulations, a, bcontrol, and bdrug were fit. For each panel, the values of the three parameters were found that minimized the sum of squared errors between the fraction of risky choices made by the animals and the model in the 8 conditions (4 probabilities of large rewards on and off the drug). The parameters were found using the simplex algorithm [33] implemented in Matlab (function fminsearch). The search was repeated 10 times with different random initial parameter values sampled from the range [0, 3].
The model reproduced the fractions of risky choices made by the animals relatively well. Importantly, the overall direction of changes in risky choices and estimated parameters is consistent with the pattern in the data. In particular, in the top panels of Fig 5, the fraction of risky choices is higher in the simulation of the agonist conditions. Furthermore, in the top left panel, estimated parameters satisfied adrug > acontrol (acontrol = 1.71, adrug = 3.13, b = 0.59), which is consistent with the excitatory effect of DA on D1 receptors, while in the top right panel, the estimated parameters satisfied bdrug < bcontrol (a = 2.72, bcontrol = 1.86, bdrug = 0.39), consistent with the inhibitory effect of DA on D2 receptors. Thus the choice behavior may become more risky due to activation of either D1 or D2 receptors, as activation of either of them decreases b − a, which reduces risk aversion in Eq 20. Analogously in the bottom panels of Fig 5, the fraction of risky choices is lower in the simulated condition with antagonists, and estimated parameters satisfy adrug < acontrol for the bottom left (acontrol = 2.67, adrug = 0.86, b = 1.04) and bdrug > bcontrol for the bottom right panels (a = 1.95, bcontrol = 0.04, bdrug = 2.16).
It is worth noting in the bottom left panel of Fig 5 that the model reproduces the cross-over of the two curves. It occurs in the simulations because as a is reduced (corresponding to the effect of D1 antagonist), the choice in the model becomes more random (recall from the Models section that a and b also control how deterministic the choice is), so that the fraction of risky choices is closer to 50%. In this task, choosing the risky lever gave higher expected reward in the 100% and 50% conditions while choosing the safe lever had higher mean reward in the 12.5% condition, and the model simulated with higher a in the bottom left panel of Fig 5 exploited the options with higher expected rewards more.
The AU model assumes particular rules for updating striatal synaptic weights, and here we consider whether these rules are consistent with the existing data concerning synaptic plasticity in the striatum. For a synaptic plasticity rule to be plausible, the change in a synaptic weight needs to depend only on the information that can be sensed by a synapse, i.e., the activity of pre-synaptic and post-synaptic neurons, the levels of neuromodulators released in the vicinity of the synapse, and the synaptic weight itself. Eqs (8) and (9) describe the change in synaptic weights between the neurons encoding current context and those encoding current movement, i.e., they describe changes in synapses between co-active neurons. This change includes two terms, which are the reward prediction error and decay. As mentioned earlier, a plethora of evidence suggests that reward prediction error (r(t) − Q(t)) is encoded in phasic changes in DA concentration, which is released in striatum.
The proposed weight update rules are consistent with the pattern of synaptic plasticity modulation by DA [34]. It has been observed experimentally that the activation of cortical neurons followed by striatal D1 neurons strengthens the synapses of D1 neurons when the DA level is elevated, and weakens these synapses when the DA level is reduced (Figs 3F and 2E in [34]). Such changes are consistent with Eq (8), because for positive prediction error, the prediction error term will dominate, so G will increase. By contrast, if the prediction error is negative, |r(t) − Q(t)|+ will be equal to 0, and the decay term will dominate, so G will decrease. Conversely, the activation of cortical neurons followed by striatal D2 neurons weakens the synapses of D2 neurons when the DA level is elevated, and strengthens the synapses of D2 neurons when DA level is reduced (Figs 1H and 3B in [34]). Such changes are consistent with Eq (9) for analogous reasons.
A critical property of the learning rules allowing encoding reward uncertainty in G(t) + N(t) is the asymmetry in how synaptic weights change for positive and negative reward prediction error. In particular, in the AU model, the change in G is only proportional to the reward prediction error if the error is positive, but not if the error is negative (analogous asymmetry holds for N). It is easy to check that if such asymmetry were not present (i.e., nonlinear functions of predictions errors were removed from Eqs (8) and (9)), then G(t) + N(t) would no longer encode the spread of reward distribution.
Such asymmetry may arise in striatal synapses from the observed differences in the affinity of DA receptors, such that a higher DA concentration is necessary to activate D1 receptors than D2 receptors [35]. Fig 6 shows how the probability of D1 and D2 receptor activation depends on DA concentration in a biophysically realistic model of DA release [36]. Simulation of that model based on activity of DA neurons in vivo [37] suggested that the baseline DA level in striatum is in a sensitive range of both D1 and D2 receptors (as illustrated by the dashed line in Fig 6). Due to the arrangement shown in Fig 6, an increase in DA level has a larger effect on the activation of D1, while a decrease in DA has a larger effect on D2 receptors.
According to Fig 6, the decrease in DA level may still have some small effect on the binding probability of D1 receptors (analogously the increase in DA may have a small effect on D2 receptors). Hence the complete lack of effect of a decrease (increase) in DA level on D1 (D2) neurons’ plasticity may seem inconsistent with the above analysis. Nevertheless below we show that for learning reward uncertainty, it is sufficient that there exist an asymmetry in the dopaminergic effects on the receptors, i.e., that the increase in DA level affect plasticity of D1 neurons more than D2 neurons (and the opposite for a decrease in DA level).
The Equations describing the AU model can be generalized to include more complex functions of reward prediction error:
G ( t + 1 ) = G ( t ) + α r ( t ) - Q ( t ) + - ϵ r ( t ) - Q ( t ) - - β G ( t ) (21) N ( t + 1 ) = N ( t ) + α r ( t ) - Q ( t ) - - ϵ r ( t ) - Q ( t ) + - β N ( t ) (22)
where ϵ is a constant such that ϵ < 1. As synaptic weights cannot be negative, whenever G(t+1) or N(t+1) computed from the above equations is negative, it is set to 0. A potential advantage of using such functions of prediction error is that after each feedback iteration, they drive changes in both G and N, and thus potentially result in faster learning. When ϵ = 0, the above model reduces to the AU model.
We now show that with these functions, the model can still encode expected reward and reward uncertainty. Subtracting the above two equations gives:
Q ( t + 1 ) = Q ( t ) + α ( 1 + ϵ ) r ( t ) - Q ( t ) - β Q ( t ) (23)
Hence at the stochastic fixed point:
Q * = α ( 1 + ϵ ) α ( 1 + ϵ ) + β E [ r ] (24)
Thus the differences in the synaptic weights of D1 and D2 neurons encode scaled relative values of actions, which are also sufficient to choose the action with the highest value. Similarly adding Eqs (21) and (22) we obtain:
S ( t + 1 ) = S ( t ) + α ( 1 - ϵ ) r ( t ) - Q ( t ) - β S ( t ) (25)
Hence at the stochastic fixed point:
S * = α ( 1 - ϵ ) β E r - Q * (26)
Using the analysis applied earlier to the AU model, we see that the sum of the weights of D1 and D2 neurons encodes a scaled version of deviation of the reward, under analogous conditions to those for the AU model (i.e., σr is relatively high with respect to μr, or β is relatively small with respect to α(1 + ϵ)). However, when ϵ > 0, the weights G(t+1) or N(t+1) computed from Eqs (21) and (22) may become negative, but negative synaptic weights are not allowed in the model, so the calculations of the fixed points above are only valid for ϵ sufficiently small so that G(t+1) and N(t+1) are not negative.
To illustrate how this generalized AU model encodes reward uncertainty, the left panel in Fig 7 shows the results of simulations in the same setting as in Fig 3, but with a fixed value of β = 0.1, for different values of parameter ϵ. The figure shows that when ϵ = 0.5, the model also encodes reward uncertainty, but the encoding is less accurate than for ϵ = 0. In particular, when S is equal to a certain value, we can infer σr more precisely from the left panel in Fig 7 for ϵ = 0, as the range of σr resulting in the certain value of S is narrower for ϵ = 0 (e.g., S = 0.75 for σ ∈ [0.6, 1]) than for ϵ = 0.5 (e.g., S = 0.75 for σ ∈ [1, 2]).
In this section we show that the actor-critic model after small extension can learn both the mean and spread of rewards associated with actions. The model uses the same rule for the update of the critic (Eq (2)), and the plasticity of synapses of D1 and D2 neurons is described by equations similar to those for the AU model, but in which the prediction error is based on the reward estimated by the critic:
G i ( t + 1 ) = G i ( t ) + α r ( t ) - V ( t ) + - α G i ( t ) (27) N i ( t + 1 ) = N i ( t ) + α r ( t ) - V ( t ) - - α N i ( t ) (28)
For simplicity, in the above equations we set the decay constant β = α, which will also allow relating the model to advantage learning [39, 40]. We will refer to a model with the actor described by the above equations, with the critic by the standard Rescorla-Wagner rule of Eq (2), and with the OpAL choice rule of Eq (6), as the Actor-Critic learning Uncertainty (ACU). We now show that the ACU model estimates both mean and spread of rewards associated with action i, which we denote by μi and σi, respectively.
To see that the mean rewards are encoded in the difference between Gi and Ni, we subtract the above equations, and using Eq (7), we obtain:
Q i ( t + 1 ) = Q i ( t ) + α r ( t ) - V ( t ) - α Q i ( t ) (29)
This update rule differs from that of the original actor-critic model of Eq (3) in that it includes a decay term, and the rule is known as advantage learning [39, 40] (for reasons that will become apparent below). Let us now find the value the vicinity of which Qi approaches, by noting that at the stochastic fixed point the following condition must hold:
E α r - V * - α Q i * = 0 (30)
Rearranging the terms in the above equation, we see:
Q i * = μ i - V * (31)
Namely, Qi at the stochastic fixed point is equal to the expected reward for action i relative to the overall average reward in the current state (this quantity has been termed the advantage of action i). Note that knowing the relative values of the actions available in a given state is sufficient for selecting the action with the highest value. The value of the state V* is equal to the average value of all actions weighted by how frequently they are selected:
V * = ∑ i P i * μ i (32)
In this model, the sum of G i ( t ) and N i ( t ) also approximates reward uncertainty. Adding Eqs (27) and (28) we obtain:
S i ( t + 1 ) = S i ( t ) + α r ( t ) - V ( t ) - α S i ( t ) (33)
At the stochastic fixed point, the expected change in the sum of weights should be equal to 0, hence:
E α r - V * - α S i * = 0 (34)
Rearranging terms, we see that the sum of weights Gi and Ni at the fixed point is:
S i * = E r - V * (35)
The above equation implies that when V* = μi, the sum of Gi and Ni is equal to the deviation of the reward from the mean. We now consider three situations when V* is close to μi.
First, when only one action is available, and chosen on all trials, then V* = μ1, and hence S 1 * ∼ σ 1. This property is illustrated in the right panel of Fig 7, where black dots show the uncertainty estimated by the ACU model in simulations with a single action. Note that S is proportional to reward uncertainty for the entire range of σr, so with a single action, the ACU model can accurately encode uncertainty for a wider range of σr than the AU model (cf. black points in left and right panels of Fig 7).
Second, when a few actions i ∈ I have similar mean rewards, while other actions j ∈ J give much lower rewards, then Pj ∈ J are close to 0. In this case, V* is equal to a weighted average of μi ∈ I, but since we assumed that all μi ∈ I are similar, then V* is close to μi for i ∈ I. Hence the ACU model estimates well the spread of reward distribution for actions with the highest mean reward, i.e., those most frequently selected. It may not estimate the spread of other actions, but this does not matter, as these actions are typically not selected anyway.
Different rows in Fig 8 show simulations of the ACU model for different reward distributions and illustrate changes in synaptic weights as learning progresses. In the first simulation, the two actions have the same mean reward, and it can be seen in the top row that the value V converges to the expected reward. For each action, Gi and Ni converge to values equal to each other, because the ACU model encodes in Gi − Ni the relative value of actions which are equal to 0 here. In the simulation, the second action has uncertainty twice as high as the first one, and indeed one can see in the top row of Fig 8 that G2 + N2 converges to a value twice as high as G1 + N1.
In the simulation illustrated in the bottom row of Fig 8, the first action has a smaller expected reward. The model learns to select the second action on a great majority of trials, which results in the expected reward V converging towards the mean reward of the second action. The model estimates well the deviation of rewards associated with the second action—note that G2 + N2 is similar in both rows of Fig 8. Finally, the model does not estimate well the deviation of reward of the first action, but this does not matter, as this action is very rarely selected.
Third, the ACU model can still estimate reward uncertainty for actions with lower mean rewards than other actions available, if the uncertainty is sufficiently large. To understand this property, it is helpful to rewrite Eq 35 as:
S i * = E ( r - μ i ) + ( μ i - V * ) (36)
When σi is sufficiently larger than |μi − V*|, the first term in the above equation will dominate over the second, and Si will be more closely proportional to σi.
In summary, the AU and ACU models differ in the conditions under which their ability to estimate reward uncertainty is limited. The AU model does not precisely estimate the reward uncertainty in situations where the standard deviation of rewards is small relative to their mean. The ACU model has a limited ability to estimate uncertainty only in a subset of these situations, i.e., when the reward uncertainty is small and additionally the mean value of the action is substantially lower than for other actions available in the corresponding state.
Finally, it is worth mentioning that the learning rule of the ACU model can be generalized as described in the previous subsection, such that the weights of the actor are modified according to Eqs 21 and 22 but with Q replaced by V. The grey dots in the right panel of Fig 7 show that the uncertainty estimated by such a generalized ACU model is still proportional to the true variability of rewards but is encoded less precisely than in the original ACU model. Furthermore, a simulation of the ACU model analogous to that shown in Fig 5 produced qualitatively similar behavior as the AU model; thus, an increased tendency to take risky options with a high level of DA is a general property of a class of models encoding reward uncertainty in G + N.
We also investigated the behavior of the OpAL model [15] in the presence of reward uncertainty. Fig 9 shows simulations of the OpAL model in the same tasks used for the ACU model in Fig 8. Top rows of Fig 9 show simulations of a task in which the two actions have the same mean reward but differ in reward deviation. In the initial trials, in which the reward prediction error is positive, Gi increase exponentially. The exponential increase arises due to the multiplication of prediction error by G or N in Eqs 4 and 5, which results in a rate of weight changes that is proportional to the weights themselves. Once the reward prediction becomes equal to 0 on average, the weights start to decay towards 0. The weights have a stochastic fixed point at Gi = Ni = 0 in the OpAL model, because when Gi = Ni = 0, there are no changes in weights according to Eqs 4 and 5. In the task simulated in the top panel of Fig 9, this fixed point was attractive, and all weights of the actor eventually approached 0. It is interesting that this decay was faster for the option with higher uncertainty, as for this option the larger fluctuations in the reward prediction error drove the weights to the fixed point faster. In the task simulated in the bottom panel of Fig 9, this fixed point was attractive only for the action with the higher value, while for the other action, Ni increased with time.
It is evident from Fig 9 that the OpAL model does not encode reward uncertainty in the weights close to convergence, and the dynamics of weight changes is much more volatile than in the ACU model (note that the range of vertical axes in Fig 9 is two orders of magnitude higher than in Fig 8). Furthermore, when two actions have equal mean reward, as in the top panels of Fig 9, after extensive training, all weights Gi and Ni converge to 0, so the probability of choosing a more risky option becomes exactly 0.5, according to Eq 6, irrespective of the values of parameters a and b. Hence in this case, the probability of a risky choice predicted by the OpAL model is not dependent on the level of DA.
The OpAL model is able to capture the effects of dopaminergic medications seen in a series of experiments [14, 41, 42], which as we will see below, are challenging for the AU and ACU models. These experiments were designed to test the effects of DA on learning from positive and negative feedback, but in these studies the feedback uncertainty also varied between choice options. During these experiments the participants were presented with Japanese characters, were asked to choose one them, and subsequently received feedback indicating whether their choice was correct. For clarity, let us consider a simplified version of the task. Assume that during the training phase, the participant is presented on each trial with 3 letters which we will refer to as A, B and C. The probability of obtaining “Correct” feedback after selecting each of the 3 options is 0.8, 0.2 and 0.5 respectively. After the training, the participant is presented with a choice between A and C, or with a choice between B and C. The fraction of A vs. C trials in which the participant chooses A has been interpreted as a measure of learning from positive feedback (as stimulus A was associated with the highest probability of “Correct” feedback). Conversely, the fraction of B vs. C trials in which the participant does not choose B has been interpreted as a measure of learning from negative feedback (as stimulus B was associated with the highest probability of “Incorrect” feedback). It has been observed that Parkinson’s patients on dopaminergic medications exhibit higher accuracy in choosing A than in avoiding B, while the opposite pattern is present off medications [14]. Furthermore, it has been suggested that this effect is dependent on the medication state during testing rather than during encoding [42].
The OpAL model is able to replicate these effects [15]. While simulating learning in this task, we assumed that the model receives a reward of r = 1 when “Correct” feedback is given, and no reward r = 0 after “Incorrect” feedback. The top left panel in Fig 10 shows the weights learned by the OpAL model. As expected, Gi increase with the probability of reward, while Ni decrease. Importantly, the relationship between weights and reward probability is non-linear. This non-linearity arises from the multiplication of prediction error by Gi or Ni in Eqs 4 and 5, which as mentioned above, results in an exponential growth of the weights and thus magnification of weights with high values. The bottom right panel in Fig 10 illustrates how the values of the weights affect behavior during test. In the simulated on medication condition, the choice is primarily affected by weights Gi (Eq 6). Thus the accuracies in choosing A and avoiding B depend on |GA − GC| and |GB − GC|, respectively. Since |GA − GC|> |GB − GC| in the top left panel, the probability of choosing A is higher than the probability of avoiding B on medications in the bottom left panel. In the simulated off medication condition, the choice is primarily affected by weights Ni, and hence the model is better at avoiding B than choosing A for analogous reasons. The choice pattern in the bottom left panel of Fig 10 is qualitatively consistent with that observed in experimental studies [14, 41, 42].
The top panel in the middle column of Fig 10 shows the weights learned by the AU model. Here also, Gi increase with reward probability, while Ni decrease. However, in the AU model the sum of weights Gi + Ni is highest for option C, which gives reward on 50% of trials and thus has highest reward variance. Consequently, the relationships between weights and reward probability are concave for the AU model, rather than convex as they were for the OpAL model. This results in the opposite effect of DA on choosing A and avoiding B relative to the OpAL model (cf. left and middle panels in the bottom row of Fig 10).
The right panels of Fig 10 illustrate that the behavior of the ACU model is qualitatively similar to that of the AU model. However, the predicted effect of medications on choice probability in ACU is smaller than in AU, because the relationships between weights and reward probability are more linear for ACU. This occurs because ACU estimates the deviation of reward from the mean across all trials (Eq 35) rather than from the mean reward for a given option, as in AU.
The OpAL model also described the dependence of learning rates α for Gi and Ni on the level of DA [15]. Simulations of the AU and ACU models indicate that increasing the learning rate for Gi (or Ni) scales up the learned values of Gi (or Ni) but does not change the convexity/concavity of the relationship between weights and reward probability, and hence does not change qualitatively the predicted effects of DA during testing on the probability of choosing A and avoiding B.
In summary, the simulations of the AU and ACU models produced qualitatively different patterns of effects of dopaminergic medications on choosing A and avoiding B than observed experimentally [14, 41, 42]. A critical feature of the OpAL model that allows it to capture the experimentally observed effects is the multiplication of prediction error by G or N in Eqs 4 and 5, but it is this very property that also caused unrealistically volatile weight changes in simulations of Fig 9.
It is interesting to ask under what assumptions the pattern of weights in the top left panel of Fig 10 (that allows reproducing the effects of medications on choosing A and avoiding B) could be obtained in a model learning reward uncertainty. In our simulations we assumed that “Correct” and “Incorrect” feedback were mapped on rewards of 1 and 0. However, it is unclear if the brain simply maps abstract feedback on the reward. It is possible that instead the brain infers that option C is unpredictable and does not engage in learning about it, which would result in relatively low GC and NC, as in the top left panel of Fig 10. This interpretation together with the AU (or ACU) model predicts that if an actual (e.g., monetary) reward is given as feedback, the effect of dopaminergic medications on choosing A and avoiding B should reverse (or be very small). This interpretation is consistent with a result of experiments employing a modified version of the Japanese letter task with more salient feedback, i.e., smiling and sad faces, in which no effect of medications was found [43]. However, to fully test this interpretation, further studies are needed that could for example use explicit monetary reward.
In this paper we presented a class of models that can learn both the mean reward and reward uncertainty. The models describe how BG can control the influence of risk on choices and choose actions that not only maximize expected rewards but also minimize risks. Below we relate the models to experimental data, state further predictions, and discuss relationships with other computational models.
We discuss here the relationships between predictions of the models and experimental data, including behavior and neural activity. Since in this paper we presented several models, it will be important to distinguish in the future which of them provide the best description of learning uncertainty in the basal ganglia. To differentiate between predictions specific to individual models and common to other models, we will use the term “the models” to refer to a class including all models introduced in this paper.
We already demonstrated in the Results section that the models account for the effect of pharmacological manipulations affecting dopaminergic receptors on risk aversion in reinforcement learning tasks in rats. The studies investigating the effect of DA on human decisions involving risks use two types of paradigms: one in which the mean and spread of rewards associated with choice options are explicitly described to the participant before each decision, and one in which they are gradually learned from feedback. Since human behavior is very different in these paradigms [44], and the models assume that the mean and deviation of rewards are learned in cortico-striatal synapses, below we only focus on studies involving learning from experience. The most commonly used paradigm in such tasks is the Iowa gambling task in which participants choose between decks of cards differing in reward variance. In agreement with the models, Parkinson’s patients receiving dopaminergic medications choose the risky decks more frequently than healthy controls, but this effect is not present in patients that have not been put on medications yet [45], or who stopped receiving medications [46].
The models introduced in this paper do not describe behavior in decision tasks in which information about risks associated with different options is explicitly presented before each trial. It is likely that processing information about uncertainty in such tasks involves different neural mechanisms and circuits than those learning about reward uncertainty over many trials.
The models are also consistent with the results of a recent study showing that optogenetic activation of striatal D2 neurons decreases the probability of choosing options with high reward variance [47]. Optogenetic activation of D2 neurons corresponds to a scenario illustrated in the bottom panel of Fig 4, where the choice is primarily driven by D2 neurons, and thus the risky option is inhibited.
The AU and ACU models differ in the predicted activity of DA neurons when the reward exactly matches the expected reward in tasks where only one action is available. In the ACU model, DA response is assumed to carry (r − V) where V* = E[r], so when r = E[r], DA neurons should not change their firing rate. By contrast, in the AU model the DA release is assumed to encode (r − Q) where Q* < E[r] (see Eq 13), so when r = E[r], DA neurons should increase their firing rate above baseline. Experimentally observed DA responses after expected rewards differed between experimental studies. For example, DA neurons were found to maintain their activity in classical conditioning in some studies [6, 7], while an increase was observed in others [48, 49]. Thus, more research is necessary to establish factors determining DA response to expected reward.
The AU model predicts that learned synaptic weights in BG are insensitive to small standard deviations of reward; thus, it predicts that an individual’s choices are not affected by small enough uncertainty in reinforcement learning tasks. By contrast, the ACU model predicts that biases in estimation of reward uncertainty should only be present for actions with mean rewards much lower than those of other actions.
The models predict that overall activity in striatum should be higher during choice between options with high reward variance than during choice between options with lower reward variance but similar mean, because in the models the spread of rewards is encoded in Gi + Ni, so higher reward variance should increase the activity of both D1 and D2 striatal neurons. This prediction could be easily tested using functional MRI.
The models predict that synaptic plasticity will depend on the current value of the weight itself (i.e., Gi or Ni), because the weight update rules include decay terms proportional to the weights themselves. Thus the models predict that the stronger the weight of a synaptic connection, the higher the amplitude of induced long-term depression. Such dependence of plasticity on the value of weights has been observed in neocortex [50], and it would be very interesting to see if it is also present in cortico-striatal synapses.
In addition to the models presented in this paper, reward uncertainty can be learned by a wide family of models in which the decay terms are proportional to the estimated uncertainty, and these models were analyzed in [28]. The models in this family can learn reward uncertainty even for small deviations. However, to implement such learning rules, the information about the uncertainty would need to be provided to a synapse, e.g., by a second neuromodulator. The models in this family predict that the release of this neuromodulator would need to be dependent on uncertainty and promote long-term depression of cortico-striatal synapses. Three different neuromodulators have been proposed to encode information about estimated (or expected) reward uncertainty: tonic DA [51], acetylcholine [52], and serotonin [30]. All of these neuromodulators have been shown to affect risky decisions [12, 53–55]. However, we have not found support in existing experimental data for predictions of our models employing multiple neuromodulators, hence we did not include them in this paper.
In previous reinforcement learning models that described learning about uncertainty [30, 56], the estimate of reward variance was updated on each trial proportionally to “variance prediction error”, which is equal to the difference between the square of reward prediction error and the current estimate of variance. An interesting model describing how such learning could be implemented in BG [57] suggested that the variance of rewards is encoded in striatal neurons co-expressing D1 and D2 receptors. This model assumed that such neurons could increase their weights both when the prediction error is highly positive (like D1 neurons) and when it is strongly negative (like D2 neurons). However, the neurons co-expressing D1 and D2 receptors form only a small proportion of striatal neurons [58], and the models we propose describe learning of reward deviation in the great majority of striatal projection neurons that express either D1 or D2 receptors.
An interesting reinforcement learning model has also been proposed in which choosing risky options can be avoided without explicitly learning the spread of reward distributions for different options [59]. In this model, the weight update rules are modified such that Qi is decreased when action i leads to a reward with higher variance. This model is efficient when the desired level of risk aversion is known and fixed before the learning starts, but unlike the models presented in this paper, it does not allow the trained system to be easily switched from risk aversion to more neutral or risk seeking behavior.
Reward uncertainty is also likely to be estimated in the cortex. A particularly interesting model [60] describes how the variance of any feature of the stimulus (including reward) can be estimated in a neural circuit with organization similar to that of the neocortex, and it has been shown how this learning about variance can be implemented with local Hebbian plasticity [61]. It is highly likely that the mechanisms of learning uncertainty in neocortex and striatum can operate in parallel. Furthermore, these two structures may estimate complementary measures of dispersion: the cortical model [60, 61] estimates variance, while the models presented here estimate the mean absolute deviation (which is less affected by outliers).
In this paper we focused on one particular type of uncertainty associated with variable rewards in a stationary environment, which is typically called “expected uncertainty” [52]. But there is also another type of uncertainty connected with rapid changes (or volatility) of mean reward, referred to as “unexpected uncertainty” [52]. It is likely that there are complementary neural mechanisms which estimate unexpected uncertainty. For example, it has been proposed that striatal cholinergic tonically active interneurons detect changes in reward contingency and increase the learning rate following such changes [62]. Areas beyond BG can also be involved in this process, as the activity in other brain regions has been shown to track reward volatility [63] and volatility prediction errors [64].
Finally, let us discuss the relationship of the ACU model to advantage learning [39, 40]. As mentioned in the Results section, the ACU model estimates the mean reward using the advantage learning rule; thus, the ACU model also provides a description of how this abstract rule may be implemented in the BG circuit. The advantage model was originally introduced to reconcile reinforcement learning models with animals’ innate tendency to approach highly rewarding stimuli [39, 40]. The central feature of the advantage model (also inherited by the ACU model) is that as learning progresses, the value V represented by the critic approaches the value of the best action in the current state, while the advantage Qi of this action approaches 0. This property describes a transition from an instrumental action selection to a stimulus-response habit, as in the trained state the action selection is implemented in the advantage model by the innate tendency to approach high value states [39, 40].
The ACU model has an analogous property that in the absence of reward uncertainty, Gi decreases towards 0 as learning progresses. Selection under such circumstances is primarily driven in the ACU model by D2 neurons, as suboptimal actions have large Ni, and thus are inhibited. This agrees with a recent proposal of D2 neurons being critical for choosing among actions [65]. It would be possible to also include in the ACU model the tendency to approach high value states, by including additional terms in the softmax choice rule, as in [66].
In the advantage and ACU models, the actor encodes the mean reward relative to the overall reward in the current state (Eq (31)). So although the actor has the information necessary to choose which action is best in the current context, it does not know whether it is worth selecting it at all (e.g., whether any μi > 0). The information on whether it is worth making a movement in the given state (i.e., on the average value of actions chosen by the actor) is encoded in the critic. Thus the models suggest that patch neurons, which the critic has been mapped onto, should also be involved in movement initiation. It is intriguing that patch neurons project to the dopaminergic neurons [19], so one could ask whether they may communicate the information on the value of making a movement via dopaminergic neurons. This idea is consistent with DA controlling the vigor of movements [67].
|
10.1371/journal.pgen.1004838 | Genetic Control of Contagious Asexuality in the Pea Aphid | Although evolutionary transitions from sexual to asexual reproduction are frequent in eukaryotes, the genetic bases of such shifts toward asexuality remain largely unknown. We addressed this issue in an aphid species where both sexual and obligate asexual lineages coexist in natural populations. These sexual and asexual lineages may occasionally interbreed because some asexual lineages maintain a residual production of males potentially able to mate with the females produced by sexual lineages. Hence, this species is an ideal model to study the genetic basis of the loss of sexual reproduction with quantitative genetic and population genomic approaches. Our analysis of the co-segregation of ∼300 molecular markers and reproductive phenotype in experimental crosses pinpointed an X-linked region controlling obligate asexuality, this state of character being recessive. A population genetic analysis (>400-marker genome scan) on wild sexual and asexual genotypes from geographically distant populations under divergent selection for reproductive strategies detected a strong signature of divergent selection in the genomic region identified by the experimental crosses. These population genetic data confirm the implication of the candidate region in the control of reproductive mode in wild populations originating from 700 km apart. Patterns of genetic differentiation along chromosomes suggest bidirectional gene flow between populations with distinct reproductive modes, supporting contagious asexuality as a prevailing route to permanent parthenogenesis in pea aphids. This genetic system provides new insights into the mechanisms of coexistence of sexual and asexual aphid lineages.
| Asexual lineages occur in most groups of organisms and arise from loss of sex in sexual species. Yet, the genomic bases of these transitions remain largely unknown. Here, we combined quantitative genetic and population genomic approaches to unravel the genetic control of shifts towards permanent asexuality in the pea aphid, which conveniently shows coexisting sexual and asexual lineages. We identified one main genomic region responsible for this transition located on the X chromosome. Also, our population genetic data indicated substantial gene exchange between these reproductively distinct lineages, potentially leading to the conversion of some sexual lineages into asexual ones in a contagious manner. This genetic system provides new insights into the mechanisms of coexistence of sexual and asexual lineages.
| While sexuality is the dominant reproductive mode in metazoans, parthenogenesis - the development of an embryo from an unfertilized egg - occurs in most branches of the animal kingdom (e.g. molluscs, insects, crustaceans, nematodes, fish, reptiles) [e.g. 1,2,3]. Cyclical parthenogenesis (CP) represents a mixed reproductive mode with an alternation of sexual reproduction and parthenogenesis, and is reported in many animal species [4]. The loss of the sexual phase in CP species - leading to permanently parthenogenetic taxa - have been shown to arise from diverse mechanisms, including microbial infection, hybridization, contagion via pre-existing parthenogenetic lineages or spontaneous mutations [5]–[9]. Nevertheless, in case of contagious or mutational origin, the precise genomic regions responsible for the transitions to obligate parthenogenesis (OP) remain largely unknown, mostly because dissecting the genetic bases of that trait using recombination-based approaches is not possible in strictly asexual species. However several species show coexisting CP and OP lineages, with OP lineages often retaining a residual production of males. Such species offer ideal systems to decipher the heredity and therefore the genetic basis of the loss of sexual reproduction. In the rare cases where it has been explored, genetic control of this trait has been shown to be rather simple, with the involvement of one to four loci, depending on the studied organisms [10]–[15]. However, the precise location and underlying function of these genetic factors have not been elucidated.
The ancestral life-cycle of aphids is cyclical parthenogenesis [16], which consists in an alternation of sexual and asexual generations. In spring and summer, CP lineages produce asexual females through apomictic parthenogenesis. In autumn, asexual females give birth to males and sexual females in response to photoperiodic cues (note that CP lineages can also produce asexual females to some extent [e.g. 17]). Sexual females are strict clones of their asexual mothers, while one of the two X chromosomes is randomly lost to generate males [17]. Eggs produced by sexual females are the only frost-resistant stage in the aphid cycle. Hence, a CP life cycle is required to survive in regions with cold winters. In addition, many aphid species also encompass OP lineages which are characterized by an altered response to sex-inducing environmental cues as they produce only asexual females (although they often produce some males [18], [19]). These lineages are thus cold-sensitive because of their inability to lay eggs. Yet, OP lineages are favoured in regions with mild winters where they have a major demographic advantage over CP lineages [20], [21]. Accordingly, CP lineages dominate in cold areas and OP lineages in warmer regions, and both coexist in regions with fluctuating winter temperatures [18]–[20]. Because male production by OP lineages is difficult to prove in the wild, it has been demonstrated in a single study which also showed that these males actually contribute to sexual reproduction with CP lineages [22], [23]. While the switch from clonal to sexual reproduction in CP aphids is triggered by photoperiodic changes, the loss of sexual form production in OP aphids is genetically determined, changes in environmental conditions having little or no effect on their reproductive phenotype [10], [19].
Here, we combined QTL and genome scan approaches to decipher the genetic bases of reproductive mode variation in the pea aphid Acyrthosiphon pisum. This species conveniently shows CP lineages (here defined as those able to produce sexual females) and OP lineages (defined as those unable to produce sexual females), and these two types of lineages locally co-occur in regions with intermediate climate conditions [24]. These independent QTL and genome scan approaches outlined the same genomic region as controlling obligate parthenogenesis, this trait being recessive and determined by an X-linked locus. Our data also indicate that asexuality is transmitted in a contagious manner, leading to the conversion of sexual lineages into asexual ones.
We produced F1 crosses between males of an obligate parthenogenetic lineage (L21V1) and sexual females of two cyclically parthenogenetic lineages (JML06 and LSR1) (Fig. 1). Five F2 crosses (families 3 to 7) involving 6 F1 lineages were performed to obtain a genetic map and to locate QTL controlling the presence and proportion of sexual females by genotypes placed under sex-inducing conditions. A total of 305 microsatellite markers (out of 394) was successfully ordered on the genetic maps. These loci clustered in four linkage groups that correspond to the four chromosomes of the pea aphid [25]. 45 loci locate on the X chromosome (LG1 following notation in [26]), and 85, 135, and 40 on LG2, LG3 and LG4, respectively. Average map length (over males and females) was 113, 95, 79 and 59 cM for LG1, LG2, LG3 and LG4, respectively (Fig. 2). Of the 89 unmapped loci (out of 394), 51 were monomorphic in the 3-generation pedigree, five were homozygous in all F1 females, and 33 showed null alleles at high frequencies or inconsistent genotypes (presumably due to difficulties to score alleles).
By contrast with the 61 F1 progeny which all produced sexual females (hence were classified as CP) segregation of reproductive phenotype was observed among the five F2 families (Fig. 1, families 3 to 7). All five families (203 F2 genotypes) comprised a mixture of genotypes expressing either an OP (no sexual females produced at all) or a CP (sexual female production ranged from 22% to 77%) phenotype. The percentage of OP F2 ranged from 7% to 35% depending on families (Fig. 1, see also S1 Figure). Contrastingly, 97% of F2 lineages produced males: only 5 out of 35 OP lineages, and 2 out of 168 CP lineages did not produce males (S1 Figure). QTLs analyses on these five F2 families revealed one candidate genomic region located on the X chromosome (LG1) for the control of reproductive mode variation (measured as the proportion of sexual females or occurrence of sexual females), as evidenced by likelihood-ratio (LR) values for these traits above the LR thresholds corresponding to the null hypothesis of no QTL. The QTL for the proportion of sexual females produced locates at 38.0 cM on LG1 based on highest LR values. The 95% confidence interval [CI] for QTL position is 34.0–43.2 cM (Fig. 2). The QTL for presence/absence of sexual females locates at 37.6 cM on LG1 and the 95% CI is 34.8–43.6 cM (Fig. 2). We then accounted for the presence of a QTL at position ∼38 cM to test for a second QTL (see Methods). No significant support for a second QTL was found for either traits, as LR values along the four chromosomes were largely below the LR thresholds corresponding to 5% significance at the genome level.
We then focused on the genomic region pinpointed by the QTL analysis (∼38 cM on LG1) and looked at the alleles inherited by F2 individuals. In three F2 families (families 3, 4 and 6, Table 1), the 29 F2 lineages that expressed an OP phenotype had the same genotype as the OP lineage L21V1 (F0) at all markers located on the X-chromosome between T_128012_2_G (34.8 cM) and T_126075_3_Y (49 cM). For simplification, we refer to this multilocus genotype as “op1/op2” (Table 1, see also S1 Figure). Contrastingly, the 89 CP F2 individuals from families 3, 4 and 6 all possessed at least one allele inherited from their CP grandmothers in this genomic region (the four possible alleles from the two CP grandmothers are collectively referred to as “CP”). Hence these individuals were either op1/CP, op2/CP or CP/CP (Table 1, S1 Figure). In these three F2 families, the op1 allele was transmitted through the F1 fathers from the OP grandfather (L21V1, of genotype op1/op2). Since chromosomes in male pea aphids do not recombine [27], the entire X-chromosome of grandfather L21V1 that carries the op1 allele was transmitted to its grandchildren. Conversely, the op2 allele was inherited from the F1 mothers, which themselves inherited the whole op2-bearing X-chromosome from their OP grandfather. Recombination of the op2-bearing X-chromosome in the F1 mothers allowed reducing the region controlling reproductive modes between markers T_128012_2_G (34.8 cM) and T_126075_3_Y (49 cM) on the op2-bearing X chromosome (S1 Figure). Based on the results from QTL analyses, we performed two additional crosses. Here only a subset of individuals per cross were phenotyped (24 and 27, respectively), chosen according to their genotype at 8 microsatellite markers in the genomic region of interest. A F3 cross (cross 9, see Fig. 1 and Table 1) confirmed the location of the QTL and allowed further narrowing down its upper boundaries to marker 111865_3 [48.5 cM] (see S1 Figure). We finally crossed op1/CP2 females with op2/CP3 males in order to recombine the op1-bearing X-chromosome (cross 8, Table 1). All the 11 lineages that harboured the op1 allele in combination with the op2-bearing X-chromosome were OP, and recombination in the op1-bearing X-chromosome showed that the region controlling reproductive phenotype lies between markers 116879_10 (39.1 cM) and D_111865_3 (48.5 cM) on the op1-bearing X copy (see S1 Figure). These different crosses revealed that op1 and op2 alleles are recessive over CP alleles (since the 76 op1/CP and the 41 op2/CP lineages are CP, and the 12 op2/op2 and 43 op1/op2 lineages are OP, Table 1). Noteworthy we observed that op1/op1 genotypes can have either a CP (11 lineages) or an OP (6 lineages) phenotype (Table 1), suggesting that other genetic or environmental factors mitigate the control of reproductive phenotype in lineages op1/op1 at the major candidate locus.
A 436-marker genome scan performed on 109 individuals from wild populations collected in environments selecting for different reproductive modes (OP or CP) revealed four loci having excessive genetic differentiation (FCT) at the α = 0.01 threshold (ARLEQUIN 3.5 analysis, Table 2, S2 Figure). FCT between populations under selection for different reproductive modes ranged from 0.14 to 0.31 at these four outlier loci, while the median FCT value estimated over the 436 markers was 0.014 (average 0.025). Among these four outliers, T_111491_2 was also identified as outlier under balancing selection in the populations from CP-selecting environment (FST among CP populations was 0.0003, and He 0.56) when ARLEQUIN analyses were performed among populations assumed to share the same reproductive mode (Table 2). This locus was not successfully genotyped in the families so its genomic location remains unknown. Interestingly, the three other outliers co-locate on the X-chromosome and within the same genomic region identified with the experimental (QTL) approach (Fig. 2). Accordingly, FCT values along the genetic map of the four chromosomes show a clear peak of genetic differentiation in the QTL region (Fig. 2). In this region, expected heterozygosity in OP populations was lower than in CP populations (S3 Figure), while heterozygosity values of the three CP populations and the three OP populations were similar along other regions of the chromosomes.
We have shown here that a key ecological trait – the variation in reproductive mode – was controlled by one main genomic region in the pea aphid. This ∼9 cM-wide X-linked region was identified by two independent and complementary approaches: the co-segregation of molecular markers and phenotypes in experimental crosses and a large scale population genomic survey (genome scans). Interestingly, 100% of phenotypic variation was explained by the genotype at the candidate locus in five crosses (crosses 3, 4, 6, 8, 9, Table 1). In the two remaining families (crosses 5 and 7), this genomic region was also strongly associated with reproductive phenotype (as all six OP F2 were op1/op1 at this candidate region) but linkage was not absolute (as 11 op1/op1 individuals are nevertheless CP) (Table 1). Two hypotheses can be invoked for this lack of association in some F2 genotypes. First, an additional locus with minor effects might contribute to the control of reproductive mode variation, its contribution being only visible in individuals op1/op1 at the major locus (all 68 individuals from crosses 5 and 7 that are not op1/op1 are CP). A second hypothesis is that the production of sexual females depends on a threshold concentration of some unknown factor (e.g. transcript, protein, hormone). Under this assumption, minor environmental variation could have drastic effect on reproductive phenotype determination around the concentration threshold. We tested for the presence of a second QTL (first hypothesis), and found no statistical support for it. Yet, power to detect such an additional QTL was low (due to the small sample size of op1/op1 genotypes) so we cannot at the moment disentangle these two hypotheses. Nevertheless, the mostly single-locus recessive inheritance of obligate parthenogenesis in the pea aphid is in line with the few similar studies which showed that the transition from sexual to asexual reproduction is determined by a small number of loci [10]–[14].
Transitions from cyclical parthenogenesis (CP, i.e. the alternation of asexual and sexual generations) to obligate parthenogenesis (OP) in aphids probably occur through loss-of-function mutations leading to an inability of lineages to produce females in response to the environmental cues that normally trigger the sexual phase. Hence, any mutation (i.e. point mutation, indel or rearrangements) that disrupts the pathway leading to the production of sexual females might be responsible for this transition. In theory, these loss-of-function mutations could occur repeatedly in the same gene, or on different genes involved in the same molecular cascade, these genes being either neighbours or scattered on the genome. Herein, the OP grand-parent used for QTL mapping harbours two distinct alleles (op1 and op2) at the identified QTL and the phenotype of homozygotes op2/op2 and op1/op1 significantly differs (all 12 op2/op2 but only 6 of the 17 op1/op1 individuals are OP, test of proportion: p = 0.0016). This indicates that at least two independent mutations in the same region are involved in the loss of sexual reproduction. Remarkably, the genome scan demonstrates that the region identified by the QTL approach also shows signatures of divergent selection between populations under different selective regimes for reproductive mode. This indicates that the QTL identified with three laboratory clones is also involved in the control of reproductive mode in wild populations originating from a large-scale geographic area (populations were collected up to 700 km apart). These population genomic data give further insights into the transitions from CP to OP. In particular, the occurrence of outliers in the QTL region, combined with their low genetic diversity in OP- compared to CP-selecting environments, reveal that only one or a few mutations leading to the OP phenotype have reached high frequencies in OP-selecting environments (otherwise this genomic region would not have been identified as FST-outlier). Outside the candidate region, populations from CP- and OP-selecting environments are weakly differentiated and show highly correlated levels of genetic diversity along chromosomes, suggesting important gene flow.
The most likely scenario to explain these genomic patterns of differentiation involves bidirectional gene exchanges between CP and OP lineages: Let us consider that the rare males produced by OP lineages successfully mate with sexual females from CP (as it is the case in laboratory conditions and presumably into the wild), producing CP offspring heterozygous at the candidate region (op/CP). These heterozygous CP lineages may produce OP progenies (those homozygote for the op alleles) that would survive if they encounter mild winter environments. Some minimal amount of gene flow can maintain a low genetic differentiation between populations from OP- and CP-selecting environments at the scale of the genome since divergence for neutral loci at a migration drift equilibrium is prevented when Nm>1, N being the effective population size and m the migration rate [28], [29]. Such bi-directional gene flow between OP and CP lineages may occur in the geographical regions with intermediate winter conditions where both CP and OP lineages coexist [22], [30]. Another scenario to consider relies on unidirectional gene flow from CP to OP. Under the hypothesis that recessive op alleles are relatively frequent in CP populations, CP lineages will regularly produce new OP lineages (those homozygous at the op alleles). If such OP linages are generated frequently, low differentiation between populations from OP- and CP-selecting environments along the genome is expected, except at the candidate region. This scenario is however less parsimonious than the former. First, it requires very frequent production of OP lineages by CP ones in order to prevent genomic differentiation between these two compartments likely to result from the strong clonal fluctuations (due to neutral factors and/or selection) typical of asexual populations [31], [32]. Second, in absence of reciprocal gene flow from OP to CP lineages, positive selection on op alleles in CP populations should be invoked to maintain these alleles. Yet, we know that op alleles are associated with a cost in CP selecting environments (homozygous op/op individuals do not survive cold winters) and therefore their frequencies are expected to decrease under these conditions. Our data are thus best explained by bidirectional gene flow between populations of distinct reproductive modes and support the hypothesis of contagious asexuality in wild pea aphid populations.
Contagious asexuality has important consequences on the evolvability of the OP lineages. Indeed, the bi-directional gene flow between CP and OP lineages allows genomes and alleles evolved under an asexual regime to enter the “sexual” pool via the few males produced by OP clones. Once introgressed in a CP lineage, a genomic region evolved under an asexual regime will recombine, allowing the purging of deleterious alleles. Later, if some of the CP individuals produce OP progeny (those homozygous at the op/op genomic region), some of the alleles evolved under the asexual regime might then reintegrate an OP lineage. Hence, contagious asexuality has the potential to combine the beneficial effects of sex (purging of deleterious mutations and combination of beneficial mutations within the same genome [33], [34]) with the advantages of clonal reproduction that avoid the two-fold cost of sex [34] and can “freeze” a genome (avoiding recombination load) [35]. This genetic system could thus favour the regular emergence of well fit OP lineages, which would be so fit because they would reuse alleles that competed and evolved under an OP-selecting environment (during their long stay within OP lineages) and that would have been separated from linked deleterious mutations during their sojourn in CP lineages.
The physical size of the ∼9 cM candidate region, that represents ∼2.6% of the whole genome in term of recombination units (cM), is still unknown because scaffolds from the pea aphid genome sequence are not yet ordered on chromosomes [36]. Hence the exact number and nature of the genes that are comprised within the candidate region are not known. Nevertheless, already 66 genes encoding proteins have been identified in the three scaffolds covering part of the 9 cM genomic region of interest (S1 Table). It is too early to designate candidate genes responsible for the CP/OP phenotypes, mostly because half of them have no predicted functions. However, recent works on the genetic programs involved in the seasonal switch from clonal to sexual reproduction in CP lineages allow highlighting in the candidate region three predicted genes putatively involved in photoperiod perception and brain signalling (e.g. rhodopsin specific isomerase, insulin), two pathways identified as differentially expressed in aphids exposed to either clonal or sex induction regimes [37]. Two genes putatively involved in the melavonate pathway (farnesyl-pyrophosphate synthase like and hydroxymetharyglutaryl-CoA synthase) upstream of the juvenile hormone synthesis, which is known as being a key regulator of reproductive orientation in aphids [38], [39], also locate within the candidate region.
To conclude, here we combined population genomics and quantitative genetics to identify the genetic bases of a key trait for aphid adaptation to climate - the loss or maintenance of sexual reproduction. We found this trait to be controlled by one main genomic region located on the X chromosome. The widespread geographical distribution of a few alleles associated with obligate asexuality suggests that these alleles might be particularly advantageous for OP lineages, and might have outcompeted previously established op alleles, a hypothesis that deserves further investigations.
We crossed individuals from three genotypes (clones LSR1, L21V1, JML06) that present contrasted reproductive phenotypes. These three F0 lineages were chosen based on their ability to produce or not sexual females under standard sex-inducing conditions (i.e. short photoperiod with 12 h light) [37]. All aphids were reared on Vicia faba (broad bean) because it is a universal host for all known host races of the pea aphid species complex [40], [41] and also because this plant is easier to grow compared to Medicago ssp. LSR1 (collected on Medicago sativa in New-York, USA in 1998 and used for complete genome sequencing [36]) produces males (21%), sexual females (54%) plus some parthenogenetic females (25%) under standard sex-inducing conditions. Under the same inducing conditions, JML06 (sampled on Medicago lupulina in Jena, Germany in 2006) produces only sexual individuals (70% males and 30% sexual females). Contrastingly, L21V1 (sampled on M. sativa in Rennes, France in 2003) produces only parthenogenetic females (89%) and a few males (11%). LSR1 and JML06 are therefore classified as cyclical parthenogens (CP) and L21V1 as obligate parthenogen (OP). Crosses between males from the OP and (sexual) females from the CP lineages were performed. For this, one L4 larva from each of the three grandparent clones was moved to a new broad bean plant and transferred to a climatic chamber with a 12 h photoperiod (18°C) to trigger the production of sexual females and males in CP lineages (and males in the OP lineage) [37]. Then, for each lineage, three larvae of the next clonal generation were isolated on three different broad bean plants. Once the larvae reached adulthood and started to give birth to nymphs, groups of 10 larvae of the next generation were isolated on new broad bean plants until the asexual female stopped reproducing and died. The morph of each individual of this second clonal generation (i.e. male, sexual females, asexual females) was determined at adult stage based on morphological characters (males are slender than females, and the legs of sexual females are longer that those of asexuals). The few individuals that died before reaching the adult stage (hence before being sexed) were also counted. Then a total of 50 males from the clone L21V1 and 50 sexual females from the clone JML06 were put together on broad bean plants (Vicia faba) to generate a F1 family (cross 1: JML06 ♀×L21V1 ♂, Fig. 1). The 50 sexual females used in the cross are clonal. However, males consist of two different genotypes because they inherit randomly one of the two X copies from their asexual mother (approximately half of males are expected to bear the first X copy of their mother and the second half the other copy) [17]. A second F1 family was generated similarly by crossing 50 L21V1 males with 50 LSR1 sexual females (cross 2, Fig. 1). In Fig. 1, dotted lines show lineages used as male and plain lines those used as female. Eggs were kept at 4°C (80% humidity) for 85 days and were then transferred at 18°C for hatching. A few days after the first eggs hatched, 50 parthenogenetic larvae for each cross were isolated on new broad bean plants. Each F1 lineage was kept for 7 to 9 months under conditions sustaining clonal reproduction (16 h light, 18°C). Reproductive phenotype of the F1 lineages was then assessed similarly.
Six F1 lineages (three per cross) were then chosen to produce the next F2 generation (Fig. 1). All F1 produced sexual females (with from 27% to 71% and 28% to 64% sexual females for cross 1 and 2, respectively), hence were CP. The 6 F1 clones were thus chosen to cover the diversity in terms of the production of males (that ranged from 0–73% and 0–55% males for cross 1 and 2, respectively) and asexual females (that ranged from 0–42% and 1–53% for cross 1 and 2, respectively). Five F2 crosses (crosses 3 to 7 in Fig. 1) were performed using the same protocol as for the F1. 44 to 47 F2 lineages per family (hence 229 F2 lineages in total) were then isolated and kept for subsequent assessment of reproductive mode phenotype (same protocol as for the F0 and F1). Twenty-six F2 lineages (out of 229) died before being phenotyped.
The three grand-parents, the six F1 parents and the 229 F2 individuals from families 3 to 7 were typed at 401 microsatellite loci (see S2 Table for loci used and [42]–[44] for primer sequences). We first checked for the presence of null alleles by looking at the inheritance of alleles in the 3-generation pedigree. Homozygous individuals originating from parents displaying a null allele were transformed into missing data. Loci located on the same chromosome were identified based on their complete linkage in males (2n = 8 in the pea aphid and chromosomes in males do not recombine) [27], [45]. Genetic maps were then constructed for each of the four chromosomes with Crimap 2.53 [46] using Kosambi mapping function. Linkage maps were drawn using MAPCHART v. 2.1 [47]. QTL detection was then performed with the interval mapping method implemented in QTLmap, using the LDLA approach [48]. The phenotypic traits analysed for each F2 lineage (from crosses 3 to 7) were the occurrence (binary variable) and the percentage (quantitative variable) of sexual females in the parthenogenetically produced offspring. We focused on these traits because the production of sexual female is the most relevant variable to predict whether a population is able to reproduce sexually or not [10], [49]. In our analyses, we set parameter ndmin to 200 so that no information from males meioses was used to locate QTLs (since chromosomes do not recombine in males, males are not informative to locate QTLs within chromosomes). QTLs were detected based on likelihood-ratio (LR) along chromosomes. LR values corresponding to a significance level of 0.05 for each chromosome were empirically determined from 1,000 simulations under the null hypothesis of the test (i.e. no QTL). Genome-wide significance levels (i.e. LR values corresponding to adjusted p-values) were computed to account for multiple testing (i.e. four tests, corresponding to the four chromosomes). The drop-off method implemented in QTLmap was applied to obtain 95% confidence intervals of the QTL location. Similarly to the reduction of x-LOD when using LOD scores, the maximum LR value was reduced by 3.84 (corresponding to a Chi2 distribution with one degree of freedom for p<0.05) to determine a threshold. Region boundaries were then defined by the LR locations crossing this threshold upstream and downstream of the peak LR [as described in 50], [51] to identify the 95% CI of the QTL. After identifying the first QTL (see Results), we tested for the presence of a second QTL. For this, genotype at locus T_121775_26 (the closest marker from the peak of the QTL on LG1) was introduced as a fixed effect in the model. This marker is highly discriminative as each of the three grandparents possesses different alleles. LR for the presence of a QTLs against the hypothesis of no QTL was then compared to LR thresholds corresponding to a 5% significance determined by 1000 simulations in QTLmap.
The QTL approach led to the identification of a single genomic region, located on the X chromosome, which controls reproductive mode (see Results). Yet, in the three F2 crosses that were highly informative, all lineages used as mother inherited by chance the same X chromosome copy from their OP father (remind that chromosomes do not recombine in male aphids). From these crosses and from recombination events, we determined that the gene(s) that control(s) reproductive mode locate(s) between markers T_128012_2_G (34.8 cM) and T_126075_3_Y (49 cM) on this X copy (see Results). The segment of this X chromosome copy is referred to as “op2 allele” hereafter. However, we had little power to test whether the same region also controls this trait on the second X copy from the OP grandparent (that we refer to as “op1 allele”). We therefore performed an additional F2 cross (cross 8: X2_25♂×X1_3 ♀) to recombine the X-chromosome bearing the op1 allele (the mother X1_3 ♀ bears one op1 allele). We also crossed X6_2♂×X3_4♀ (that each possesses an op2 allele, cross 9) to produce homozygous individuals at this X chromosome region (i.e. op2/op2) to assess the dominance status of the different alleles in the candidate region (i.e. op1, op2, and those inherited from CP clones, referred to as CP1, CP2, CP3 and CP4). Since these two crosses were performed after we had identified the genomic region controlling reproductive mode variation, only a subset of individuals were phenotyped (24 and 27, respectively), chosen accordingly to their genotype at 8 microsatellite markers in the genomic region of interest (see S1 Figure for markers used).
Pea aphid individuals were collected in alfalfa fields from six sampling sites (S3 Table). All A. pisum individuals were sampled from the same plant species (Medicago sativa) to prevent confounding effects of plant or reproductive mode specialization on genetic divergence [52]. Three of the sites locate in north-east France or Switzerland and correspond to regions with cold winters (“temperate continental climate” as defined in [53]). In these areas, pea aphid populations consist mainly of CP lineages, because eggs are the only stage that survives cold winters [19] (we thus consider these areas as CP-selecting environment). Individuals were collected in spring 2008, a few weeks after egg hatching to maximize the probability to sample locally overwintering populations (these samples have been used in [43], [44]). The three other sampling sites locate in south-west France, and correspond to regions characterized by mild winters (i.e. “temperate oceanic climate” as defined in [53]). These areas are considered as OP-selecting environment. Here, sampling took place in winter 2008–2009 because at this season, OP lineages can be discriminated from CP ones, since the former overwinter as parthenogenetic females while the latter spend winter as eggs. Parthenogenetic females were collected from the six populations (see [43] for further details). To obtain sufficient amounts of DNA for genotyping hundreds of microsatellites, field-collected aphids were grown individually in controlled conditions ensuring continuous clonal reproduction (16 h light/day, 18°C). In each of the 6 geographic populations we then kept 20 individuals (except for one population for which only 15 individuals successfully established in the lab to provide sufficient DNA) on which all further analyses were conducted. These 115 individuals were genotyped at 443 microsatellite loci (301 of them were positioned in the genetic maps, see S2 Table). Six individuals (out of 115) with more than 15% missing genotypes were removed as well as seven markers (out of 443) with more than 30% missing data.
To detect loci that depart from neutral expectation, and which are therefore potentially involved in reproductive mode variation, we used a hierarchical method [54] implemented in ARLEQUIN 3.5 [55]. The distribution of the genetic differentiation among populations characterized by different reproductive mode expected under neutrality was estimated by means of coalescent simulations. The among-reproductive mode differentiation was characterized by the parameter FCT, which accounts for the geographical structure within populations (three populations originate from OP-selecting environments and the three other from CP-selecting environments). 100 000 coalescent simulations were performed conditionally on the multilocus estimate of FCT at the 436 microsatellite loci, assuming 50 groups and 100 demes per group. The observed data from each locus were compared with the simulated distribution, and a particular locus was classified as a significant outlier if it fell outside the 99% confidence envelope. We focused here on loci putatively involved in divergence between populations with contrasted reproductive mode; hence, we considered in this analysis only the loci falling above the upper confidence limit. As we were interested in identifying outlier loci involved in the variation of reproductive mode, and not in adaptation to local environmental conditions, we checked that the outliers identified from this global analysis (in which the two types of populations were included simultaneously) were not classified as outliers (either under divergent or balanced selection) among populations sharing the same reproductive phenotype. To that end, we ran two independent analyses for the detection of outliers within populations sharing the same reproductive mode. We also checked that the outcomes of genome scan analysis were not affected by the inclusion of markers with >10% missing data (58 loci). Since the confidence interval was similar when including or not markers with >10% missing data and because our aim was to screen the genome with the highest number of markers, we present only the analysis based on the whole dataset (436 markers).
|
10.1371/journal.ppat.1002205 | A Diverse Population of Cryptococcus gattii Molecular Type VGIII in Southern Californian HIV/AIDS Patients | Cryptococcus gattii infections in southern California have been reported in patients with HIV/AIDS. In this study, we examined the molecular epidemiology, population structure, and virulence attributes of isolates collected from HIV/AIDS patients in Los Angeles County, California. We show that these isolates consist almost exclusively of VGIII molecular type, in contrast to the VGII molecular type isolates causing the North American Pacific Northwest outbreak. The global VGIII population structure can be divided into two molecular groups, VGIIIa and VGIIIb. Isolates from the Californian patients are virulent in murine and macrophage models of infection, with VGIIIa significantly more virulent than VGIIIb. Several VGIII isolates are highly fertile and produce abundant sexual spores that may serve as infectious propagules. The a and α VGIII MAT locus alleles are largely syntenic with limited rearrangements compared to the known VGI (a/α) and VGII (α) MAT loci, but each has unique characteristics including a distinct deletion flanking the 5′ VGIII MATa alleles and the α allele is more heterogeneous than the a allele. Our studies indicate that C. gattii VGIII is endemic in southern California, with other isolates originating from the neighboring regions of Mexico, and in rarer cases from Oregon and Washington state. Given that >1,000,000 cases of cryptococcal infection and >620,000 attributable mortalities occur annually in the context of the global AIDS pandemic, our findings suggest a significant burden of C. gattii may be unrecognized, with potential prognostic and therapeutic implications. These results signify the need to classify pathogenic Cryptococcus cases and highlight possible host differences among the C. gattii molecular types influencing infection of immunocompetent (VGI/VGII) vs. immunocompromised (VGIII/VGIV) hosts.
| Infections that were once uncommon have become considerable public health threats due to the increased number of patients with weakened immunity, including a large HIV infected population. Among the emerging infections in the HIV/AIDS population are those caused by fungi. Fungi are distinct from viral or bacterial pathogens and often cause skin (athletes foot), mouth (thrush), or vaginal (yeast) infections. Life-threatening infections of the lungs and brain are caused by Cryptococcus, which infects >1,000,000 patients annually causing ∼1/3 of all AIDS-related deaths. Many infections are caused by C. neoformans, but the sibling species C. gattii infects both healthy individuals and HIV/AIDS patients and is causing an outbreak in the Pacific Northwest. Often, Cryptococcus is not identified at the species level, and C. gattii may therefore be more common than appreciated. We studied C. gattii isolates infecting HIV/AIDS patients in southern California and found they are unrelated to those causing the Pacific Northwest outbreak (molecular type VGIII rather than VGII). We show that within VGIII there are two subgroups, and one is more virulent than the other in animals. Our studies establish a foundation for research on C. gattii in HIV/AIDS patients. Given the global magnitude of infections, these findings have significant public health implications.
| The pathogenic Cryptococcus species complex is comprised of two common fungal pathogens of humans and other animals: C. neoformans and C. gattii [1]. The more prevalent C. neoformans is ubiquitously distributed worldwide and a common cause of meningitis in immunocompromised hosts [1], [2], [3]. C. gattii is more geographically restricted to tropical and subtropical regions, associated with eucalypts, Douglas fir, and other trees, and has a greater predilection for infecting immunocompetent hosts [4], [5]. However, the geographic distribution of this species has been expanding, with an outbreak occurring in the Pacific Northwest region of North America [4], [5], [6], [7], [8], [9], [10], [11], [12]. C. gattii can be subdivided into two serotypes (B and C) [13] and four molecular types (VGI, VGII, VGIII, VGIV) that appear to represent genetically isolated cryptic species [14], [15], [16]. VGI and VGII cause the majority of cases in otherwise healthy hosts. VGIII and VGIV appear to more commonly infect immunocompromised patients, including those with HIV/AIDS, similar to C. neoformans [3], [5], [17], [18], [19].
Compared to C. neoformans, less is known about the epidemiology and ecology of C. gattii, especially molecular types VGIII and VGIV. VGIV is rare globally but has been reported to cause infections in sub-Saharan Africa AIDS patients [18], while VGIII has been isolated from a number of regions worldwide [14], [16]. The limited epidemiological data may be due to a lack of laboratory species distinction, particularly in Cryptococcus cases among HIV/AIDS patients, although sporadic C. gattii infections in HIV/AIDS patients have been reported from many global regions [18], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30].
There are limited accounts for HIV/AIDS or VGIII/VGIV associated C. gattii in North America. However, C. gattii has been reported in southern California among an HIV/AIDS patient cohort and one AIDS patient in Mexico, and C. gattii VGIII/VGIV have been reported in clinical cases from Mexico [21], [22], [31]. Additionally, there are three reported VGIII isolates from the Pacific Northwest outbreak [6], [7], [10]. These reports suggest that C. gattii may be endemic in western North America.
Here we compared isolates from HIV/AIDS patients in southern California [22] with a global collection. Based on MLST analyses, >93% (28/30) of the C. gattii isolate cohort are VGIII, and genotypic diversity is much greater than in the Pacific Northwest VGII outbreak. The genotypes cluster into two distinct groups (VGIIIa, VGIIIb). These findings suggest that VGIII may have been endemic in southern California for a longer period of time compared to the more clonal VGII Pacific Northwest population, or that VGIII is more actively recombining and/or mutating. The high levels of diversity, together with results suggesting recombination, support ongoing genetic exchange.
The VGIII lineage is highly fertile compared to other C. gattii molecular types [32]. Isolates from our cohort include both mating types (α and a), unlike the exclusively α mating type Pacific Northwest outbreak [7], [16]. This is noteworthy, as a-α mating can promote recombination and yield infectious spores [33], [34], [35], [36], [37], [38]. Sexual recombination is critical in evolution of eukaryotic microbial pathogens, including both parasites and fungi [39], [40], [41], [42], [43]. For these reasons, we extended the analysis of the VGIII group to include extensive mating assays to determine genotypes associated with high fertility. Additionally, sequenced and characterized the C. gattii VGIII MAT locus alleles. These findings expand on previous studies of the Cryptococcus MAT locus [44], [45], [46], [47], [48] yielding new insights into plasticity of this genomic region involving rearrangements, gene truncation, and loss, which may impact fertility.
While previous studies have shown C. gattii is highly virulent in mice, the focus has been on genotypes causing disease in otherwise healthy hosts [16], [49], [50], with no studies to date comprehensively examining virulence of VGIII. Here we examined murine virulence and macrophage intracellular proliferation for both the VGIIIa and VGIIIb. The isolates were also selected based on mating type criteria to examine what roles, if any, the MAT locus plays in VGIII virulence. This was of interest, as previous studies have shown virulence differences in C. neoformans congenic isolates that differ only at MAT [51], [52]. Our results demonstrate the VGIIIa lineage has increased levels of virulence compared to VGIIIb.
Our studies reveal a complex population structure within the highly fertile VGIII molecular type, suggesting recombination in both the United States and globally via opposite and/or same-sex mating. We posit that genetic exchange and the formation of infectious spores may be contributing to an underlying endemic level of C. gattii infections in southern California. Additionally, we show that these isolates are virulent in both murine and macrophage infection models [10]. Overall, the VGIII isolates from HIV/AIDS patients show decreased levels of virulence in comparison to the Pacific Northwest outbreak VGII isolates. This study highlights the need for isolate typing at a resolution sufficient to distinguish both species and molecular type. Accurate isolate identification will advance understanding of the epidemiology, ecology, and health burden of this emerging pathogen, with potential prognostic and therapeutic benefits for the clinical management of cryptococcal infections.
To characterize the molecular type of C. gattii isolates collected from the southern California C. gattii cohort [22], we applied multilocus sequence typing (MLST) analysis at eight unlinked genomic loci. In total, the patient cohort consisted of 30 isolates. Of these, one was VGI (3.3%), another isolate was VGII (3.3%), and the remaining 28 isolates were molecular type VGIII (93.3%). Examination of the 28 VGIII isolates revealed a high level of diversity (13 unique genotypes based on seven MLST markers), isolates of the two mating types (α and a) (Figure 1), and no evidence for heterozygosity (consistent with FACS analysis showing they are haploid, data not shown). There are two genetically distinct groups within the VGIII molecular type. One was termed VGIIIa (predominantly orange shading in Figures 1, 2), and the other termed VGIIIb (predominantly green shading in Figures 1, 2). Additionally, two diploid isolates from the cohort (CA1388 and CA2355) were excluded due to high levels of heterozygosity in the majority of sequences, indicating they are likely hybrids. Hybrids between C. neoformans and C. gattii have been previously reported [53] (see discussion).
To further examine the isolates, their MLST profiles were compared to an additional 32 VGIII isolates collected from multiple locations and sources (Figure 2). Based on an examination of 60 VGIII isolates, the VGIIIa and VGIIIb clusters form two distinct groups. While many alleles were VGIIIa or VGIIIb specific, several were shared. Alleles shared between VGIIIa and VGIIIb are colored in fuchsia (Figures 1 and 2). In the VGIIIa cluster, all isolates originate from North America and Australia, while in the VGIIIb cluster, isolates originate from North America, South America, and Asia (Figure 2). Thus while there are possible geographic niches for each individual group, some geographic regions (Mexico and the US) harbor isolates from both groups. This finding suggests that in at least North America, the two groups might occupy similar environmental niches with potential for cross-hybridization.
We next examined the distribution of the sequence types by constructing maximum likelihood (ML) dendrograms (Figure 3A). This analysis also supports two distinct clusters. In total, 28 sequence types are represented in the seven-loci dendrogram 12 in VGIIIa and 16 in VGIIIb. One VGIIIa genotype, ST28, is an intermediate between the two subgroups, and consists of two clinical isolates from Mexico (97/426 and 97/427). This sequence type harbors two alleles shared between the two groups, two alleles common and exclusive to the VGIIIa group, and four alleles unique to this sequence type. The ancestry of this isolate remains unclear; however, both ST28 isolates originate from Mexico, a region that harbors isolates of both VGIIIa and VGIIIb, and thus these isolates may represent VGIIIa/VGIIIb hybrids.
The global population was examined by constructing Neighbor Joining phylograms (Figure 3B, Figure S1). This analysis also illustrates two defined lineages, with the ST28 genotype representing a VGIIIa/b intermediate. Bootstrap support for the separation of the VGIIIb lineage from the other genotypes is robust (100) (Figure 3B). Support for all VGIIIa genotypes other than ST28 is at a level of 86, indicating that ST28 may be a distinct lineage, or conversely be a divergent genotype within the VGIIIa lineage (Figure 3B). To address the ancestry of the VGIII subgroups, we examined the isolates in the context of the three other C. gattii molecular types (VGI, VGII, VGIV) (Figure S1). From this analysis, there is strong support (bootstrap value of 100) that the VGIIIa subgroup is ancestral to the VGIIIb subgroup, with ST28 as the closest genotype to the VGIIIb clade (Figure S1). While the two clusters (VGIIIa/VGIIIb) are distinct from one another, when compared with VGI, VGII, VGIV, all 60 VGIII isolates lie within the VGIII lineage, clearly delineated from the other molecular types with a bootstrap support of 100 (Figure S1).
Based on an examination of the 60 total isolates, the VGIIIa and VGIIIb clusters may represent early speciation. To address aspects of the evolutionary history of the groups, haplotype network analysis was applied using TCS phylogenetic estimation for allele ancestry and evolution. The primary alleles focused on in this analysis were those shared between the subgroups. This examination addresses if shared alleles have a probable ancestral origin, or conversely if they were more recently introgressed. Of the shared alleles, SXI2a was not informative as there is only one allele. For the other three loci (CAP10, TEF1, PLB1), TCS analysis revealed that the CAP10 locus and TEF1 locus shared alleles appear ancestral, while the most parsimonious hypothesis is that the PLB1 shared allele was introgressed between VGIIIa and VGIIIb because another allele (PLB1-20) is assigned as the ancestral root (Figure 4A–C).
In addition to markers encompassing shared alleles, haplotype networks were also constructed to examine the evolutionary history of the remaining five markers (Figure S2). For this analysis, four of the five markers (LAC1, MPD1, GPD1, and IGS) showed that the VGIIIa and VGIIIb subgroups are separated (Figure S2). This provides additional evidence for genetic isolation. The remaining marker, SXI1α, shows evidence for separation, but the history is less well resolved, and shared introgressed alleles may remain to be discovered in the population (Figure S2). This marker is highly variable and many of the proposed intermediates have not been found to date. This could be the result of: 1) under-sampling of isolates, 2) missing alleles that are either no longer extant in the population or never existed in cases where mutations occurred simultaneously, or 3) recombination within the allele. Overall, this analysis supports a genetic separation, with several hypothesized but likely rare introgression events having occurred.
To examine the role that recombination may have contributed to population structure, we conducted a paired allele analysis among the global genotypes (Figure 5, Figure S3). The discovery of all four possible allele combinations between two unlinked loci (AB, ab, Ab, aB), in the absence of parallel mutations, serves as evidence for recombination [54], [55]. In total, 18 of 28 different molecular marker pairs showed evidence for recombination, with 25 examples in which all four allele combinations were observed. To further classify the analysis of the VGIIIa and VGIIIb subgroups, we examined how frequently allele combinations indicating recombination were associated with a specific lineage. In total, 1/25 locus pairs involved only the VGIIIa lineage, whereas 20/25 pairs involved only the VGIIIb lineage; the four remaining pairs involved shared alleles. Of the four pairs involving shared alleles, two involved shared and VGIIIa alleles and two involved shared and VGIIIb alleles. No pairs contained both VGIIIa and VGIIIb unique alleles (Figure 5, Figure S3). These studies demonstrate that while recombination may be present within both subgroups, the VGIIIb subgroup may be more actively undergoing recombination in its environmental niche.
To further characterize the role recombination may play in population structure, we employed MultiLocus software to examine the percentages of compatible loci and indices of association (IA). These analyses were completed for the global population, and also individually for the VGIIIa and VGIIIb subgroups. Additionally, the dataset was analyzed using all 60 isolates, and then was analyzed as a clone-corrected set with only unique genotypes represented (n = 28). Clone correction is commonly used to examine fungal populations that are known to often undergo asexual reproduction and clonal blooms, with less frequent sexual reproduction and meiosis [56]. The analysis of both the percentage of compatible loci and IA produce p-values that if significant reject the null hypothesis of recombination. If the p-values are not significant, the null hypothesis cannot be rejected and recombination can be inferred. The analysis illustrates that although the null hypothesis is rejected in the analysis of all isolates, VGIIIa individual, VGIIIb individual, and the overall clone corrected dataset, the null hypothesis cannot be rejected in the analysis of the individual VGIIIa and VGIIIb subgroups when the dataset is clone corrected (Table 1). As we cannot reject the null hypothesis for recombination in the clone corrected subgroup analysis, it suggests recombination may be occurring within each subgroup. These results support the hypothesis of active recombination within the VGIII molecular type and that these events more likely occur between isolates within the same subgroup, rather than between subgroups, similar to studies supporting recombination in VGI and VGII populations in Australia [35], [36].
In addition to population-based studies, we conducted extensive mating assays among VGIII isolates, showing highly fertile isolates from both subgroups. In total, 182 mating pairs were examined by laboratory mating assays, including all eight mating type a isolates, and representatives from each of the α genotypes. When observed via light microscopy, 138 pairs showed no signs of fertility, while 44 pairs were able to undergo sexual reproduction, with representative high resolution SEM imaging of a VGIII×VGIII cross (NIH312×B4546) shown in Figure 6 (see also Tables 2 and S1). The SEM imaging, similar to previous studies [32], illustrated key hallmarks of mating including hyphae, fused clamp cells, basidia, and elongated basidiospores (Figure 6). When the strains that were fertile with the greatest number of mating partners were separated into the top four from each respective mating type, three of four from both mating types are from the VGIIIb subgroup (Table S1). This shows that while both groups are fertile, levels of fertility (based on the number of fertile partners) may be higher in the VGIIIb subgroup, consistent with the increased number of unique VGIIIb genotypes (n = 16 out of 27 isolates or 59%) compared to VGIIIa (n = 12 out of 33 isolates or 36%). Overall, a higher percentage of a isolates are fertile compared to α isolates (Table 2). When both mating types are combined VGIIIb shows an increased percentage of fertile isolates compared to VGIIIa (Table 2). These findings support the hypothesis that mating may also occur in the environment.
Sequencing of the α and a MAT locus alleles from two representative strains of the VGIII lineage shows that overall, the general structure, size, and characteristics are similar to previously sequenced C. gattii VGI and VGII MAT loci (Figure 7). Both SXI1α or SXI2a and the pheromone receptor and pheromone genes are present in the MAT locus, further supporting that C. gattii VGIII has a bipolar mating system.
Although there are marked similarities for the MAT locus alleles of VGI/VGII with VGIII, there are also distinct characteristics for each VGIII MAT allele. There are two rearrangements in the MATa allele 3′ region compared to the VGI isolate E566. Moreover, a partial deletion of UAP1 and loss of the FCY1 and FAO1 genes from the VGIII MATa allele 5′ flanking region was identified via sequencing of the VGIII a isolate B4546 and confirmed to be present in all eight VGIII a isolates examined, although no phenotypic consequences were observed (Figures 7, S4, and S5). Southern blot analysis of FCY1 and PCR analysis of all three full length genes with primers selected based on regions conserved between C. neoformans and C. gattii shows that they might not be present in the genome, or alternatively that they are rapidly evolving and too diverged to hybridize with probes and primers used (Figure S5). All isolates remain sensitive to the antifungal agent 5-Fluorocytosine (5-FC) indicating that the FCY1 gene is functioning or that another gene acts in a similar manner. Additionally, an ∼1.1 kb truncation of the FAO1 gene, a putative iron oxidoreductase, was found in all VGIII α strains analyzed. Using PCR, we could not detect an intact FAO1 gene elsewhere in the genome of the VGIII α isolates. Four pheromone gene copies were previously found within the VGII and VGI MATα loci, arranged in two pairs of divergently oriented, linked genes. However, although VGIII has been shown to be more fertile than VGI/VGII, based on the sequence assembly for strain NIH312, we found only two copies of the MFα pheromone gene, although one gap remains in this region of the locus. Additionally, we discovered a remnant of the a specific gene SXI2a in the VGIIIα mating type locus, and through sequence comparisons show that it is present in C. gattii molecular types VGI, VGII, and VGIII, but not present in C. neoformans var. neoformans (serotype D) or C. neoformans var. grubii serotype A (Figure S6).
Fingerprinting analysis of the MAT locus alleles revealed no size polymorphisms within MAT for VGIII a isolates, but did reveal diversity within the α allele (Figure 7, Table S2). Distinct alleles for fingerprint products 9 and 18 of the α locus, were identified that are correlated with the VGIIIa and VGIIIb molecular types (Figures 7, Table S2). Additionally, a 120 bp polymorphism within the CID1 and LPD1 intergenic sequence (fingerprint 8 of the α locus) was correlated with VGIIIa and VGIIIb molecular types with the exception of one VGIIIb strain harboring the SXIα allele 38, which contained the VGIIIa genotype at this region and could reflect recombination within MAT as a result of α-α mating (Figure 7, Table S2). Other VGIII isolates with SXI1α allele 38 were also confirmed to have this genotype (data not shown) and fingerprints 2 and 7 of the α locus showed multiple genotypes that were not correlated with the subgroups.
To address the virulence characteristics of the VGIII isolates, and correlate these with genotype and mating type, we conducted murine intranasal instillation challenges. In total, we chose 11 isolates from the patient cohort and one environmental VGIII control strain isolated in San Diego, California from an E. camaldulensis tree in 1992 [57]. Of the eleven clinical isolates, six were from the VGIIIa subgroup and five from the VGIIIb subgroup, with eight α and three a isolates (Figure 8A). Of the isolates examined, CA1499 was more virulent than all other isolates tested (Figure 8A). Additionally, when we compared overall virulence between the two subgroups, the VGIIIa subgroup showed a significantly higher mortality rate when compared to the VGIIIb subgroup (p<0.025). When isolate CA1089 was excluded from this analysis, the support increased (p<0.005). These results suggest that the molecular differentiation between the two lineages within VGIII is associated with a dichotomy in murine mortality post cryptococcal infection. We also show that the environmental isolate WM161 (VGIIIa) was less virulent than the five clinical VGIIIa isolates, but more virulent than all of the VGIIIb isolates. Isolates with increased levels of mortality were also found to be associated with rapid declines in total body weight during the course of infection (Figure 8B).
Histological examinations of lung tissues from mice infected with the highly virulent C. gattii strain CA1499 revealed widespread dissemination of cryptococci throughout the alveoli and airways with little to no inflammatory response. The expansion of alveoli with confluent clusters of budding yeasts and breakdown of alveolar walls were seen (Figure 9A–9B). The lungs of mice infected with strain CA1292 (moderate virulence) also revealed minimum inflammation with alveoli engorged with rapidly dividing yeasts. Also evident were a few alveoli with one to two C. gattii yeasts. Despite rapid multiplication, this strain appeared to be less disseminated as some of the alveoli and airways were devoid of any yeast cells (Figure 9C–9D). In contrast, lungs of mice infected with the low virulence strain WM161 revealed influx of inflammatory response surrounding infected alveoli and airways. The majority of the infected alveoli contained only one to two yeasts with a few alveoli containing three or four yeasts (Figure 9E–9F). No yeast cells were seen in any of the lung sections of mice infected with avirulent strain CA2339; however, vigorous tissue response in the form of neutrophil influx was observed throughout the lung section (Figure 9G–9H). The brain sections of all infected mice did not reveal any apparent lesions except for the highly virulent strain (CA1499), which showed occasional single or budding yeasts on the meninges (data not shown).
Although there were some differences in the dissemination pattern of CA1499 (high virulence) and CA1292 (moderate virulence) by histopathology, the quantification of yeasts from infected lungs at weekly intervals did not reveal any significant difference (Figure 10). This could be due to heavy colonization by CA1292 in the infected sites (Figure 9C–9D). When the lung organ loads of these two strains were compared with that of lung organ load of mice infected with low virulent strain WM161, the results were statistically significant (P<0.05). In comparison, the avirulent strain was cleared rapidly by one week post-infection and was not found in the lungs at two weeks post-infection (Figure 10). Overall, these results indicated that C. gattii rapid multiplication, dissemination, and invasion of the host defense of the lungs dictate the outcome of pulmonary cryptococcosis. Only small numbers of yeasts were recovered from brain cultures of mice infected with virulent strains of C. gattii (data not shown). These results indicated that pulmonary cryptococcosis and not CNS dissemination was the primary disease manifestation in this model and the likely cause of mice mortality observed (Figure 8).
Intracellular proliferation rates (IPRs) were determined for each strain examined in the murine model of infection. This assay is correlated positively with murine virulence, based on previous studies Pacific Northwest outbreak isolates [10], [49]. To further examine the virulence characteristics for the VGIII isolates, intracellular proliferation of cells within macrophages was directly tested for the seven VGIIIa and five VGIIIb isolates that were examined in the mouse model (Figure 11 A–B). Similar to the murine experiments, the VGIIIa subgroup showed significantly higher IPR levels than the VGIIIb subgroup (p<0.005, Figure 11 A, C). Additionally, as in the previous studies, IPR values and survival time in vivo showed strongly positive correlations upon regression analysis (Figure 11 B, D). A single VGIIIa isolate, CA1089, was an outlier with high IPR but low virulence. Therefore, the regression analysis was conducted both with (Figure 11 A, B) and without this isolate (Figure 11 C, D). While the significance levels were better supported when CA1089 was excluded, they were statistically significant in both cases (Figure 11 B, D). The IPR analysis combined with the in vivo model shows that the isolates from the Californian patient cohort show virulence differences in which the VGIIIa subgroup was more virulent overall than the VGIIIb subgroup.
Globally the burden of cryptococcal disease is significant with approximately one million annual cases [19]. Although >99% of AIDS related infections and >95% of overall cases are attributed to C. neoformans serotype A [1], C. gattii also has been shown to cause disease among AIDS patients in both sub-Saharan Africa and the US [18], [22]. In the study of Litvintseva et al., an African AIDS cohort was found to be infected with C. gattii serotype C VGIV molecular type [18]. Here C. gattii isolates were examined from a previously reported HIV/AIDS patient cohort in southern California [22], and found to be almost exclusively (93%) VGIII molecular type. The finding of C. gattii VGIII and VGIV isolates associated with HIV/AIDS patients is in stark contrast to numerous C. gattii infections among immunocompetent individuals caused by VGI/VGII isolates, and the ongoing VGII outbreak in the North American Pacific Northwest [5], [7], [9], [10], [16], [58], [59], [60], [61], [62].
The Californian HIV/AIDS patient cohort C. gattii strains were analyzed in the context of a global VGIII isolate collection to further characterize this molecular type. Analysis of 60 isolates at eight loci resolved two distinct subgroups designated VGIIIa and VGIIIb. In contrast to the Pacific Northwest VGII outbreak that is largely clonal with one genotype causing the majority of illness [9], [10], [16], [63], based on analysis of seven non sex-linked loci we found a total of 13 unique genotypes in the 28 VGIII isolates from the Californian patients. This increased level of diversity suggests this molecular type was introduced into the region longer ago than the VGII lineage to the Pacific Northwest, or alternatively that the VGIII population in California is more rapidly diverging. Our studies, in addition to a recent study documenting VGIII isolates in Mexico, lead to the hypothesis that the VGIII molecular type may be endemic to a large area of western North America [16], [31]. These findings are significant, as there may be unrecognized cases of C. gattii among HIV/AIDS patients.
Phylogenetic analysis of the molecular types revealed that VGIIIa is basal to VGIIIb (Figure S1). Given that VGIIIa is more virulent in mice and murine macrophages, it may be that VGIIIb has become more specialized to infect hosts with increased susceptibility to infections. It currently remains unclear as to whether these two groups have naturally evolved to display differences in virulence or if there is some form of selection. Regardless, increased surveillance of VGIII cases, aided by clinical studies to determine the types of infection each group causes, should shed further light as to whether these two groups also have distinct clinical features. Clustering and phylogenetic analyses of the VGIII global collection revealed 28 unique genotypes among 60 isolates. Subsequent phylogenetic analysis revealed that the two observed groups were indeed well supported lineages. Haplotype mapping was then conducted, supporting the discrimination into VGIIIa and VGIIIb. Additionally, the mapping of markers harboring shared alleles between the subgroups revealed that two of the three shared alleles are ancestral in origin, while one may result from introgression between VGIIIa and VGIIIb. This type of introgression event is not unprecedented between distinct lineages, cryptic species, or species within the fungal kingdom, with examples from model fungi and both plant and animal pathogens [64], [65], [66], [67].
The ability for microbial pathogens to undergo sexual reproduction involving meiosis is unique to the eukaryotic lineage, including the parasites and fungi. Sex may play a significant role in several aspects of pathogenesis, including the generation of genetic diversity, the formation of invasive hyphae or spores, and direct links between the mating type locus and mating pathways with virulence [34], [35], [37], [39], [52], [68], [69], [70], [71], [72], [73], [74], [75], [76]. For these reasons, and to understand the population dynamics of VGIII, we conducted two types of recombination analyses: informative paired allele graphs and statistical analyses of percentages of compatible loci and IA. Our analyses support the hypothesis of recombination within the distinctive VGIII subgroups. This is consistent with the observed molecular differences between the two groups. Our population results, taken together with an increased number of highly fertile VGIIIb isolates in comparison to VGIIIa, suggest that the VGIIIb group may be more actively recombining. Furthermore, the two excluded hybrid isolates appear to be C. neoformans var. neoformans/C. gattii molecular type VGIIIb hybrids (based on the GPD1 MLST allele specifically amplifying C. gattii sequence and the IGS allele specifically amplifying C. neoformans sequence, while the other MLST loci were highly heterozygous). The enhanced fertility ofn the VGIIIb subgroup may contribute to hybridization with C. neoformans.
The ability to undergo sexual reproduction also has implications for the formation of infectious spores. Recent studies show that spores can initiate disseminated cryptococcal disease in the murine inhalation model of infection [34], [37]. Laboratory studies have also shown that Cryptococcus can complete a full sexual cycle in association with plants, leading to the production of infectious spores [77]. It has been previously shown that the VGIII lineage shows high levels of fertility [32] and our examination of compatible isolate pairs supports that mating and spore formation may play a significant role in the formation of small aerosolized particles that could be readily inhaled. Furthermore, in California, both mating types are present and several of these pairs are fertile under laboratory conditions.
The MAT alleles of C. gattii molecular types VGI and VGII are highly conserved, based on both DNA sequence similarity and gene synteny [48]. The sequenced MATα locus alleles of VGI WM276, VGII R265, and VGIII NIH312 share full synteny (Figure 7). A comparison of the three loci reveals only expansions, contractions, and translocation of intergenic sequences, which account for the majority of the structural variation. Additionally, there is also a truncation of the FAO1 gene in all VGIII α isolates examined. The sequenced MATa locus alleles of VGI E566 and VGIII B4546 show a higher degree of rearrangement in comparison to MATα (Figure 7). Additionally, a partial deletion of the UAP1 gene, along with a complete deletion of the FCY1 and FAO1 genes from the 5′end of the VGIII MATa locus is fixed in the VGIII a population. UAP1 is predicted to encode a candidate uric acid transporter whose closest homolog is the A. nidulans uapA gene [78], [79], [80] and Fcy1 is a cytosine deaminase that has been shown to confer resistance to 5-FC when deleted in C. albicans, S. cerevisiae, and C. neoformans [81], [82]. Although both PCR and southern blot analyses could not detect a copy of the FCY1 gene in the genome of the VGIII MATa isolates, phenotypic tests showed these isolates remain sensitive to 5-FC and they may therefore harbor a more diverged FCY1 gene elsewhere in the genome. Phenotypic assays on media with uric acid as a sole nitrogen source to examine the possible loss of UAP1 also showed no distinct phenotype from control isolates, i.e., all isolates were able to efficiently grow (data not shown). Based on findings in C. neoformans, recombination hotspots may flank the MAT locus alleles and thus have contributed to the flanking region deletions [83]. Overall, analysis of the C. gattii MAT locus alleles showed a more dynamic MATa locus compared to the α locus. Given the significant rearrangement of the MATa locus, and the deletion or exclusion of genes from the locus and flanking region, it appears that the VGIII MATa population has experienced increased genetic variation. Deletion or exclusion of genes from the MAT locus and flanking regions may be an indication of expansion or contraction of the VGIII MATa locus. The finding of a SXI2a remnant conserved in C. gattii may signal a previous tetrapolar mating type state, consistent with the prevailing models and studies in related Cryptococcus species (Figure S6 B) [46], [47], [48], [84], [85], or result from a C. gattii lineage specific gene conversion of SXI2a into the α allele (Figure S6 C). Alternatively, this remnant may be functional and influence mating.
Within VGIII, fingerprinting analysis revealed a monomorphic MATa allele and a mostly biallelic MATα locus correlating with the VGIIIa and VGIIIb lineages. Two fingerprints revealed size polymorphisms correlated with VGIIIa and VGIIIb isolates, serving as additional evidence for ongoing genetic separation between these two groups. VGIIIb strains harboring SXIα allele 38 contained the VGIIIa genotype for the intergenic sequence between CID1 and LPD1, suggesting that these strains harbor either the ancestral VGIII MATα allele or contain a hybrid MATα locus that may be the result of a gene introgression event between the two groups or a recombination event between same-sex isolates of opposite VGIII lineages (i.e., VGIIIa/VGIIIb). However, an additional polymorphic marker within the MATα locus is needed to address the latter hypothesis.
Following the definition of two well-supported lineages including isolates from both mating types, we examined virulence characteristics associated with these properties. We found significant differences between phylogenetic subgroups. In both whole animal murine in vivo intranasal instillation and proliferation assays in murine derived macrophages, the VGIIIa subgroup is more virulent than the VGIIIb subgroup. These findings are significant and serve as a foundation for future studies to determine the molecular basis for these observed phenotypes. Additionally, from a public health and epidemiological standpoint, it may be useful to determine if isolates are VGIIIa or VGIIIb. Clinical studies would have to be conducted in coordination to ascertain if the molecular subgroup is associated with altered clinical manifestations or outcomes. Based on the histological examinations, our findings indicate that host tissue responses in combination with yeast cell multiplication and capsule induction are associated with the outcome of pulmonary cryptococcosis, similar to previous studies documenting that upon serial in vivo passage of C. neoformans, strains increase in virulence with an associated decrease in capsule size [86].
Our study provides a comprehensive molecular and phenotypic overview of the C. gattii VGIII molecular type, which has historically been less studied than both the VGI and VGII molecular types. Our findings support two distinct lineages that might each be recombining, with the VGIIIa lineage showing higher levels of virulence in the models examined. Of significance, many of the isolates examined were from a cohort of HIV/AIDS patients in southern California. The high level of VGIII observed is in stark contrast to the Pacific Northwest VGII outbreak of VGII, in which the vast majority of cases reported are not associated with HIV/AIDS infected patients [7], [9], [10], [16], [63]. This suggests that C. gattii may occur in two general patient settings: VGI/VGII in otherwise healthy hosts (>50%) or those treated with steroids, vs. VGIII/VGIV predominantly in HIV/AIDS patients. Moreover, C. gattii infections may cause a substantial unrecognized health burden. To address these aspects, both retrospective and prospective studies should be conducted to: 1) survey global isolate collections from HIV/AIDS patients and 2) assign species and molecular types to newly collected isolates from HIV/AIDS patients with cryptococcal infections.
All isolates from southern California and other global isolates were screened to confirm that they were C. gattii. Melanin production was assayed by growth and dark pigmentation on Staib's niger seed medium; urease activity was detected by growth and alkaline pH change on Christensen's agar. These tests established that isolates were Cryptococcus (C. neoformans or C. gattii). Additionally, isolates were assayed for resistance to canavanine and utilization of glycine on L-canavanine, glycine, 2-bromothymol blue (CGB) agar. Growth on CGB agar indicates that isolates are canavanine resistant, and able to use glycine as a sole carbon source, triggering a bromothymol blue color reaction indicative of C. gattii. All CGB positive isolates were then grown under rich culture conditions prior to genomic DNA extraction and storage at −80°C in 25% glycerol. For genomic DNA isolation, the MasterPure Yeast DNA purification kit from Epicentre Biotechnologies was used.
Each VGIII isolate examined in the analysis was subjected to multilocus sequence typing (MLST) [87] at a total of eight loci (Table S3). This marker set was selected to include loci that have been validated in other analyses of C. gattii [7], [16], [88], [89], [90]. For each isolate, genomic regions were PCR amplified, purified (ExoSAP-IT, Qiagen), and sequenced. Sequences from both forward and reverse strands were assembled with complete double strand coverage and manually edited using Sequencher version 4.8 (Gene Codes Corporation). Based on BLAST analysis of the GenBank database (NCBI), each allele was assigned a corresponding number, or given a new number if the sequence was not already in the database. GenBank accession numbers with corresponding allele numbers are listed in the supplementary information (Table S4).
For each isolate, a sequence type was defined as a sequence exhibiting a unique sequence profile, based on concatenation of the MLST markers. Each sequence type was confirmed to be unique by BLAST analysis of the NCBI GenBank database [91]. A multiple alignment of the sequences was carried out using Clustal W software [92]. Clustering analysis of the sequences was conducted using PhyML, which applies the maximum likelihood model for analysis [93]. The phylogenetic analysis was conducted using the Neighbor Joining algorithm. Haplotype network modeling was conducted using TCS software (version 1.21) [94]. The statistical recombination analysis was completed using MultiLocus software (version 1.2.2).
Mating assays were conducted on V8 media (pH = 5). Isolates were incubated at room temperature in the dark for 2–4 weeks in dry conditions. Fertility was assessed by light microscopic examination for hyphae, fused clamp cells, basidia, and basidiospore formation at the periphery and surface of the co-incubated mating patch. All mating assays were conducted in duplicate. If there were no signs of fertility after the four-week period, isolate pairs were scored as having no fertility when paired together.
SEM analysis was completed using protocols similar to those previously published [34]. Mating cultures (NIH312×B4546) were excised from V8 medium agar plates and fixed at 4°C in 3% glutaraldehyde that was buffered in 0.1 M Na cacodylate (pH = 6.8). Samples were then washed in triplicate using cold 0.1 M Na cacodylate buffer. This was followed by a graded dehydration series of 1 hr changes in cold 30% and 50% ethanol and held overnight. Dehydration was completed with 1 hr changes of cold 95% and 100% ethanol at 4°C and warming to room temperature in 100% EtOH. Two additional 1 hr changes of room temperature 100% EtOH completed the dehydration series. The samples were then critical point dried in liquid CO2 (Samdri-795; Tousimis Research Corp., Rockville, MD) for 15 min at the critical point. The agar pieces were mounted and sealed with silver paint to ensure good conductivity. The samples were sputter coated with 50 Å of Au/Pd (Hummer 6.2; Anatech U.S.A., Hayward, CA). Samples were held under vacuum conditions until viewed with a Jeol JSM 5900LV scanning electron microscope at 15 kV.
The strains used for the construction of the Bacterial Artificial Chromosome (BAC) and fosmid libraries and analysis of the MAT locus alleles were NIH312 (MATα) and B4546 (MATa). Isolates were grown on YPD media at 30°C, and library construction was performed according to previous studies [48], [84]. Sequencing reactions were performed using BigDye 3.1 (Applied Biosystems, Foster City, California, United States) and analyzed on an ABI3100 sequencer. Sequence reads were assembled using the PHRED/PHRAP/CONSED package [95], [96] and Sequencher 4.8 (Gene Codes Corporation). Additional analysis of the data was performed using BLASTn [91]. Based on the initial assembly of sequences, oligonucleotides were selected to close gaps in the sequence coverage by primer walking. Genes on the MAT locus of C. gattii were annotated based on homology to the existing annotation in C. gattii VGII (strain R265) and VGI (strains WM276 and E566), and C. neoformans (H99 and JEC21). Certain gene annotations were modified using the FGENESH program (http://linux1.softberry.com/berry.phtml). Comparison of the MAT locus alleles among VGI, VGII, and VGIII strains was performed using BLASTn. The BLASTn results were parsed using a PERL script and imputed into the ACT program to construct synteny diagrams of the MAT locus alleles among C. gattii strains. In order to evaluate size polymorphisms within the MAT locus, 14 MATα strains representing all SXI1α alleles and 8 MATa strains were amplified using primers complementary to conserved sequences. Primers used for fingerprinting are listed in Table S3. The PCR products were digested using appropriate enzymes selected on the basis of DNA sequences using NEBcutter version 2.0 (http://tools.neb.com/NEBcutter2/index.php). Fingerprints showing variability were sequenced to determine the nature of any size polymorphisms. Sequences from both forward and reverse strands were assembled with complete double strand coverage and manually edited using Sequencher version 4.8 (Gene Codes Corporation).
The pathogenic potential of C. gattii strains belonging to the VGIII genotype was tested in a murine model of pulmonary cryptococcosis [97]. Briefly, C. gattii strains were grown in YPD broth for 16 hours and washed with sterile physiological saline. Cells were counted with a hemocytometer and suspended at a concentration of 3.3×106 cells/ml. Five male BALB/c mice (approximately six-weeks old, 15–20 g; Charles River) in each group were first anesthetized with xylazine-ketamine mixture and then 30 ul of infectious inocula was gently dripped into their nares. The injected animals were observed for any overt signs of illness, and all morbid animals were promptly sacrificed by CO2 inhalation to minimize pain and suffering.
Data from all infected animals were used to determine Kaplan-Meyer survival curves using SAS software (SAS Institute, Inc., Cary, N.C.). For this statistical analysis with genotype information, the non-parametric Mann-Whitney U-test was applied, with further details in the statistical analysis section (http://faculty.vassar.edu/lowry/utest.html). Progression of disease was also determined by weighing infected animals once every alternate day at the start of the infection and then every day until they were moribund. One half portion of the lung and brain tissues from sacrificed animals was cultured on YPD and Staib's niger seed agar for the recovery of cells to determine that infections were cryptococcal in origin.
To characterize virulence properties, representative C. gattii strains from four virulence groups comprising high virulence strains (CA1499), moderate virulence strains (CA1292), low virulence strains (WM161), and avirulent strains (CA2339), were chosen for further intranasal infection per procedure described previously. A total of six mice were infected for each test strain and groups of three mice were sacrificed at one and two-weeks post infection. The lungs and brains were removed aseptically, one half of each organ was fixed in 10% buffered formalin and Bouin's fixative, processed into paraffin blocks, sectioned, and stained with hematoxylin and eosin (H & E) and Mayer's mucicarmine for histopathological examination. The other halves of the lungs and brains were weighed, homogenized, diluted in PBS, and plated on YPD agar. Colonies were counted after incubation of the plates at 30°C for 4 days, and the results were expressed as colony forming unit (CFU) per gram of infected tissue. Results from organ load experiment were analyzed by student t-test, with significance determined at P≤0.05.
The animal studies conducted were in full compliance with all of the guidelines set forth by the Wadsworth Center Institutional Animal Care and Use Committee (IACUC) and in full compliance with the United States Animal Welfare Act (Public Law 98–198). The Wadsworth Center IACUC approved all of the vertebrate studies. The studies were conducted in facilities accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC).
A proliferation assay was previously developed to monitor the intracellular proliferation rate (IPR) of individual strains (for a 72-hour period) following phagocytosis [49]. For this assay, J774 macrophage cells were exposed to cryptococcal cells that were opsonized with 18B7 antibody for 2 hr as described previously [98]. Each well was washed with phosphate-buffered saline (PBS) in quadruplicate to remove as many extracellular yeast cells as possible and 1 ml of fresh serum-free DMEM was then added. For time point T = 0, the 1 ml of DMEM was discarded and 200 µl of sterile dH2O was added into wells to lyse macrophage cells. After 30 minutes, the intracellular yeast were released and collected. Another 200 µl dH2O was added to each well to collect the remaining yeast cells. The intracellular yeast were then mixed with Trypan Blue at a 1∶1 ratio and the live yeast cells were counted. For the subsequent five time points (T = 18 hrs, T = 24 hrs, T = 48 hrs, T = 72 hrs), intracellular cryptococcal cells were collected and independently counted with a hemocytometer.
For each strain tested, the time course was repeated at least three independent times, using different batches of macrophages. The IPR value was calculated by dividing the maximum intracellular yeast number by the initial intracellular yeast number at T = 0. We confirmed that Trypan Blue stains 100% of the cryptococcal cells in a heat-killed culture, but only approximately 5% of cells from a standard overnight culture. Compared to a conventional colony counting method, this method was shown to be more sensitive in detecting the clustered yeast population or yeast cells undergoing budding.
Time to 50% lethality (LT50) and median IPR values were used to assess the statistical significance between the VGIII subgroups and these respective values. For this statistical analysis, the non-parametric directional Mann-Whitney U-test was applied and values of p<0.025 were considered as statistically significant (http://faculty.vassar.edu/lowry/utest.html). Regression analysis was used to measure the correlation between LT50 and IPR values, and an F-value of p<0.05 was considered to be a significant correlation, with R2, p values, and Pearson correlations also calculated and represented in the analysis.
All strains used in this study are listed in Table S6, including C. gattii, C. neoformans, and S. cerevisiae isolates.
|
10.1371/journal.pbio.0060301 | Senescence-Associated Secretory Phenotypes Reveal Cell-Nonautonomous Functions of Oncogenic RAS and the p53 Tumor Suppressor | Cellular senescence suppresses cancer by arresting cell proliferation, essentially permanently, in response to oncogenic stimuli, including genotoxic stress. We modified the use of antibody arrays to provide a quantitative assessment of factors secreted by senescent cells. We show that human cells induced to senesce by genotoxic stress secrete myriad factors associated with inflammation and malignancy. This senescence-associated secretory phenotype (SASP) developed slowly over several days and only after DNA damage of sufficient magnitude to induce senescence. Remarkably similar SASPs developed in normal fibroblasts, normal epithelial cells, and epithelial tumor cells after genotoxic stress in culture, and in epithelial tumor cells in vivo after treatment of prostate cancer patients with DNA-damaging chemotherapy. In cultured premalignant epithelial cells, SASPs induced an epithelial–mesenchyme transition and invasiveness, hallmarks of malignancy, by a paracrine mechanism that depended largely on the SASP factors interleukin (IL)-6 and IL-8. Strikingly, two manipulations markedly amplified, and accelerated development of, the SASPs: oncogenic RAS expression, which causes genotoxic stress and senescence in normal cells, and functional loss of the p53 tumor suppressor protein. Both loss of p53 and gain of oncogenic RAS also exacerbated the promalignant paracrine activities of the SASPs. Our findings define a central feature of genotoxic stress-induced senescence. Moreover, they suggest a cell-nonautonomous mechanism by which p53 can restrain, and oncogenic RAS can promote, the development of age-related cancer by altering the tissue microenvironment.
| Cells with damaged DNA are at risk of becoming cancerous tumors. Although “cellular senescence” can suppress tumor formation from damaged cells by blocking the cell division that underlies cancer growth, it has also been implicated in promoting cancer and other age-related diseases. To understand how this might happen, we measured proteins that senescent human cells secrete into their local environment and found many factors associated with inflammation and cancer development. Different types of cells secrete a common set of proteins when they senesce. This senescence-associated secretory phenotype (SASP) occurs not only in cultured cells, but also in vivo in response to DNA-damaging chemotherapy. Normal cells that acquire a highly active mutant version of the RAS protein, which is known to contribute to tumor growth, undergo cellular senescence, and develop a very intense SASP, with higher levels of proteins secreted. Likewise, the SASP is more intense when cells lose the functions of the tumor suppressor p53. Senescent cells promote the growth and aggressiveness of nearby precancerous or cancer cells, and cells with a more intense SASP do so more efficiently. Our findings support the idea that cellular senescence can be both beneficial, in preventing damaged cells from dividing, and deleterious, by having effects on neighboring cells; this balance of effects is predicted by an evolutionary theory of aging.
| Cancer is a multistep disease in which cells acquire increasingly malignant phenotypes. These phenotypes are acquired in part by somatic mutations, which derange normal controls over cell proliferation (growth), survival, invasion, and other processes important for malignant tumorigenesis [1]. In addition, there is increasing evidence that the tissue microenvironment is an important determinant of whether and how malignancies develop [2,3]. Normal tissue environments tend to suppress malignant phenotypes, whereas abnormal tissue environments such at those caused by inflammation can promote cancer progression.
Cancer development is restrained by a variety of tumor suppressor genes. Some of these genes permanently arrest the growth of cells at risk for neoplastic transformation, a process termed cellular senescence [4–6]. Two tumor suppressor pathways, controlled by the p53 and p16INK4a/pRB proteins, regulate senescence responses. Both pathways integrate multiple aspects of cellular physiology and direct cell fate towards survival, death, proliferation, or growth arrest, depending on the context [7,8].
Several lines of evidence indicate that cellular senescence is a potent tumor-suppressive mechanism [4,9,10]. Many potentially oncogenic stimuli (e.g., dysfunctional telomeres, DNA damage, and certain oncogenes) induce senescence [6,11]. Moreover, mutations that dampen the p53 or p16INK4a/pRB pathways confer resistance to senescence and greatly increase cancer risk [12,13]. Most cancers harbor mutations in one or both of these pathways [14,15]. Lastly, in mice and humans, a senescence response to strong mitogenic signals, such as those delivered by certain oncogenes, prevents premalignant lesions from progressing to malignant cancers [16–19]. Interestingly, some tumor cells retain the ability to senesce in response to DNA-damaging chemotherapy or p53 reactivation; in mice, this response arrests tumor progression [20–22].
Despite support for the idea that senescence is a beneficial anticancer mechanism, indirect evidence suggests that senescent cells can also be deleterious and might contribute to age-related pathologies [10,23–25]. The apparent paradox of contributing to both tumor suppression and aging is consistent with an evolutionary theory of aging, termed antagonistic pleiotropy [26]. Organisms generally evolve in environments that are replete with extrinsic hazards, and so old individuals tend to be rare in natural populations. Therefore, there is little selective pressure for tumor suppressor mechanisms to be effective well into old age; rather, these mechanisms need to be sufficiently effective only to ensure successful reproduction. Further, tumor suppressor mechanisms could in principle even be deleterious at advanced ages, as predicted by evolutionary antagonistic pleiotropy. Consistent with this view, senescent cells increase with age in mammalian tissues [27], and have been found at sites of age-related pathologies such as osteoarthritis and atherosclerosis [28–30]. Moreover, in mice, chronically active p53 both promotes cellular senescence and accelerates aging phenotypes [31,32].
How might senescent cells be deleterious? Senescent cells acquire many changes in gene expression, mostly documented as altered mRNA abundance, including increased expression of secreted proteins [33–41]. Some of these secreted proteins act in an autocrine manner to reinforce the senescence growth arrest [37,38,40,41]. Moreover, cell culture and mouse xenograft studies suggest that proteins secreted by senescent cells can promote degenerative or hyperproliferative changes in neighboring cells [35,39,42,43]. Thus, although the cell-autonomous senescence growth arrest suppresses cancer, factors secreted by senescent cells might have deleterious cell-nonautonomous effects that alter the tissue microenvironment. To date, a comprehensive analysis of the secretory profile of senescent cells is lacking, as is knowledge regarding how this profile varies with cell type or senescence inducer, or how it relates to the tumor suppressor proteins that control senescence.
To fill these gaps in our knowledge, we modified antibody arrays to be quantitative and sensitive over a wide dynamic range and defined the senescence-associated secretory phenotype (SASP). We show that this phenotype is complex, containing elements associated with inflammation and tumorigenesis, and is induced only by genotoxic stress of sufficient magnitude to cause senescence. SASPs are expressed by senescent human fibroblasts and epithelial cells in culture. Moreover, epithelial tumor cells exposed to DNA-damaging chemotherapy senesce and express a SASP in vivo. The arrays allowed us to identify two new malignant phenotypes promoted by senescent cells (the epithelial–mesenchyme transition and invasiveness), and the SASP factors responsible for them (interleukin [IL]-6 and IL-8). Strikingly, the SASP was markedly amplified by oncogenic RAS or loss of p53 function. Our results identify a mechanism by which p53 acts as a cell-nonautonomous tumor suppressor, and RAS as a cell-nonautonomous oncogene, and provide a novel framework for understanding how age-related cancers might progress.
To determine whether tissue of origin, donor age, or genotype affected secretory phenotypes, we first studied five human fibroblast strains, derived from embryonic lung (WI-38, IMR-90), neonatal foreskin (BJ, HCA2), or adult breast (hBF184). We cultured the cells under standard conditions, and either atmospheric (∼20%) O2 or 3% O2, which is more physiological [44]. We made presenescent (PRE) cultures (>80% of cells capable of proliferation) quiescent by growing the cells to confluence in order to compare them to nondividing senescent (SEN) cultures (<10% proliferative) (Table S1). We induced senescence by either repeatedly passaging the cells (REP, replicative exhaustion) or by exposing them to a relatively high dose (10 Gy) of ionizing radiation (XRA [X-irradiation]) (see Materials and Methods; Table S1; and Text S1).
To identify proteins secreted by PRE and SEN cells, we generated conditioned media (CM) by incubating each culture in serum-free medium for 24 h. After normalizing for cell number, we analyzed CM using antibody arrays designed to detect 120 proteins selected for roles in intercellular signaling (Table S2). We modified the detection protocol (see Materials and Methods and Text S1), thereby rendering the arrays linear over two to three orders of magnitude; accurate, as determined by concordance with enzyme-linked immunosorbent assays (ELISAs) of recombinant proteins; and reliable, as determined by comparing triplicate samples analyzed separately to pooled samples (Text S2). We quantified the signals, normalizing intensities to controls on the arrays to facilitate interexperiment comparisons, and calculated secreted protein levels as log2-fold changes relative to averages of all samples for each cell strain (baseline). We used these values for quantitative data analyses (Datasets S1–S4) and for visual display (Figure 1A). For the visual display, values over baseline are displayed in grades of yellow, and values under baseline are displayed in grades of blue (Figure 1A).
Although the display gives only a semiquantitative assessment of how secretion levels vary (see accompanying scale in Figure 1A, with log2-fold changes indicated), the data (Datasets S1–S4) show that SEN cells secrete significantly higher levels of numerous proteins compared to PRE cells (Figure 1A). We term this phenomenon a senescence-associated secretory phenotype (SASP). Of 120 proteins interrogated by the arrays, 41 were significantly altered in the SEN CM and were oversecreted in comparison to PRE CM (Figure 1A and Dataset S4). However, the SASPs did not result from a general stimulation of secretion. Seventy-nine proteins showed no significant differences in secreted levels between SEN and PRE cells, although many of these proteins were easily detectable by the arrays (Dataset S4).
The SASPs were complex, and their biological effects could not be predicted a priori. SASP components included inflammatory and immune-modulatory cytokines and chemokines (e.g., IL-6, −7, and −8, MCP-2, and MIP-3a). They also included growth factors (e.g., GRO, HGF, and IGFBPs), shed cell surface molecules (e.g., ICAMs, uPAR, and TNF receptors), and survival factors (Figure 1A and Table S2). However, the SASP was not a fixed phenotype. Rather, it was a wide-ranging profile because each cell strain also displayed unique quantitative or qualitative features. In addition, within strains, PRE cultures secreted higher levels of some factors in 20% versus 3% O2, and SEN cultures secreted higher levels of some factors in 3% versus 20% O2. Thus, although human cells are less sensitive than mouse cells to hyperphysiological O2 [44], human cells are not unaffected by the ambient O2 level. Nonetheless, the secretory phenotype was highly conserved (r > 0.75) in human senescent cells cultured in 3% versus 20% O2. By contrast, the ambient O2 level strongly affects the secretory phenotype of mouse senescent cells (J. P. Coppe, C. K. Patil, F. Rodier, A. Krtolica, S. Parrinello, et al., unpublished data).
We verified the secretion levels of several SASP proteins by ELISAs (Figure S1 and Text S1). Further, because secretion increased greater than 10-fold for some SASP factors, we could verify up-regulation by intracellular immunostaining. For example, IL-6 and IL-8 were barely visible in PRE cells but clearly detectable in SEN cells (Figure 1B, Figure S2, and Text S1). We performed the immunostaining on cells in 10% serum, which allowed us to rule out the possibility that the SASP was a senescence-specific response to the serum-free incubation needed to collect CM. Moreover, the SASPs of SEN cells induced by REP and XRA were highly correlated (r > 0.9; Figure 1C), indicating that the phenotype was not specific to one senescence inducer. The secretory profiles of fibroblast strains from the same tissues (e.g., BJ and HCA2 from neonatal foreskin; and IMR90 and Wi-38 from fetal lung) were highly correlated (Datasets S3 and S4). In subsequent figures, data from these related cell strains, as well as from REP and XRA samples from cells of the same type, were pooled and averaged in order to simplify the display.
Because REP and XRA induce senescence primarily by causing genomic damage (from telomere shortening and DNA breaks, respectively), we asked whether the SASP was a primary DNA damage response. We irradiated cells using either 0.5 or 10 Gy. As expected, both doses initiated a DNA damage response, as determined by p53 stabilization and phosphorylation (see Figure S3 and Text S1). However, cells that received 0.5 Gy transiently arrested growth for only 24–48 h before resuming growth, whereas cells that received 10 Gy underwent a permanent senescence growth arrest (Figure 1D). Antibody arrays performed on CM collected between 2 and 10 d after irradiation showed that only 10 Gy induced a SASP (Figure 1D and Figure S3). Moreover, cells that senesced owing to DNA damage developed the SASP slowly, requiring 4–7 d after irradiation before expressing a robust SASP. These findings indicate the SASP is not a DNA damage response per se. However, it is induced by DNA damage of sufficient magnitude to cause senescence, after which it requires several days to develop.
We also determined that proteins comprising the SASP were, in general, up-regulated at the level of mRNA abundance (Figure 1E and 1F, red symbols and line; Figure S4, and Text S1). However, for detectable proteins that showed little or no senescence-associated change in secretion, mRNA levels were a poor predictor of secreted protein levels (Figure 1F, blue symbols and line; and Figure S4). Thus, antibody arrays provide a more accurate assessment of the senescence-associated secretory signature than mRNA profiling.
To determine whether the SASP is limited to fibroblasts, we studied the secretory activity of epithelial cells. Normal human prostate epithelial cells (PrECs) underwent a classic senescence growth arrest in response to X-irradiation (see Table S1). We collected CM from PRE and SEN PrECs, and analyzed the factors secreted by these cells using antibody arrays. Normal PrECs expressed a robust SASP upon senescence (Figure 2A and Datasets S5–S8). Like fibroblasts, SEN PrECs secreted many factors at significantly higher levels compared to PRE PrECs. To compare the SASPs of normal human epithelial and stromal cells senesced under similar conditions, we analyzed factors that showed a significant change (p < 0.05) upon senescence in PrECs or fibroblasts induced by XRA only (Figure 2A). Using the hypergeometric distribution, we determined that the SASPs of both normal cell types overlapped highly significantly (p = ∼10−3; see Materials and Methods). The trend analysis of all 120 factors on the arrays (Figure 2B) demonstrated that the secretory profiles were correlated (r = 0.53) and also indicated >66% overlap between normal fibroblasts and normal epithelial cells. More specifically, both SASPs included inflammatory or immune factors such as IL-6, IL-8, or MCP-1, growth modulators such as GRO and IGFBP-2, cell survival regulators such as OPG or sTNF RI, and shed surface molecules such as uPAR or ICAM-1 (Figure 2A). Not surprising, there were also differences between the SASPs of fibroblasts and PrECs. In contrast to fibroblasts, three factors (Acrp30, BTC, and IGFBP-6) were significantly down-regulated by SEN in PrECs. Moreover, IL-1α or HGF were SASP factors unique to either normal epithelial SASP or normal fibroblast SASP, respectively. This result indicates that the SASP is not limited to normal stromal cells, and that a substantial overlap between normal senescent cells of different tissue origins exists.
Some tumor cells retain the ability to senesce in response to DNA damage, including DNA-damaging chemotherapy [20–22]. We therefore asked whether prostate cancer cells also developed a SASP. We studied three prostate tumor cell lines (BPH1 [45], RWPE1 [46], and PC3 [47], which differ in their degree of malignancy as follows: PC3 > RWPE1 > BPH1). As with normal epithelial cells (PrECs), we induced senescence by XRA, and analyzed CM using antibody arrays (Datasets S5–S8). SEN epithelial cells secreted significantly higher levels of numerous proteins compared to PRE counterparts (Figure 2C). The SASPs of prostate epithelial cells showed striking overlap between normal and transformed cells (Figure 2C and Dataset S8), and there was also striking similarities between the SASP of fibroblasts and all epithelial cells, transformed and not transformed (Figure 2C, asterisks indicate common secreted proteins, and Datasets S5–S8). Twenty-four proteins were shared between the SASPs of all fibroblasts (Figure 1A) and all epithelial cells (Figure 2C); this overlap was highly significant relative to the overlap predicted from chance (p = ∼10−5; see Materials and Methods). We conclude that normal fibroblasts and both normal and transformed epithelial cells can develop SASPs that significantly overlap, displaying many common, but also some distinct, features.
Many human tumor cells retain the ability to senesce, in culture and in vivo, in response to DNA-damaging chemotherapeutic agents [48,49]. Epithelial cell lines, as well as normal fibroblasts, underwent senescence in culture in response to mitoxantrone (MIT) (see Table S1), a topoisomerase 2β inhibitor that causes DNA breaks and is used to treat prostate cancer [50]. Antibody arrays (Figure 3A and Datasets S5–S8) and ELISAs for IL-6, IL-8, and GRO-α (Figure S1) showed that MIT induced a SASP that correlated well (r = 0.89) with the XRA-induced SASP (Figure 3A).
The finding that human prostatic tumor cells express a SASP in response to MIT in culture allowed us to determine whether MIT induced a SASP in vivo. We laser captured approximately 1,000 tumor epithelial cells in biopsies from human prostate cancer patients before MIT chemotherapy and in tissues removed after chemotherapy and prostatectomy [50]. By microscopic inspection, the captured cells were devoid of stromal cells and leukocytes. Since mRNA and secreted protein levels correlated well for significantly up-regulated SASP factors (Figure 1E and 1F and Figure S4), we used quantitative PCR to quantify mRNAs encoding senescence and proliferation markers and SASP factors.
After chemotherapy, most of the tumors contained significantly higher levels of p16INK4a and p21 mRNAs, which are typically up-regulated in senescent cells (Figure 3B). They also contained significantly lower levels of proliferation-associated mRNAs encoding cyclin A, MCM-3, and PCNA (Figure 3B). These results suggest that MIT induced tumor cells to senesce in vivo. Importantly, most of the tumors contained significantly higher levels of mRNAs encoding the SASP components IL-6, IL-8, GM-CSF, GRO-α, IGFBP-2, and IL-1β (Figure 3C). However, the levels of mRNA encoding IL-2, which is not a SASP component, did not significantly change on average (Figure 3D). These findings (summarized in Figure 3E) suggest that the SASP is not limited to cultured cells, but also occurs when human cells senesce in vivo.
The epithelial–mesenchymal transition (EMT) confers invasive and metastatic properties on epithelial cells, and is an important step in cancer progression that presages the conversion of carcinomas in situ to potentially fatal invasive cancers [51,52]. We found that the fibroblast SASP induced a classic EMT in two nonaggressive human breast cancer cell lines (T47D and ZR75.1). Secreted factors from SEN, but not PRE, fibroblasts caused dose-dependent epithelial cell scattering, a mesenchymal characteristic (Figure 4A). Moreover, immunostaining showed that PRE CM preserved surface-associated β-catenin and E-cadherin, and strong cytokeratin 8/18 expression (Figure 4B), and western analysis showed that PRE CM preserved low expression of vimentin (see below). These features are epithelial characteristics frequently retained by nonaggressive cells [51,52]. By contrast, CM from SEN cells markedly decreased overall and cell surface β-catenin and E-cadherin and reduced cytokeratin expression (Figure 4B), consistent with a mesenchymal transition. Further, SEN CM down-regulated the tight junction protein claudin-1, leaving the remaining protein localized primarily to the nucleus (Figure 4B), a hallmark of an EMT and a feature of metastatic but not primary tumors [53]. Finally, SEN CM increased vimentin expression (see below), another mesenchymal marker and hallmark of an EMT [52].
Consistent with a SASP-induced EMT, CM from SEN, but not PRE, cells stimulated premalignant MCF-10A and malignant T47D, ZR75.1, CAMA1, and HCC1187 cells to invade a basement membrane (Figure 4C), as well as MDA-MB-231 and MDA-MB-453 (data not shown and unpublished data). The antibody arrays guided us in identifying the highly secreted SASP components IL-6 and IL-8 as candidates for this activity [54,55]. Recombinant IL-6 and IL-8 added to PRE CM stimulated invasion to varying degrees depending on the epithelial line. Importantly, IL-6 and IL-8 blocking antibodies reduced the invasion stimulated by SEN CM (Figure 4C), indicating a substantial contribution from IL-6 and IL-8. These results support the idea that paracrine activities of the SASP can promote malignant phenotypes in nearby premalignant or malignant cells, and identify new SASP activities: the ability to induce an EMT and to invade a basement membrane.
Certain oncogenes, of which oncogenic RAS is the prototype [56], induce senescence in part by indirectly causing DNA damage [57,58]. RAS and related oncogenes are best known for generating mitogenic signals that promote cell-autonomous malignant phenotypes, although RAS-transformed cells are also known to secrete specific factors that contribute to tumorigenesis [59–61]. Our finding that SASPs can promote malignant phenotypes in nearby cells suggested that RAS-like oncogenes might also promote malignancy cell-nonautonomously via a complex SASP. To test this possibility, we expressed oncogenic RAS in fibroblasts and epithelial cells using lentiviruses, allowed the cells to senesce (Table S1), and analyzed CM using antibody arrays.
RAS induced a SASP that had both common and unique features relative to SASPs induced by REP or XRA (Figure 5A–5E and Datasets S9–S12). The RAS-induced SASP subsumed a subset of proteins that showed increased secretion upon REP- or XRA-induced senescence (Figure 5A). To simplify the visual comparison, we averaged the highly correlated data (Figure 5A and 5C) from cells originating from the same tissue (WI-38 + IMR90 from embryonic lung, and BJ + HCA-2 from neonatal foreskin), and from cells induced to senesce by REP or XRA (see Datasets S9–S12 for details of the averaging, and the raw and processed data). Overall, there was good correlation between the SASPs of fibroblasts induced to senesce by RAS, XRA, or REP (Figure 5C). Correlations between SEN (XRA or REP) (averaged) and SEN(RAS) were 0.75 for WI-38 and IMR-90 fibroblasts (averaged) and 0.84 for HCA-2 fibroblasts.
A striking feature of the RAS-induced SASP was that all fibroblasts induced to senesce by RAS secreted multiple proteins at levels significantly and dramatically higher than other SEN cells. Because the visual display (Figure 5A) is only semiquantitative, the quantitative nature of the amplified SASP is best illustrated by the bar graph (Figure 5D), which plots the log2-fold increases in factors secreted by SEN(RAS) cells compared to their SEN (REP or XRA) counterparts (nine proteins were significantly more secreted in SEN(RAS) versus SEN(REP;XRA) across all fibroblasts). In addition, the RAS-induced SASPs had unique features because this SASP included five proteins that were not secreted at significantly elevated levels by other SEN (REP or XRA) cells (Figure 5E). We refer to the overall secretory response of cells induced to senesce by RAS, including the quantitative increase in secretion of specific proteins and the secretion of proteins not present in REP or XRA SASPs, as the amplified SASP. We confirmed the robust SASP induced by RAS by immunostaining (Figure 5B and Figure S2) and ELISAs (Figure S1). Further, we confirmed that oncogenic RAS induced an amplified SASP in prostate epithelial cells (Figure 5F and 5G).
Taken together, these results suggest that oncogenic RAS, despite inducing a tumor-suppressive senescence arrest, might promote tumorigenesis in a cell-nonautonomous manner by inducing an amplified SASP.
The p53 pathway is important for establishing and maintaining the senescence growth arrest caused by genotoxic stress, although cells that lack p53 can undergo senescence providing they express p16 [62]. We therefore asked whether p53 established or maintained the SASP. To inactivate p53, we expressed genetic suppressor element 22 (GSE22, also designated GSE), a peptide that prevents p53 tetramerization and causes inactive monomeric p53 to accumulate (detectable by immunostaining, Figure S2) [63]. We obtained similar results using a short hairpin RNA (shRNA) that reduces p53 expression by RNA interference. We induced WI-38 fibroblasts to senesce by REP or XRA and then inactivated p53. Because WI-38 and IMR90 cells senesce with high levels of p16INK4a, they do not resume proliferation when p53 is inactivated [62]. To simplify the visual comparison of antibody array readouts, we averaged data from highly correlated samples, as described for Figure 5 (see Datasets S13–S16 for details of the averaging, and the raw and processed data). The SASPs of SEN WI-38 in which p53 was either wild type or inactivated after senescence were similar by visual display (Figure 6A, compare row 1 with row 11), and by the graphical plot of the log2-fold changes that occurred in specific factors (Figure 6B, green bars showing variations obtained using the appropriate p53 wild-type baseline (e.g., SEN(REP>GSE) versus SEN(REP) in WI-38) and Datasets S13–S16). This finding indicates that p53 is not required to maintain an established SASP.
To determine whether p53 is needed to initiate a SASP, we inactivated p53 in PRE WI-38 cells and then induced senescence by XRA, REP, or RAS (Figure 6A, rows 8–10). p53 inactivation did not induce a SASP in PRE cells (Figures 6A, rows 5–7, Figures S1 and S2, and Datasets S13–S16). Upon senescence by either REP, XRA, or RAS, however, p53-deficient cells not only developed a SASP, but the magnitude of the SASP was markedly enhanced (Figure 6A, rows 8–10 versus 1–4), similar to the amplified SASP induced by RAS. The quantitative effect of p53 deficiency on SASPs is best illustrated by the bar graph, which plots the log2-fold increases of significantly altered factors secreted by cells made p53 deficient and then induced to senesce, compared to cells with wild-type p53 and induced to senesce (Figure 6C, red and grey bars). We confirmed the robust SASP by immunostaining (Figure 6D and Figure S2) and ELISA (Figure S1). Together, these findings indicate that p53 is not required to initiate the SASP, and further that it restrains development of an amplified SASP.
Importantly, the combined loss of p53 and gain of oncogenic RAS resulted in the most amplified SASP (Figure 6A, rows 9–10 versus rows 1–4; Figure 6E and 6F, Figure S5A, and Text S1). In addition, when p53 was inactivated prior to XRA, the SASP developed much earlier—between 2 and 4 d after irradiation (Figure 6G and Figures S5B and S5C), compared to 4–7 d in cells with wild-type p53 (Figure 1D). Interestingly, cells induced to senesce by RAS also developed the amplified SASP earlier—within 2–4 d after irradiation (Figures 6G, Figures S5B, S5C, and S6, and Text S1). Thus, loss of the p53 tumor suppressor, or gain of oncogenic RAS, not only amplified the SASP, but also accelerated its development.
We also inactivated p53 in fibroblasts that senesce with low p16INK4a levels: HCA2, BJ, and WI-38 expressing a shRNA (shp16) that reduces p16INK4a expression by RNA interference. In these cells, whether the cells were induced to senesce by REP or XRA, the SASP was also markedly amplified (Figure 6A, rows 1–4 versus rows 11–14). The quantitative outcome of p53 loss on established SASPs is best illustrated by the bar graph, which lists the significantly altered factors secreted by cells made to senesce and then induced to lose p53 function, compared to cells induced to senesce and keeping a functional p53 (Figure 6B, blue and pink bars). As reported [62], p53 inactivation reversed the growth arrest of these cells, and the reverted cells resumed growth (Figure S6). We refer to these cells as REV, and the reversal of the growth arrest verified the efficacy of the p53 inactivation. These findings indicate that, once established, the SASP cannot be suppressed despite reversion of the senescence growth arrest. Together, the results indicate that p53 activation by genotoxic stress not only restrains cell proliferation, but also restrains the SASP.
The SASPs of p53-deficient cells were qualitatively similar to those of SEN cells with wild-type p53, resulting in tightly clustered profiles (Figure 6F). The main influence of p53 status was quantitative. As was the case for RAS-induced senescence, a subset of SASP proteins was secreted at 5- to 30-fold higher levels after p53 inactivation (Figure 6B, 6C, and 6E, and Datasets S13–S16). However, there were also unique features of the p53-deficient SASP (Figure 6B and 6C, bottom). Interestingly, many factors that were further or uniquely up-regulated by cells made senescent by RAS were amplified in a similar fashion in p53-deficient cells (Figures 6B, 6C, 6E, 5D, and 5E).
p53 also restrained the SASP in prostate epithelial cells. Factors identified as part of the epithelial SASP (Figure 2A and 2C) were amplified in the PC3, BPH1, and RWPE1 cancer cells, which are p53 deficient, compared to normal PrECs, which have wild-type p53 (Figure 6H, top cluster). In addition, p53-deficient epithelial cells that underwent senescence oversecreted most factors that were restrained by p53 in normal fibroblasts (compare Figure 6H, bottom cluster versus listed factors in Figure 6B and 6C). Taken together, these findings indicate that fibroblasts and epithelial cells that are induced to senesce by genotoxic stress develop a SASP that is restrained by p53 activity.
To determine the possible biological consequences of the amplified SASPs, we compared the ability of CM from fibroblasts with unamplified or amplified SASPs to induce an EMT and invasiveness in relatively nonaggressive human cancer cells. CM from cells with an amplified SASP was significantly more potent than CM from SEN cells with wild-type p53 at inducing an EMT, as determined by cell scattering (compare Figure 7A with Figure 4A), immunostaining (compare Figure 7A with Figure 4B), and robust expression of vimentin (Figure 7A), an important quantitative marker of the EMT [52]. These comparisons were made on cells from the same single experiment. Further, the amplified SASP was significantly more potent at stimulating cancer cell invasiveness (Figure 7B). Senescent cells have been shown to stimulate the growth of premalignant or malignant epithelial cells [42,43]. We found the amplified SASP stimulated this epithelial cell growth to a significantly greater extent than the unamplified SASP (Figure 7C). These findings support the idea that p53 restrains the cell-nonautonomous promalignant activities of the SASP.
We identified a hallmark of cellular senescence—the senescence-associated secretory phenotype or SASP—that confers cell-nonautonomous paracrine functions on cells, and is markedly exacerbated by gain of oncogenic RAS or loss of p53 function.
To study the SASP, we modified a commercially available antibody array protocol, substituting radioactivity for chemiluminescence as a final detection method. This modification greatly improved the dynamic range of the arrays, and rendered them highly quantitative, reliable, and accurate. Using this modified array protocol, we were able to study both qualitative and quantitative aspects of the SASP, and compare similarities and differences among individual donors, cell types, and tissues of origin. Importantly, we uncovered quantitative differences caused by oncogenic RAS or loss of p53 function.
The SASP was a general feature of senescent fibroblasts from different tissues, donors, and donor ages, as well as prostate epithelial cells, both normal and transformed. There were distinct quantitative and qualitative differences among the different cell strains and lines, as expected given their different genotypes and tissue origins, indicating that the SASP is not an invariant phenotype. However, the striking feature of the phenotype was the marked similarities among the SASPs from diverse donors, cell types, and tissues, suggesting the existence of a conserved core secretory program that any cell undergoing senescence would trigger. Notably, all the SASPs featured high levels of secreted inflammatory cytokines, immune modulators, and growth factors, suggesting that SASPs might have myriad biological activities in addition to those we describe here. Many SASP factors were up-regulated at the level of mRNA abundance, suggesting that the phenotype may be controlled transcriptionally.
The correspondence between mRNA levels and SASP factors allowed us to probe human biopsy samples for the expression of SASP components before and after DNA-damaging chemotherapy. Our results showed that human tumor cells very likely undergo senescence in response to DNA damaging chemotherapy in vivo, as reported for mice [49]. Moreover, human tumor cells very likely express a SASP after chemotherapy. We speculate that components of chemotherapy-induced SASPs, particularly the high levels of inflammatory cytokines, might contribute to the debilitating effects of DNA-damaging chemotherapy. These SASPs might also fuel development of secondary cancers by creating a local tissue environment that is permissive for the growth and progression of cells that acquire therapy-induced mutations, and fail to senesce or die.
Senescent human fibroblasts have been shown to stimulate the proliferation of premalignant and malignant epithelial cells in culture, and the tumorigenicity of premalignant epithelial cells in mouse xenografts [35,42,43]. However, the mechanisms responsible for these stimulatory activities are incompletely understood. We identified two new biological activities of SEN cells mediated by the SASP: the ability to induce an EMT in relatively nonaggressive carcinoma cells, and the ability to stimulate their invasion through a basement membrane. The antibody arrays allowed us to identify two SASP factors, IL-6 and IL-8, which explained much of these two biological activities. In addition to stimulating an EMT and invasiveness, IL-6 and IL-8 promote inflammation, as do other SASP components. Further, senescent cells secreted or shed cytokine receptors, which could act as decoys and allow nearby premalignant or malignant cells to avoid immune surveillance. As they persist in tissues, senescent cells likely create a proinflammatory tissue environment, which is known to be protumorigenic [3,64]. Taken together, our findings support the idea that senescent cells can create a tissue microenvironment that promotes multiple stages of tumor evolution.
Recent findings show that tumors induced to senesce in mice gradually regress [21,22], owing perhaps to infiltration by cells of the innate immune system [22]. Inflammatory cytokines and chemokines, such as IL-6, IL-8, GRO-α, MCP-1, or GM-CSF, which are core features of the SASP, might contribute to this infiltration and eventual clearance. Why, then, are senescent cells found with increasing frequency during aging and at sites of age-related pathology? Some SEN cells might be refractory to immune clearance either because they are intrinsically different or they produce higher levels of factors that promote immune evasion. Alternatively, aging or age-related pathologies may dampen immune responses or increase the rate at which senescent cells are produced. Whatever the case, there is mounting evidence that senescent cells increase with age [65–68] and that chronic inflammation is a prominent feature of aging [69]. If the senescence response is an example of antagonistic pleiotropy, the senescent microenvironment created by SASPs might contribute to degenerative diseases of aging, such as osteoarthritis or atherosclerosis [28–30], in which senescent cells are found, as well as fuel the development of late-life cancers.
The evolutionary theory of antagonistic pleiotropy provides an explanation for the apparent dilemma of how the senescence response, or any biological process, can be both beneficial and deleterious, depending on the age of the organism. It is now recognized that aging is a consequence of the declining force of natural selection with age [26,70]. This decline is due to the high mortality caused by extrinsic hazards in natural environments, resulting in the relative scarcity of older individuals. Thus, natural selection can favor a trait that contributes to early life fitness (e.g., protection from cancer), even if that trait is deleterious in older individuals (e.g., promoting cancer development). We speculate that both the growth arrest and the secretory phenotype of senescent cells can be both beneficial and deleterious.
The senescence-associated growth arrest is beneficial because it arrests the growth of cells at risk for neoplastic transformation (cell-autonomous tumor suppressor function). It can be deleterious, however, because an accumulation of nondividing senescent cells can diminish the ability of renewable tissues to repair or regenerate. Although some aged tissues contain less than one or only a few percent of senescent cells [29,66,71], others can accumulate as many as 15% senescent cells [65,72]. Likewise, the senescence-associated secretory phenotype might have both beneficial and deleterious effects. The SASP can be beneficial because some SASP components reinforce the senescent growth arrest by an autocrine cytokine network [37,38,40,41], thereby contributing to maintenance of the senescence growth arrest. In addition, many SASP components are predicted to stimulate tissue repair and regeneration, and act as “danger signals” within the vicinity of tissues or systemically at the organism level. Thus, cells undergoing senescence may initially signal tissue damage, and initiate tissue repair via the SASP. Such effect would be the beneficial cell-nonautonomous function of cellular senescence. When chronically present, however, the secretory activity of senescent cells may be deleterious, disrupting normal tissue structure and function, and eventually stimulate age-associated tissue degeneration or promote malignant phenotypes (e.g., cancer progression, as described here).
Oncogenic RAS induced a SASP that was more robust than other senescence inducers, even when p53 function was intact. Oncogenic RAS is a cell-autonomous driver of cell proliferation in many cancer cells. In normal cells, however, oncogenic RAS causes genotoxic stress and senescence [57,58], inducing a SASP and thereby conferring complex cell-nonautonomous oncogenic activities. Thus, oncogenes such as RAS, which are known to activate protumorigenic paracrine mechanisms during transformation [59–61], might also exert cell-nonautonomous protumorigenic effects through nontransformed cells during the process of inducing senescence (Figure 7D).
How does oncogenic RAS induce a SASP? One possibility is that this activity of RAS is the result of the genotoxic stress caused by RAS-stimulated hyperproliferation. Alternatively, oncogenic RAS might induce a SASP more directly by stimulating the MAP kinase or other signaling pathway. Whatever the case, many aspects of the SASP induced by RAS resembled the SASP of p53-deficient cells.
Genotoxic stress sufficient to cause senescence both activates p53 and stimulates a SASP. Our data indicate a dual role for p53 (Figure 7D). First, in responding to genotoxic stress, p53 imposes the senescence growth arrest, consistent with its role as a cell-autonomous tumor suppressor. Second, p53 restrains the SASP because loss of p53 function, in combination with senescence-causing damage, greatly amplifies the SASP. p53 might restrain the SASP in part by rapidly arresting growth after cells experience DNA damage (unpublished data), thereby preventing the accumulation of further damage that could ensue should cells attempt to replicate the damaged DNA template. Additionally, p53 optimizes DNA repair, so cells that lack p53 might accumulate more DNA damage than cells with wild-type p53, which in turn might result in a more robust (amplified) SASP. Thus, the p53 tumor suppressor may act as an early sensor of oncogenic stress, and ultimately operate as a molecular catalyst preventing tissue inflammation. The strong correlation between DNA damage and development of a SASP suggests the SASP might be activated by the mammalian DNA damage response (DDR). Indeed, our preliminary data suggest that some components of the DDR are important for establishing and maintaining the SASP (F. Rodier, J-P. Coppé, C. K. Patil, W. A. M. Hoeijmakers, D. P. Muñoz, et al., unpublished data). However, the SASP does not develop immediately after DNA damage and therefore is not a simple or classic DDR. Rather, the SASP is a slow and persistent response to severe or irreparable damage of sufficient magnitude to cause senescence.
The persistence of the SASP might have important biological consequences. For example, cells that express low p16INK4a levels (e.g., SEN(REP) or SEN(XRA) HCA2) senesce in response to severe damage by activating the p53 pathway; when p53 is subsequently inactivated in these cells, they resume proliferation [62], but do not lose the SASP. Moreover, they eventually amplify the SASP as they acquire additional damage owing to proliferation in the absence of a functional checkpoint. Proliferating p53-deficient cells that senesced in response to genotoxic stress also developed a highly amplified secretory phenotype. These cells are at greater risk for escaping senescence (unpublished data) and would pose a danger to the tissue, not only by virtue of their proliferation, but also by virtue of their amplified SASP. Moreover, human cells that bypass oncogene-induced senescence [58], as well as cells in some human premalignant lesions [73,74], show signs of a persistently activated DDR. It is possible, if not likely, that these cells also express a SASP and therefore greatly increase the risk of cancer progression in vivo. By restraining the SASP, p53 acts as a cell-nonautonomous tumor suppressor, dampening the protumorigenic activities of the SASP. This activity might explain why a p53-deficient stroma promotes epithelial cancer progression [75,76]. We therefore propose that, in addition to its cell-autonomous ability to suppress cancer by inhibiting cell growth, p53 might further suppress cancer by restraining development of an inflammatory tissue milieu caused by a SASP.
Our broad, quantitative assessment of factors secreted by senescent cells revealed a highly complex secretory phenotype. We show here that this phenotype can promote cellular behaviors associated with malignancy, and suggest that cells that acquire mutations such as those that inactivate p53 and/or activate RAS functions can be particularly malignant owing to the paracrine activities of the SASP. It is very likely, though, that additional consequences of the SASP will be uncovered as the many SASP components are tested for specific activities.
Cells were obtained, cultured, and made quiescent or senescent as described in Text S1 [42,66].
Cultures were washed and incubated in serum-free Dulbecco's modified Eagle medium (DMEM) for 24 h to generate CM, which was collected and cells counted. CM was filtered (0.2 μm pore), frozen at −80 °C, and analyzed using antibody arrays (RayBiotech or Chemicon; Human cat #AA1001CH-8; Mouse cat #AA1003M-8) essentially as per the manufacturer's instructions. Briefly, CM was thawed and concentrated 2- to 3-fold at 4 °C (3 kDa cutoff). Volumes equivalent to 2 × 105 cells were diluted to 1.2 ml with DMEM and mixed with 300 μl of blocking solution. Array membranes were preincubated with 1.5 ml of blocking solution, incubated with CM mixture (overnight, 4 °C), washed 5×, then incubated with biotin-conjugated antibody cocktail (1 h 45 min, room temperature). After five washes, detection solution containing 0.265 μCi 35S-streptavidin (732 Ci/mmol; 0.1 mCi/ml) in blocking solution was added (1 h 45 min, room temperature), followed by five washes. Radioactivity bound to the filters was detected and quantified using a phosphorimager. Signals were analyzed as described in Text S2.
Assays were performed as described [77,78], using kits and antibodies described in Text S1.
Recombinant proteins and blocking antibodies were obtained as described in Text S1. Vectors to express oncogenic RAS (Ha-RASv12), TIN215C, and GSE22 were described [62,77,79].
Patients with high-risk localized prostate cancer enrolled and treated on a phase I–II clinical neoadjuvant chemotherapy trial at the Oregon Health & Science University, Portland VA Medical Center, Kaiser Permanente Northwest Region, Legacy Health System, and University of Washington [50]. Patients provided signed informed consent. From each patient, prostate biopsies were obtained prior to chemotherapy. At the time of radical prostatectomy following chemotherapy, cancer-containing tissue samples were obtained and frozen. Frozen sections were processed as described in Text S1. Cancerous epithelium from pretreated biopsy and posttreated prostatectomy specimens were captured separately and histology of acquired cells verified by review of hematoxylin and eosin (H&E)-stained sections from each sample and review of the laser confocal microscopy (LCM) images.
RNA was isolated from cultured or laser-captured cells and analyzed as described in Text S1.
Correlation coefficients were evaluated using Pearson correlation. Statistical significance between distributions of protein or mRNA signals was evaluated using a Student t-test with two tails, and an assumption of equal variance. For determination of the significance of overlap between epithelial and fibroblast SASPs, we used the hypergeometric distribution with the following parameters: population size = 120 (total proteins on the array), sample size = 41 (fibroblast SASP; see Figure 1A), successes in population = 39 (epithelial SASP; see Figure 2C), and successes in sample = 24 (overlap between the fibroblast and epithelial SASPs; see Figure 2C, asterisks). The same statistical analysis was used to compare SEN(XRA) normal epithelial cells (PrECs) versus SEN(XRA) normal fibroblasts (the following parameters were used: 120, 29, 25, and 12; see Figure 2A and Results).
|
10.1371/journal.pgen.1007984 | A missense variant in FTCD is associated with arsenic metabolism and toxicity phenotypes in Bangladesh | Inorganic arsenic (iAs) is a carcinogen, and exposure to iAs via food and water is a global public health problem. iAs-contaminated drinking water alone affects >100 million people worldwide, including ~50 million in Bangladesh. Once absorbed into the blood stream, most iAs is converted to mono-methylated (MMA) and then di-methylated (DMA) forms, facilitating excretion in urine. Arsenic metabolism efficiency varies among individuals, in part due to genetic variation near AS3MT (arsenite methyltransferase; 10q24.32). To identify additional arsenic metabolism loci, we measured protein-coding variants across the human exome for 1,660 Bangladeshi individuals participating in the Health Effects of Arsenic Longitudinal Study (HEALS). Among the 19,992 coding variants analyzed exome-wide, the minor allele (A) of rs61735836 (p.Val101Met) in exon 3 of FTCD (formiminotransferase cyclodeaminase) was associated with increased urinary iAs% (P = 8x10-13), increased MMA% (P = 2x10-16) and decreased DMA% (P = 6x10-23). Among 2,401 individuals with arsenic-induced skin lesions (an indicator of arsenic toxicity and cancer risk) and 2,472 controls, carrying the low-efficiency A allele (frequency = 7%) was associated with increased skin lesion risk (odds ratio = 1.35; P = 1x10-5). rs61735836 is in weak linkage disequilibrium with all nearby variants. The high-efficiency/major allele (G/Valine) is human-specific and eliminates a start codon at the first 5´-proximal Kozak sequence in FTCD, suggesting selection against an alternative translation start site. FTCD is critical for catabolism of histidine, a process that generates one-carbon units that can enter the one-carbon/folate cycle, which provides methyl groups for arsenic metabolism. In our study population, FTCD and AS3MT SNPs together explain ~10% of the variation in DMA% and support a causal effect of arsenic metabolism efficiency on arsenic toxicity (i.e., skin lesions). In summary, this work identifies a coding variant in FTCD associated with arsenic metabolism efficiency, providing new evidence supporting the established link between one-carbon/folate metabolism and arsenic toxicity.
| Chronic exposure to arsenic through food and drinking water is a serious global health issue, as arsenic can increase risk for cancer, cardiorespiratory diseases, and other chronic conditions. Ingested arsenic absorbed into the blood stream is metabolized (through reduction and methylation reactions) in order to facilitate excretion in urine and removal from the body. Individuals differ with respect to the efficiency of this metabolism, in part due to inherited genetic variation. The only region of the genome known to contain variation that impacts arsenic metabolism efficiency is 10q24.32, and these variants likely alter the function of the nearby gene AS3MT (arsenite methyltransferase). In order to identify new genetic variants that affect arsenic metabolism, we measured variation in protein-coding regions across the entire genome for >4,800 individuals with varying levels of exposure to arsenic through naturally-contaminated drinking water in Bangladesh. Using this data, we identified a variant in the FTCD gene (formiminotransferase cyclodeaminase) that is associated with arsenic metabolism efficiency and risk for arsenic-induced skin lesions. This genetic variant alters the FTCD amino acid sequence, potentially disrupting a cryptic protein translation start site in exon 3. FTCD codes for an enzyme involved in histidine catabolism and one-carbon/folate metabolism; thus, our result provides new evidence supporting the well-established hypothesis that the folate/one-carbon cycle plays an important role in arsenic-related disease.
| Exposure to inorganic arsenic (iAs) through consumption of contaminated drinking water is a major global health problem. Over 130 million individuals worldwide are exposed at levels >10 μg/L, including ~50 million in Bangladesh, where natural contamination of ground water is a well-known public health issue [1]. Arsenic is a human carcinogen [2], and chronic exposure to iAs through drinking water exceeding 50–100 μg/L is associated with various types of cancer in multiple populations [3,4] including the United States [5]. Arsenic exposure has also been linked to diabetes [6], cardiovascular disease [7], non-malignant lung disease [8], and overall mortality [9]. Arsenic-induced skin lesions are an early sign of arsenic exposure and toxicity [10] and are a risk factor for subsequent cancer [11].
Once absorbed into the blood stream, iAs can be converted to mono-methylated (MMA) and then di-methylated (DMA) forms of arsenic, with methylation facilitating the excretion of arsenic in urine [12]. This metabolism is believed to occur primarily in the liver [13]. The relative abundance of these arsenic species in urine (iAs%, MMA%, DMA%) varies across individuals and represents the efficiency with which an individual metabolizes arsenic. Arsenic metabolism is influenced by lifestyle and demographic factors [14], as well as inherited genetic variation. Prior genome-wide association (GWA) [15,16], linkage [17], and candidate gene studies [18] have shown that variation in the 10q24.32 region near the AS3MT gene (arsenite methyltransferase) influences arsenic metabolism efficiency, with two independent association signals observed in this region among exposed Bangladeshi individuals. These metabolism-related single nucleotide polymorphisms (SNPs) appear to impact the production of DMA (not the conversion of iAs to MMA) [14], and DMA%-increasing alleles are also associated with reduced risk for arsenic-induced skin lesions via a SNP-arsenic (i.e., gene-environment, GxE) interaction [16].
Other than 10q24.32/AS3MT, we currently know of no other regions of the human genome that contain variants that show robust and replicable evidence of association with arsenic metabolism efficiency [14], although studies of heritability suggest that additional variants are likely to exist [19,20]. In order to identify additional genetic variants that influence arsenic metabolism efficiency, we conducted a whole-exome study of associations between nonsynonymous, protein coding variation and arsenic metabolism efficiency.
Using DNA from individuals participating in HEALS (Health Effects of Arsenic Longitudinal Study), we conducted exome-wide association analyses for each of the three major arsenic species measured in urine, using percentages of total arsenic as our primary phenotypes (iAs%, MMA%, and DMA%). For this analysis, we restricted to 1,660 genotyped HEALS participants (among 2,949 HEALS participants with Illumina exome array data) with available data on arsenic species in urine. After SNP QC (see methods), we had data on 19,992 variants with MAF >1%, and ~90% of these were missense variants. Among these SNPs, rs61735836 (chr21:47572637 based on hg19) showed a clear association with all three arsenic species percentages (Fig 1A–1C). P-values for this association were P = 8x10-13 for iAs%, P = 2x10-16 for MMA%, and P = 6x10-23 for DMA%. The minor allele (A) was associated with decreased DMA% and increased MMA% and iAs% (Fig 1D–1F), consistent with the directions of association previously observed for SNPs in the AS3MT region. Results for all 19,992 variants are in Supporting Files S1-S3.
Like AS3MT, this association was most relevant to the second methylation step, as it showed a strong association with the secondary methylation index (SMI = DMA/MMA), but not the primary methylation index (PMI = MMA/iAs) (S1 Table). Similarly, after applying principal components (PC) analysis to arsenic species percentages as previously described [14], rs61735836 showed strong association with PC1 (representing production of DMA) but not PC2 (representing conversion of iAs to MMA) (S1 Table). Individuals carrying two minor alleles (AA) as compared to one (AC) appear to have even lower DMA%, suggesting a potential additive effect of the A allele; however, our sample size of minor allele homozygotes was small (n = 12), limiting our ability to examine differences between these two groups (S1 Table). The association of rs61735836 with arsenic species was similar across groups stratified by sex and age (S2 Table), and rs61735836 did not show evidence of interaction with either of the AS3MT SNPs previously identified in this population (rs9527 and rs11191527) in relation to DMA% or skin lesions status (S3 Table). The probe intensity data for rs61735836 is shown in S1 Fig, with very distinct clusters indicating high-quality data for this SNP.
We then conducted exome-wide association analyses of arsenical skin lesion status (the most common sign of arsenic toxicity) using data on 2,401 cases and 2,472 lesion-free controls (from both HEALS and BEST, the Bangladesh Vitamin E and Selenium Trial). While there was no notable departure from the expected null distribution, the low-efficiency allele for FTCD SNP rs61735836 (A) was associated with increased skin lesion risk (per allele OR = 1.25; P = 5x10-4; risk allele carrier OR = 1.35, P = 1x10-5) (S2 Fig). Results for all 19,992 variants are in S4 File. This observation is similar to what has been observed for metabolism-related variants in the AS3MT region and suggests rs61735836 impacts arsenic toxicity risk through its impact on arsenic metabolism efficiency. In this manner, this variant would be expected to reduce urinary arsenic elimination and thereby increase the internal or biologically effective dose of arsenic.
The MAF for rs61735836 was 0.077 in our data, highly consistent with the MAFs of 0.064 and 0.079 observed in the 1,000 Genomes Project (1KG) Bangladesh (BEB) population and South Asian (SAS) super-population, respectively. The MAF for this variant is less than <21% in all human populations with available data in the Geography of Genetic Variants browser [21] and is most common in East Asian populations (S3 Fig).
After combining our exome array results with our previously reported GWA results for genome-wide SNPs [15,16] (HumanCytoSNP-12 array imputed to ~8.2 million SNPs using 1KG phase 3 v5), we observed that rs61735836 is the only variant in this region showing strong evidence of association (Fig 2). This is consistent with the observation that rs61735836 is not in linkage disequilibrium (LD) (r2>0.1) with any nearby variant in 1KG South Asian (SAS) populations. This SNP is in mild LD with nearby variants in the 1KG African (AFR) super-population (r2~0.27) (S4 Fig), with the strongest LD observed in the ESN (Esan in Nigeria) population (r2 = 0.43 with rs184976755). SNP rs61735836 was not genotyped in our prior GWA study [15,16], and therefore could not be imputed due to the lack of LD with nearby variants. Among the 5 additional exonic variants in FTCD that passed QC (all missense), none showed association with any of our arsenic species measures (P>0.01).
Using data from HEALS, we tested rs61735836 for evidence of interaction with baseline arsenic exposure in relation to risk for arsenic-induced skin lesions (which were primarily incident lesions diagnosed after baseline). As an exposure measure, we used the arsenic concentration measured in the drinking well that each individual reported as their primary water source at baseline (prior to arsenic mitigation efforts in the HEALS cohort [22]). A test of multiplicative interaction produced a non-significant sub-multiplicative interaction estimate (OR = 0.86, P = 0.42), while a test of additive interaction produced a non-significant supra-multiplicative interaction (RERI = 0.49; P = 0.10) (Table 1).
To further assess the impact of rs61735836 on arsenic metabolism, we obtained data on arsenic species in blood (as opposed to urine) for 155 of our genotyped HEALS cohort members. These HEALS participants had existing data on arsenic metabolites in blood due to their participation in additional arsenic-related studies focused on folic acid and/or creatinine supplementation [23,24] and oxidative stress [25]. Consistent with our observed association with arsenic species in urine, the minor allele of rs61735836 (A) showed evidence of association with decreased DMA% (P = 0.02), increased MMA% (P = 0.41), and increased iAs% (P = 0.02), with arsenic species measured prior to any intervention (S4 Table). Among these 155 participants, 109 also had data on arsenic species in blood collected 12 weeks after the start of a supplementation intervention. Under the assumption that the interventions do not modify the impact of rs61735836 on arsenic metabolism efficiency (an assumption we make with considerable uncertainty), we can also examine these associations using these post-intervention measures. Using a mixed-effects model to analyze data from both time points, we observed that the A allele is associated with decreased DMA% (P = 0.005), increased MMA% (P = 0.01), and increased iAs% (P = 0.15) (S4 Table), consistent with results based on arsenic species measured in urine.
SNP rs61735836 resides in exon 3 of FTCD (Formiminotransferase cyclodeaminase), a gene predominantly expressed in liver [26,27] (S5 Fig), the tissue in which the majority of arsenic metabolism is believed to occur [13]. FTCD codes for a 541-amino-acid protein that forms a homo-octameric enzyme involved in histidine catabolism. SNP rs61735836 codes for a valine to methionine substitution at codon 101 (p.Val101Met) (Fig 3). The major (G) and minor (A) alleles correspond to valine and methionine, respectively. Codon 101 codes for an amino acid in the formiminotransferase N-subdomain and resides between secondary structure elements β4 and α4. This codon is highly conserved [27] with methionine being the predominant amino acid in all other vertebrates, including the Neanderthal and Denisovan sequences (with the exception of lamprey, which is Valine) (Fig 3). This suggests the derived Valine codon (G allele) has gone to near fixation in humans at some point after the modern-archaic human split, suggesting selection on a functional mutation (G) that confers a selective advantage.
We do not yet understand the mechanism by which rs61735836 presumably affects arsenic metabolism; however, there are several mechanisms by which rs61735836 may affect FTCD function. First, because the minor/ancestral allele A produces a start codon (Met), this allele may create an alternative translation start site that would produce a truncated FTCD protein. The minor allele A/Met creates the first 5´-proximal Kozak consensus sequence in the FTCD gene ([A/G]xxAUGG). While translation generally initiates at the first 5´ AUG, the efficiency with which this AUG is recognized is influenced by the presence of a Kozak consensus sequence [28]. For 5´-proximal AUG codons that do not reside in a Kozak consensus sequence, ribosomes can fail to initiate translation at that site, and continue scanning for downstream start codons (i.e., “leaking scanning”) [29]. There are three start codons upstream of rs61735836, but none are a Kozak consensus sequence, including the canonical start site (S6 Fig).
Second, the V → M amino acid substitution may alter the structure of the protein, potentially through protein folding or octamer formation, thereby altering the efficiency with which the FTCD enzyme functions. However, this substitution is not strongly predicted to be damaging according to SIFT (“tolerated” with a score of 1.0), PolyPhen-2 (benign with a score of 0.029), CADD (0.77 with a PHRED-like scaled score of 9.3), and ClinVar (likely benign).
Third, exon 3 is just downstream of several transcription factor binding sites and chromatin marks indicative of enhancers and promoters, and the exon itself is contained within a weak promoter in the HepG2 liver cancer cell line (S7 Fig). This suggests that it is possible that rs61735836 could affect initiation of transcription or represent a translation start site specific to an FTCD isoform that lacks the canonical start codon. However, among the 14 FTCD isoforms observed in GTEx liver tissue, no transcripts lacking exon 1 include exon 3 (S8 Fig). Furthermore, rs61735836 is not associated with FTCD expression in any GTEx tissue, including liver, and is not reported to be an FTCD isoform QTL, suggesting that the effect of this SNP is likely due to the amino acid substitution.
The enzyme encoded by FTCD catalyzes the two consecutive final reactions of the L-histidine degradation pathway, which links histidine catabolism to one-carbon/folate metabolism (Fig 4) [27]. First, the formiminotranserase domain of FTCD catalyzes the transfer of a formimino group from N-formiminoglutamate (FIGLU) to tetrahydrofolate (THF), freeing glutamate and adding a one-carbon substituent at the oxidation level of formic acid to THF. Second, the cyclodeaminase domain catalyzes the removal of ammonia from formimino-THF, generating 5,10-methenylTHF [30,31]. MTHFD1 catalyzes the interconversion of 5,10-methenylTHF to either 5,10-methyleneTHF or to THF (via 10-formylTHF), both of which can enter the folate cycle and be used for synthesis of 5-methylTHF. Histidine has been proposed as a potential source 5,10-methenyl-THF in some tissues [32]; however, the relative contribution of histidine to the one-carbon pool is currently unclear, and contribution may vary across tissues [33]. Additional potential roles of FTCD include catalyzing the conversion of THF to 5-formyl-THF and conversion of 5-formyl-THF to 5,10-methenyl-THF [34,35].
The one-carbon cycle is critical for arsenic metabolism, because 5-methyl-THF (primarily originating from dietary sources, but also generated from histidine catabolism) is essential to the production of S-adenosylmethionine (SAM) which provides methyl groups for methyltransferase reactions, including methylation of arsenic (Fig 4). Methylation of arsenic is catalyzed by AS3MT, a known arsenic susceptibility/metabolism gene [15,18]. The methionine cycle is also linked to the production of glutathione (GSH), which may increase the speed of arsenic reduction (i.e., arsenate (AsV) to arsenite (AsIII)), which occurs prior to methylation of arsenic by AS3MT. Variation in folate status/intake and one-carbon metabolism have long been hypothesized to influence arsenic metabolism [36], and randomized studies have provided strong evidence that folate supplementation increases arsenic metabolism efficiency and reduces blood arsenic concentrations [23,37]. However, prior candidate gene association studies of polymorphisms in one-carbon metabolism genes and arsenic metabolism have provided only suggestive or null findings [38,39], and no prior studies examined variation in FTCD.
Interestingly, a recent GWA study of 124 arsenic-exposed women living in the northern Argentinean Andes identified associations between SNPs in the 21q22.3 region and urinary DMA% (P = 1.2x10-5) and MMA% (P = 1.2x10-5) (Schlebusch et al [40]). The SNPs showing the strongest associations reside in the LSS, MCM3AP, and YBEY genes, which are in the range of ~30 to ~150 kb upstream of (and telomeric to) FTCD. While this previously reported signal is nearby the signal we report, the two signals appear distinct. Our association involves a single coding SNP in FTCD that is in very low LD with all surrounding SNPs, while the Schlebusch et al. association involves many SNPs in a LD block that spans several genes (with no association observed for SNPs within FTCD itself). Thus, it appears unlikely these two signals are due to the same causal variant. However, it is possible that the causal variants underlying these associations impact the function of the same gene(s).
As of January 31, 2019, the FTCD gene has not been reported in any GWA study of human traits (according to the NHGRI-EBI GWAS catalog). Due to the very weak LD between rs61735836 and nearby variants, this variant cannot be accurately imputed in most populations; it must be directly genotyped. However, commercially available arrays that lack “exome content” (https://genome.sph.umich.edu/wiki/Exome_Chip_Design) do not include rs61735836. Among arrays used in prior GWA studies, 25 (out of 56) Illumina arrays and 1 (out of 20) Affymetrix array include rs61735836 (based on LDlink [41]). Thus, a large fraction of prior GWA studies have not measured or imputed rs61735836, including all studies conducted prior to the development of the exome content.
Rare mutations in FTCD cause various forms of FTCD deficiency (OMIM: 229100), an autosomal recessive disorder which is the second most common inborn error of folate metabolism [31,42]. Severe forms have been reported to cause mental and physical retardation, anemia, and elevated serum folate, while less severe cases have been reported to have developmental delay and elevated levels of FIGLU in urine [30], which accumulates due to FTCD deficiency (Fig 4). Recent work has demonstrated that individuals homozygous for putative loss-of-function mutations in FTCD have clearly detectable levels of FIGLU in urine in the absence of histidine loading (which is normally very low or undetectable), in the range of 5 to 195 mmol per mol creatinine [43].
To assess the potential impact of rs61735836 on urine FIGLU, we measured FIGLU in baseline urine samples for 60 of our HEALS participants (20 for each of the three rs61735836 genotype categories) using tandem mass spectrometry in the laboratory of Dr. Devin Oglesbee as described previously [43]. We observed no evidence for elevated FIGLU among carriers or non-carriers of the G allele, with no participant having a FIGLU >0.25 mmol/mol creatinine (S9 Fig). This finding suggests that impact of rs61735836 on FTCD function is less severe than the impact of loss of function mutations on FIGLU.
Combining data on FTCD SNP rs61735836 with the two previously-reported arsenic metabolism SNPs in the AS3MT region (rs9527 and rs11191527) [15,16], we can explain ~10% of the phenotypic variation in DMA% for our HEALS participants. Mendelian randomization analyses of all three variants (using the inverse-variance weighted meta-analysis method [44]) provides strong evidence of a causal effect of arsenic metabolism efficiency (as measured by DMA%) on skin lesion (OR = 0.89 for a 10% increase in DMA%; P = 6x10-8) (Fig 5). We observe similar results when using (a) either iAs% or MMA% as a measure metabolism efficiency and (b) alternative MR methods implemented in the MendelianRandomization R package [44] (S10 Fig).
These MR results are consistent with prior observational studies [14,45–48] showing that high DMA% (and SMI) are generally associated with decreased skin lesion risk, while high iAs%, MMA%, and PMI are generally associated with increased skin lesion risk. These observational studies also indicate that, among the various arsenic metabolism measures, MMA% is most consistently associated with increased risk for skin lesions and several types of cancer [49]. Consistently, in vitro studies indicate MMAIII is likely to be the most toxic of all metabolites of inorganic arsenic [50,51]. Thus, the primary finding from this work and our prior studies—that producing DMA more efficiently (and therefore depleting iAs and MMA) reduces skin lesion risk—could be attributed to a) enhanced excretion of arsenic from the body in the form of DMA and/or b) lower percentages of the most toxic metabolites (e.g. MMAIII) among all arsenic species in the body.
FTCD SNP rs61735836 showed suggestive evidence of additive GxE interaction, results that are directionally consistent with previously reported additive interaction results for AS3MT genotypes [16]. For both loci, the expected interaction between SNP and arsenic exposure in relation to skin lesions is much more apparent on the additive scale compared to the multiplicative scale. This is an important observation considering these SNPs must modify the effect arsenic on skin lesion risk, a conclusion we draw based on the fact that these lesions do not occur in the absence of arsenic exposure. In other words, this variant cannot affect skin lesion risk among unexposed individuals, so GxE must be present. However, because we have few truly unexposed individuals in our study, we are unable to assess GxE on the present vs. absent exposure scale. In addition, it is possible that we are not well-powered to detect GxE due to the low MAF of rs61735836 and the relatively small number of genotyped cases having exposure data obtained prior to arsenic mitigation efforts (n = 443).
In summary, this work identifies a protein-altering variant in FTCD (rs61735836) that is associated with both arsenic metabolism efficiency and risk for arsenic-induced skin lesions, the most common sign of arsenic toxicity. Future studies can use this variant, in conjunction with AS3MT variants, to study the effects of arsenic exposure (through food, water, or other sources) and metabolism efficiency on health outcomes believed to be affected by arsenic (e.g., cancer and cardiovascular disease), even in the absence of data on arsenic exposure. This work provides evidence of links among histidine catabolism, one-carbon/folate metabolism, and arsenic metabolism, which is intriguing in light of the strong prior evidence supporting a role for folate status and one-carbon metabolism in arsenic metabolism efficiency [36], including randomized studies of folate supplementation in humans [23,37]. However, additional research is needed to understand (1) if and how this SNP impacts the relative distribution of folate metabolites and (2) the potential mediating role of folate on the association between rs61735836 and arsenic metabolism efficiency. A better understanding of these effects could enable the use of rs61735836 as a tool for studying the many human diseases with hypothesized connections to folate and one-carbon metabolism (e.g., cancer, vascular disease, cognitive decline, neural tube defects) [52–54].
This research was approved by the Institutional Review Board of the University of Chicago (IRB16-1236). Verbal informed consent was obtained from all participants.
The DNA samples used in this work were obtained at baseline interview from individuals participating in one of the two following studies: the Health Effects of Arsenic Longitudinal Study (HEALS) [55] and the Bangladesh Vitamin E and Selenium Trial (BEST) [56]. HEALS is a prospective study of health outcomes associated with arsenic exposure through drinking water in a cohort of adults in Araihazar, Bangladesh, a rural area east of the capital city, Dhaka. A cohort of ~12,000 participants was recruited in 2000–2002, and ~8,000 additional participants were recruited in 2006–2008. Over 6,000 wells in the study area have been tested for arsenic using graphite furnace atomic absorption spectrometry and individuals reported the primary well from which they drank. Trained study physicians conducted in-person interviews, clinical evaluations (including ascertainment of skin lesions), and spot urine collection at baseline and follow-up visits (every two years). BEST is a 2×2 factorial randomized chemoprevention trial (n = 7000) evaluating the effects of vitamin E and selenium supplementation on non-melanoma skin cancer (NMSC) risk. BEST participants are residents of Araihazar (the same geographic area as HEALS), Matlab, and surrounding areas. BEST uses many of the same study protocols as HEALS, including arsenic exposure assessment and biospecimen collection. All BEST participants had existing arsenic-related skin lesions at baseline.
The exome-wide association study of arsenic species percentages was conducted using urinary arsenic metabolite and exome chip SNP data on 1,660 individuals randomly selected from HEALS. Exome-wide association analyses of arsenic-induced skin lesions were conducted using exome chip SNP data on 2,401 cases and 2,472 lesion-free controls (from both HEALS and BEST). This case-control sample includes 1,660 HEALS participants with arsenic metabolite data. Analyses of blood arsenic metabolites were conducted using 155 cohort members for whom we had existing data on arsenic species measured in blood. These data on blood arsenic species were generated in the context of various HEALS ancillary studies: the Nutritional Influences on Arsenic Toxicity (NIAT) Study [23], the Folate and Oxidative Stress (FOX) Study [25], and the Folic Acid and Creatinine Trial (FACT) [24] (data courtesy of Gamble, MV and Graziano, JH). Among these 155 participants, 147 were included in the case-control analysis of skin lesions, and 87 were included in the analysis of arsenic metabolites in urine.
We assessed SNP-arsenic interaction using data on HEALS participants with individually-measured arsenic exposure (i.e., arsenic concentration of their primary drinking well at baseline). These exposure measures were taken prior to arsenic mitigation efforts [22]; thus, these measures represent longer-term, historical exposure levels. The majority of the HEALS participants (~95%) were lesion-free at baseline. Similarly, among our genotyped HEALS participants, only 66 of the 443 skin lesion cases were prevalent cases. The remaining 377 were incident skin lesions cases (ascertained at biennial follow-up visits by trained study physicians using a structured protocol [55]). All BEST participants had skin lesions at baseline, because a skin lesion diagnosis was part of the BEST eligibility criteria [56]. In this study, skin lesion cases were defined as individuals with any type of arsenic-induced lesion, including keratosis, melanosis, and/or leukomelanosis.
Study protocols were approved by the Institutional Review Boards of The University of Chicago and Columbia University, the Ethical Review Committee of the International Center for Diarrheal Disease Research, Bangladesh, and the Bangladesh Medical Research Council. Informed consent was obtained from all participants.
Using DNA from individuals participating in HEALS (Health Effects of Arsenic Longitudinal Study) and BEST (the Bangladesh Vitamin E and Selenium Trial), we genotyped 4,939 Bangladeshi individuals (HEALS n = 2,949; BEST n = 1,983) using Illumina’s exome array v1.1. Prior to QC, our dataset consisted of 242,901 variants. We removed samples with >3% missing SNPs (n = 6), gender mismatches (n = 22), and duplicate individuals (n = 25). We removed SNPs with call rate <97% (176 SNPs), monomorphic SNPs (n = 27,687), and 166 SNPs deviating from Hardy-Weinberg Equilibrium (P<10−10). None of the SNPs that pass this HWE threshold show HWE P-values <10−7 when relative pairs are removed from the dataset. We removed SNPs with a minor allele frequency (MAF) <1% (n = 178,015). Among the 19,992 post-QC variants, there were 17,919 missense, 141 nonsense, 1,260 synonymous, and 672 non-exonic variants. All post-QC variants were included in our analysis. A similar QC procedure for our participants’ existing genome-wide data on ~300,000 SNPs measured using the Illumina HumanCytoSNP-12 v2.1 array has been described previously [15,16].
As previously described [45], arsenic species in HEALS urine samples were separated using high-performance liquid chromatography (HPLC) and detected using inductively coupled plasma-mass spectrometry (ICP-MS) with dynamic reaction cell (DRC). Percentages of iAs, MMA and DMA among all arsenic species were calculated after subtracting arsenobetaine and arsenocholine (i.e., nontoxic organic arsenic from dietary sources) from total arsenic. All data on arsenic species in blood were generated using ICP-MS-DRC coupled to HPLC, as described previously for NIAT and FOX [23,57] (the FACT data is not yet published). Blood samples were processed in the same way for each of these studies, and this processing has been described previously in detail [23] and follows the method of Csanaky and Gregus [58]. For quality control purposes, samples with known concentrations of arsenic species were regularly analyzed. Two samples were run at the beginning of every working day and throughout the day, after every 10 samples, as previously described [23]. The limit of detection for each metabolite of interest was 0.2 μg/L. We have previously reported intra-assay CVs for this assay (from FOX) for AsIII, AsV, MMA, and DMA (0.9%, 11.5%, 3.6%, and 2.6%, respectively) as well as inter-assay CVs (3.7%, 23.2%, 2.9%, and 3.5%, respectively) [57]. Arsenic exposure in HEALS was assessed by measuring total arsenic concentration in individuals’ urine and their primary drinking well at baseline (2000–2002) [55].
We conducted exome-wide association analyses for each of the three arsenic species measured in urine (iAs%, MMA%, and DMA%) restricting to 1,660 HEALS participants with available data on arsenic species in urine. We conducted exome-wide association analyses of arsenical skin lesion status (the most common sign of arsenic toxicity) using data on 2,401 cases and 2,472 lesion-free controls (from both HEALS and BEST). All participants included in these analyses have existing genome-wide data on ~300,000 SNPs based on the Illumina HumanCytoSNP-12 v2.1 array, as described previously [15,16]. For association analysis, we used GEMMA (Genome-wide Efficient Mixed Model Association) [59] to account for cryptic relatedness, as many of our participants have a relative in the study. For the random effects model implemented in GEMMA, we used a kinship matrix based on ~260,000 genome-wide SNPs, as described previously [15]. We also used GEMMA for case/control association testing; we approximated odds ratios (ORs) by first dividing the beta coefficient by [x(1 –x)], where x is the proportion of cases in our sample, in order to estimate the beta from a logistic model. This quantity was exponentiated to obtain an OR.
Multiplicative interaction was tested by including an interaction between arsenic exposure tertiles (coded 0, 1, 2) and rs61735836 (coded 0, 1, or 2 minor alleles) in a logistic regression. Using the results from this logistic regression, additive interaction was estimated as the relative excess risk for interaction (RERI) using the delta method for confidence interval estimation [60,61]. SNP-SNP interaction was tested by including an interaction between two SNPs, coded as minor allele counts, in linear or logistic regression models. In order to analyze the effect of SNPs on arsenic species in blood, including measures taken at multiple time points for the same individuals, we used a mixed-effects model with a random intercept for each individual to account for the fact that 109 individuals appear twice in the dataset (having both baseline and follow-up/post-intervention measurements). Mendelian randomization analyses based on summary statistics were conducted using the inverse-variance weighted meta-analysis method as implemented in the MendelianRandomization R package [44], in addition to a maximum likelihood method, the median methods, and Egger regression [44]. Allele frequencies and linkage disequilibrium (LD) patterns were examined using LDlink [41] and the Geography of Genetic Variants browser [21].
|
10.1371/journal.pbio.0060257 | Proteomic Profiling of γ-Secretase Substrates and Mapping of Substrate Requirements | The presenilin/γ-secretase complex, an unusual intramembrane aspartyl protease, plays an essential role in cellular signaling and membrane protein turnover. Its ability to liberate numerous intracellular signaling proteins from the membrane and also mediate the secretion of amyloid-β protein (Aβ) has made modulation of γ-secretase activity a therapeutic goal for cancer and Alzheimer disease. Although the proteolysis of the prototypical substrates Notch and β-amyloid precursor protein (APP) has been intensely studied, the full spectrum of substrates and the determinants that make a transmembrane protein a substrate remain unclear. Using an unbiased approach to substrate identification, we surveyed the proteome of a human cell line for targets of γ-secretase and found a relatively small population of new substrates, all of which are type I transmembrane proteins but have diverse biological roles. By comparing these substrates to type I proteins not regulated by γ-secretase, we determined that besides a short ectodomain, γ-secretase requires permissive transmembrane and cytoplasmic domains to bind and cleave its substrates. In addition, we provide evidence for at least two mechanisms that can target a substrate for γ cleavage: one in which a substrate with a short ectodomain is directly cleaved independent of sheddase association, and a second where a substrate requires ectodomain shedding to instruct subsequent γ-secretase processing. These findings expand our understanding of the mechanisms of substrate selection as well as the diverse cellular processes to which γ-secretase contributes.
| All cells face the challenge of removing transmembrane proteins from the lipid bilayer for the purpose of signaling or degradation. One molecular solution to this problem is the multiprotein enzyme complex γ-secretase, which is able to hydrolyze several known transmembrane proteins within the hydrophobic lipid environment. Due to its central role in the pathogenesis of Alzheimer disease, modulation of γ-secretase activity has become a therapeutic goal. However, the number and diversity of proteins that can be cleaved by this protease remain unknown, and the attributes that target these proteins to γ-secretase are unclear. In this study, we used an unbiased approach to substrate identification and surveyed the proteome for targets of γ-secretase. Of the thousands of proteins detectable, only a relative few were substrates of γ-secretase, all of which were type I transmembrane proteins. In addition to validating several of these novel substrates, we compared them to other proteins that we identified as nonsubstrates and determined that there are specific domains that can activate or inhibit γ-secretase processing. These findings should advance our understanding of the many cellular processes regulated by γ-secretase and may offer insights into how γ-secretase can be exploited for therapeutic purposes.
| In the recently discovered process of regulated intramembrane proteolysis, activated transmembrane proteins are liberated from the lipid bilayer in a two-step mechanism. The first cleavage by a class of proteases dubbed secretases or sheddases releases the ectodomain, leaving the protein with a short lumenal stub, a transmembrane domain, and a cytoplasmic domain. The second scission occurs when a protease uses an unusual active site within the hydrophobic lipid environment to recognize and cleave the truncated target protein, releasing both the lumenal fragment and the cytoplasmic domain from the membrane. The released intracellular domain (ICD) may then signal as a transcription factor or by other means [1,2]. This process was first elucidated in studies of the pathogenesis of Alzheimer disease, in which the amyloid precursor protein (APP) is initially cleaved by β-secretase to generate an APP C-terminal fragment (CTF) that is subsequently cleaved by the intramembrane aspartyl protease γ-secretase, releasing amyloid β-protein (Aβ) from the membrane. Secreted Aβ initiates the amyloidogenic cascade that is widely believed to drive pathogenesis [3].
γ-secretase is a multiprotein complex consisting of presenilin (PS), nicastrin, Aph-1, and Pen-2, with PS containing the two catalytic aspartates that mediate peptide bond scission [4]. PS is synthesized as a holoprotein that is post-translationally cleaved into an N-terminal fragment (NTF) and a CTF, which remain bound as a heterodimer. More than 160 different missense mutations have been identified within the two human presenilin genes that cause an aggressive, early-onset form of Alzheimer disease, largely by producing longer and thus more aggregation prone species of Aβ [5]. For its key role in Aβ generation, γ-secretase has become a principal target for Alzheimer disease therapeutics aimed at inhibiting or modulating the protease's activity. In addition, γ-secretase inhibition may prove therapeutic for some forms of cancer by decreasing intracellular signaling molecules, e.g. the Notch ICD, that are generated by γ cleavage [6].
Through extensive investigation, principally of the prototypical substrates Notch and APP, a general model of γ-secretase function and activity has emerged [4]. γ-secretase processing is preceded by shedding of the substrate's ectodomain by either an α- or β-secretase, generating a CTF with a short N-terminal extracellular domain. Once this conversion to a CTF has occurred, the γ-secretase complex can bind the substrate and then translocate it to the active site, where intramembrane proteolysis can occur within an interior hydrophilic chamber [7,8]. There remain several major gaps in our understanding of γ-secretase proteolysis. It is unclear what number and spectrum of substrates are processed by γ-secretase. From the existing candidate-based substrate studies, it has been proposed that γ-secretase may function as a type of “proteasome of the membrane,” with loose substrate specificity [9]. It is thought that γ-secretase can only cleave type I transmembrane proteins. However, substrates of other topologies have been proposed [10,11], and preference for type I proteins has never been addressed in an unbiased manner. Finally, it is unclear which regions of a substrate are important for binding and subsequent proteolysis by γ-secretase, and whether a sheddase may, beyond reducing ectodomain size, contribute to substrate specificity.
To address these issues, we have performed a novel unbiased proteomic screen to identify the range of substrates in the proteome that are regulated by γ-secretase processing. Coupling a selective γ-secretase inhibitor with the method of stable isotope labeling with amino acids in cell culture (SILAC) and subsequent mass spectrometry, we have identified a relatively small cohort of novel γ-secretase substrates among thousands of other proteins that are unchanged by γ-secretase inhibition. In agreement with the general model of γ-secretase processing, all substrates we identified were type I transmembrane proteins. Using genetic and pharmacological manipulations, we validated a subset of substrates and nonsubstrates and confirmed that γ-secretase cleavage is preceded by ectodomain shedding. By generating chimeric proteins containing substrate and nonsubstrate regions, we identified two potentially independent mechanisms for targeting a protein for proteolysis. Finally, we determined that γ-secretase binding of a truncated type I protein is not sufficient to induce proteolysis, and that permissive transmembrane and cytoplasmic domains are required for γ-secretase cleavage to occur. These results demonstrate that γ-secretase regulates a relatively small subset of the membrane proteome, and that substrates have specific determinants that enable their recognition and proteolysis.
The development of tools to compare proteomes quantitatively has enabled identification of phosphorylation cascades [12], DNA damage responses [13], and changes in tumor cell lines [14]. The SILAC method takes advantage of the ability of mass spectrometry to differentiate between heavy and light isotopic variants. By metabolically labeling cells with amino acids containing either heavy or light isotopes, the entire proteome can be quantitatively compared, and differences arising from genetic disparity or experimental treatments can be determined. In the present study, we sought to identify substrates of γ-secretase using SILAC coupled with treatment by a selective γ-secretase inhibitor.
HeLa cells were labeled with light or heavy lysine and arginine. Once labeled, the γ-secretase inhibitor DAPT was applied to the light labeled cells (Light+DAPT) and control DMSO solvent to the heavy labeled cells (Heavy+DMSO). After this 16 h treatment, equal numbers of cells were combined and then fractionated. As evidence of successful γ-secretase inhibition, we examined endogenous APP levels by Western blot (Figure 1A). In whole-cell lysates (lanes 1 and 2) DAPT treatment did not change the levels of full-length APP, but as expected increased the levels of the APP CTF, the immediate substrate of γ-secretase. Proteins from the combined Light+DAPT and Heavy+DMSO treatment conditions were fractionated into cytosolic and membrane fractions, with the APP and APP CTFs found in the membrane fraction and absent from the cytosol. Purified membrane proteins were separated by SDS-PAGE and stained with Coomassie (Figure 1B). For mass spectrometry analysis, the lane was divided into ten horizontal slices, and each subjected to tryptic digestion and LC-MS/MS (see Materials and Methods for details).
The expected spectral pattern for full-length APP and APP CTF peptides, as well as other putative γ-secretase substrates, is shown in schematic form in Figure 1B. Peptides derived from full-length proteins would not be expected to change in relative abundance in response to γ-secretase inhibition (upper spectrum), regardless of whether they are true substrates or nonsubstrates. In contrast, peptides derived from γ-secretase substrates, following ectodomain shedding that reduces the protein's molecular weight, should be higher in relative abundance in the light-labeled condition treated with DAPT (lower spectrum). Data arising from all quantitative peptide comparisons were analyzed for this spectral pattern to identify individual γ-secretase substrates expressed under endogenous conditions. In total, over 16,400 peptides representing more than 2,500 proteins were quantitatively compared (Table S1).
The APP protein sequence is shown in Figure 1C to exemplify the findings of this proteomic analysis. An APP tryptic peptide derived from the higher–molecular weight region of the gel and present in equal abundance under the two labeling conditions is indicated in blue, and two peptides from the lower–molecular weight region of the gel with significantly increased abundance in the Light+DAPT condition are indicated in red. As expected for APP, the peptides enriched in the Light+DAPT condition map to the C-terminal region of the protein, which remains embedded in the membrane after ectodomain shedding and represents the substrate for γ-secretase. Figure 1D displays all the proteins identified by this quantitative proteomic screen that fit the criteria for a potential γ-secretase substrate. Listed next to the protein in the left column are the higher–molecular weight, equal-abundance peptides, and in the right column are the relatively lower–molecular weight peptides enriched by DAPT treatment. APP and its homolog amyloid precursor-like protein 2 (APLP2) [15], as well as CD44 [16], are previously described γ-secretase substrates. The other APP homolog, APLP1, was not identified in this screen, consistent with it being expressed in the nervous system [17] and thus not present in the epithelial HeLa cell line. Several novel substrates were identified in this screen that have closely related homologues previously identified as γ-secretase substrates: the human leukocyte antigen (HLA) [18], the low-density lipoprotein receptor (LDLR) [19], and syndecan-1 and −2 (Synd) [20]. The remaining novel γ-secretase substrates we identified are: dystroglycan (DG), the Delta/Notch-like EGF-related receptor (DNER), desmoglein-2 (DSG2), natriuretic peptide receptor-C (NPR-C), plexin domain-containing protein 2 (PLXDC2), and vasorin. Table S2 lists the sequence and quantitative information of the peptides derived from these putative γ-secretase substrates.
To confirm that the proteins identified in the SILAC proteomic screen are indeed γ-secretase substrates, DG, DNER, DSG2, NPR-C, PLXDC2, and vasorin were each cloned and C-terminally tagged with a FLAG epitope. To also validate two of the many proteins identified by the screen that were not implicated as γ-secretase substrates, the proteins integrin β-1 (Itgβ1) and natriuretic peptide receptor-A (NPR-A) were similarly cloned and FLAG-tagged. These two apparent nonsubstrates were chosen based on their similarities to known substrates, including type I transmembrane topology and function in cell adhesion (Itgβ1) or as a peptide receptor (NPR-A). As a genetic method of validation (Figure 2), we used a fibroblast cell line that contains deletions of the Presenilin-1 and -2 genes and thus completely lacks γ-secretase activity. In this PS double-knockout (PS-DKO) cell line, we expressed one of three constructs by viral transduction to generate a substrate validation system: (1) a control cell line expressing just empty vector; (2) a cell line expressing a catalytically inactive PS construct with one active site aspartate mutated to alanine (PS1-D257A), which thus restores γ-secretase complex formation but prevents catalytic activity; and (3) a cell line expressing wild-type PS1 that forms a proteolytically active γ-secretase complex (Figure 2A). Due to low expression of some proteins, not all cloned substrates could be tested in these PS-DKO cells, but they were validated separately by DAPT inhibitor treatment (see below).
Expression of either PS1-D257A or wild-type PS1 allowed the formation of the γ-secretase complex, as confirmed by the mature glycosylation of nicastrin [21] (Figure 2A, top panel). Probing for PS1 by Western blot with an N-terminally directed antibody (middle panel) showed both PS1 holoprotein and NTF in the wild-type PS1 cells, whereas the PS1-D257A construct remained as a holoprotein, failing to be post-translationally cleaved into NTF and CTF. These data confirm that PS holoprotein processing into NTF and CTF occurs by an endoproteolytic mechanism dependent on PS activity. As a readout for establishment of γ-secretase activity, these various cell lines were probed for the presence of endogenous APP CTFs (lower panel), whose levels were reduced in the PS1-expressing cells, thus indicating functional γ-secretase activity only with expression of wild-type presenilin.
As a second means of validation, the cloned substrates and nonsubstrates were expressed stably in HEK cells, and the presence or absence of regulation by γ-secretase was confirmed by treatment with DAPT (Figure 3). In further analyses of these cells, we searched for evidence of the sheddase responsible for ectodomain secretion by applying the β-secretase inhibitor C3, the metalloprotease (α-secretase) inhibitor GM6001, or the phorbol ester PMA, which stimulates ectodomain shedding. To probe for the γ-secretase–mediated release of intracellular domains, many of which have been shown to be degraded rapidly by the proteasome [18,22–24], we treated cells with the proteasome inhibitor epoxomicin. The genetic (Figure 2) and pharmacological (Figure 3) experiments will be described together for each potential substrate or nonsubstrate.
Vasorin is an inhibitor of TGF-β signaling and may play a role in vascular remodeling [25]. In the PS-DKO cells, vasorin CTF levels were potently reduced by active γ-secretase (Figure 2B and 2H). Vasorin achieved high overexpression in HEK cells, with significant CTF accumulation in response to DAPT treatment, and a reduction in CTF levels with a 6-h GM6001 treatment, indicating the vasorin sheddase to be a metalloprotease (Figures 3A and S1A). A prominent vasorin ICD was observed with vasorin expression, which was further stabilized by proteasome inhibition (Figure 3A, far right lane). In both the PS-DKO and HEK cells, several distinct vasorin CTF bands were observed. Similarly, when expressed at comparatively low levels in HeLa cells, multiple vasorin CTF bands were still generated (unpublished data). This CTF pattern is unlikely to be due to multiple sheddase cleavages (see below), but may be due to post-translational modification or SDS-stable multimer formation. Of interest, we noticed that stable vasorin expression in HEK cells altered cell morphology and reduced cell adherence (Figure S1I), although it did not produce overt cytotoxicity. Such a morphological change was not observed upon expression of any of the other γ-secretase substrates.
DG, a member of the multiprotein dystrophin–glycoprotein complex, provides a physical connection between the extracellular matrix and the intracellular cytoskeleton [26] and is implicated in several diseases, including muscular dystrophies [27]. DG is synthesized as a long precursor protein and is post-translationally cleaved into an extracellular α-DG fragment that noncovalently associates with the membrane-anchored β-DG fragment [28]. In our PS-DKO cell lines, β-DG was observed to undergo shedding to generate a DG CTF, whose levels were then regulated by γ-secretase activity (Figure 2C and 2H). In the HEK cells, DAPT treatment led to a significant increase in two distinct DG CTFs, one of which was produced by metalloprotease cleavage (Figures 3B and S1B–S1D). A DG ICD was observed without DAPT treatment, which significantly accumulated with proteasome inhibition (Figure 3B, right panel). To compare the complex banding pattern observed in HEK cells, suggesting more than one ectodomain proteolytic pathway, we also analyzed DG processing in HeLa cells. When DG was expressed in HeLa cells, only a single CTF was produced from β-DG, and it accumulated upon γ-secretase inhibition (Figures 3C and S1B). Thus, different cell types apparently use different β-DG secretory processing pathways.
DNER is a Notch ligand expressed in neurons and involved in cerebellar development and function [29,30]. In the PS-DKO cell lines, full-length DNER was expressed as a doublet, perhaps arising from differential glycosylation, and produced a CTF that was regulated by γ-secretase (Figure 2D and 2H). In HEK cells, DNER expression resulted in the production of a CTF whose levels also increased with DAPT treatment. After a 16-h treatment, the metalloprotease inhibitor GM6001 modestly reduced CTF levels by inhibiting α-secretase mediated ectodomain shedding (Figure S1E). The DNER ICD could be detected upon inhibition of the proteasome (Figure 3D, far right lane).
NPRs bind to circulating natriuretic peptides (atrial natriuretic peptide, brain natriuretic peptide, and C-type natriuretic peptide) [31]. NPR-A and NPR-B contain a guanylate cyclase domain, and upon binding natriuretic peptides generate cyclic GMP. NPR-C, while similar to NPR-A and -B in the function of its extracellular domain, has a shortened cytoplasmic domain with no guanylate cyclase activity. Thus, with no identified signaling function [32], NPR-C has been termed a “clearance receptor” whose principal role may be to remove excess natriuretic peptides [33,34]. However, recent studies have identified regions in the cytoplasmic tail of NPR-C that modulate G-protein activity [35], suggesting that NPR-C's short cytoplasmic tail has a signaling function. In our PS-DKO cell lines, expression of full-length NPR-C resulted in an NPR-C CTF, the levels of which were significantly regulated by γ-secretase activity (Figure 2E and 2H). Expression in HEK cells revealed robust CTF accumulation with DAPT treatment, and metalloprotease inhibition by 16-h GM6001 treatment reduced CTF levels (Figures 3E and S1F).
Desmogleins are structural components of desmosomes, which form intercellular junctions and direct tissue morphogenesis [36]. Overexpression of DSG2 has been noted in several forms of cancer, and transgenic overexpression of DSG2 drives tumorigenesis by altering multiple signaling pathways [37]. In HEK cells, DSG2 expression produced a full-length protein and a CTF that was processed by γ-secretase. CTF levels decreased significantly upon 6-h treatment with either C3 or GM6001, indicating that both β-secretase and an α-secretase–like metalloprotease can shed the DSG2 ectodomain (Figures 3F and S1G). The DSG2 ICD was observed when its degradation was prevented by the proteasome inhibitor epoxomicin (Figure 3F, far right lane).
PLXDC2 is a poorly characterized protein expressed in the developing nervous system [38] and found to be up-regulated in cancerous tissue [39]. Full-length PLXDC2 is robustly expressed in HEK cells, though the levels of CTF produced by ectodomain shedding are relatively low. Shedding is apparently due to a metalloprotease, as a 16-h GM6001 treatment significantly reduces CTF levels, while the CTF levels increase with PMA treatment (Figures 3G and S1H).
In addition to the novel substrates discussed above, analysis of proteins that we identified to not be processed by γ-secretase may yield insight into substrate requirements. Integrins are cell adhesion molecules that also play substantial roles in signal transduction. All integrin isoforms are single-pass membrane proteins that form αβ heterodimers [40]. Itgβ1 achieved modest expression in PS-DKO cells (Figure 2F), and stronger expression in HEK cells (Figure 3H), although no CTF was observed and protein levels were not modulated by sheddase activators or inhibitors (Figure 3H, right panel). We also analyzed endogenous Itgβ1 levels in the PS-DKO cells using a C-terminally directed antibody, and again found no evidence for CTF production (unpublished data). The second nonsubstrate examined was NPR-A, and despite being in the same family as NPR-C, NPR-A did not yield a CTF and was not proteolytically regulated by sheddase or γ-secretase activity, both in PS-DKO cells (Figure 2G) and HEK cells (Figure 3I). Thus, as first identified in our unbiased proteomics screen, these two proteins are not regulated by γ-secretase processing, despite having type I topology and functions in common with some other known γ-secretase substrates.
Our finding that Itgβ1 and NPR-A are not γ-secretase substrates suggests that protease activity is directed only at specific type-I transmembrane proteins. In an effort to establish which regions within a substrate are important for γ-secretase recognition and proteolysis, we used recombinant methods to fuse the domains of a γ-secretase substrate (vasorin) and a nonsubstrate (Itgβ1), thereby producing various chimeric type I proteins. By expressing these chimeras in HEK cells and inhibiting γ-secretase activity with DAPT, it becomes feasible to determine which regions of a protein are permissive or inhibitory to γ-secretase processing. First, we examined chimeras of full-length proteins. A construct containing the full-length ectodomain of integrin fused to the transmembrane and cytoplasmic domains of vasorin (Int/Vas) was expressed at high levels, although no CTF was produced. This is presumably due to the observed lack of ectodomain shedding of the integrin extracellular domain (above), and thus Int/Vas is not a substrate of γ-secretase (Figure 4A), even though the transmembrane domain of intact vasorin undergoes robust γ cleavage. In contrast, when the ectodomain of vasorin is fused to the transmembrane and cytoplasmic domain of Itgβ1 (Vas/Int), ectodomain shedding occurs to produce a CTF that is processed by γ-secretase (Figure 4B). Thus, ectodomain shedding is a requirement of γ-secretase processing, even when a protein contains a transmembrane and cytoplasmic domain normally recognized and cleaved by γ-secretase.
To bypass the requirement for ectodomain shedding, chimeric CTFs can be produced directly by fusing the signal peptide to the ectodomain 13 amino acids before the transmembrane domain (designated CTF13). Signal peptide removal produces a protein that resembles identically a CTF generated by sheddase cleavage [41]. Interestingly, the Vas/Int CTF13 protein was not regulated by γ-secretase (Figure 4C). This is in contrast to the Vas/Int CTF derived from ectodomain shedding of the full-length Vas/Int (Figure 4B). Taken together with data from other chimeric CTFs discussed below, these data suggest that there may be an alternative pathway for γ-secretase recognition of a CTF, in addition to the canonical stepwise cleavages by a sheddase and then γ-secretase, which may depend on the context in which the CTF is generated and presented to γ-secretase (see Discussion).
It has been hypothesized that due to the lack of a consensus recognition sequence in the known γ-secretase substrates, the protease has loose sequence specificity and simply requires that a protein's ectodomain is shed to enable γ-secretase processing. To test this hypothesis, we made artificial CTFs of Itgβ1, which normally does not have its ectodomain shed and is not regulated by γ-secretase cleavage (Figures 2F and 3H). We first examined Itgβ1 CTF25, which after signal peptide removal should produce a CTF with an ectodomain containing the first 25 amino acids of the integrin lumenal domain. Expression in HEK cells with versus without DAPT treatment (Figure 4D) and in the PS-DKO cell system (Figure 4E) produced no evidence for processing of Itgβ1 CTF25 by γ-secretase. Similar results were observed for the Itgβ1 CTF13, which has a shorter ectodomain as would be generated by an α-secretase (Figure 4F). Thus, γ-secretase requires that a substrate has features more than just a short ectodomain. In this regard, dimerization of CTFs has been hypothesized to be important for γ-secretase processing of APP [42] and activated tyrosine kinase receptors [43], although forcing dimerization of the ectodomain can inhibit γ cleavage [44]. To investigate whether intracellular dimerization of a CTF is sufficient to confer γ-secretase cleavage, we appended the dimerization-inducing leucine zipper domain to the C terminus of Itgβ1 CTF13 to produce Itgβ1-LZ CTF13. However, Itgβ1-LZ CTF13 levels were not changed by DAPT treatment (Figure 4K), suggesting that homo-dimerization is not sufficient to promote γ-secretase proteolysis.
The vasorin CTF13, as expected from previous results with the full-length vasorin protein (Figures 2B and 3A), was strongly regulated by γ-secretase processing in that DAPT treatment increased vasorin CTF13 levels approximately 10-fold (Figure 4G and 4L). Also consistent with γ-secretase cleavage of the vasorin CTF13, a vasorin ICD was observed in the absence of DAPT (Figure 4G, lower panel). Of note, accumulated vasorin CTF13 produced higher–molecular weight bands similar to those seen with expression of the full-length protein, again suggesting either complex post-translational modification or a very stable multimerization of the CTF. A CTF chimera of integrin and vasorin composed of a short integrin ectodomain plus the vasorin transmembrane and cytoplasmic domains (Int/Vas CTF13) was processed by γ-secretase (Figure 4H), with significant increases in CTF levels caused by DAPT treatment (Figure 4L). Int/Vas CTF13 constructs also produced a detectable ICD following γ-secretase cleavage (Figure 4H, lower panel), which further accumulated with proteasome inhibition (unpublished data). That Int/Vas CTF13 (Figure 4H), but not the full-length Int/Vas (Figure 4A), was processed by γ-secretase further supports ectodomain shedding as a necessary event occurring prior to a substrate's recognition by γ-secretase.
Integrin and vasorin CTF chimeras that contain only the vasorin cytoplasmic domain (Int/Vascyto CTF13) (Figure 4I) or only the vasorin lumenal and transmembrane domains (Vas/Intcyto CTF13) (Figure 4J) were not substrates of γ-secretase. Due to the higher–molecular weight bands observed in the Int/Vascyto CTF13 (Figure 4I), the vasorin cytoplasmic domain is likely the site of post-translational modification that produces the banding pattern also observed in natively produced vasorin CTF, vasorin CTF13, and Int/Vas CTF13. Taken together, the data suggest that both a permissive transmembrane and a permissive cytoplasmic domain are necessary to confer γ-secretase cleavage upon a substrate, and that the ectodomain must be short but does not by itself dictate substrate processing.
To determine why NPR-C is regulated by γ-secretase processing but not by the NPR-A protein, despite their being closely related receptors, we generated various chimeric constructs in a similar manner as those described above. The NPR-A CTF13 construct itself, like Itgβ1 CTFs, was not processed by γ-secretase (Figure 5A); again demonstrating that more than ectodomain shedding is required for a protein to be a γ-secretase substrate. Moreover, when the large intracellular domain of NPR-A (containing the kinase homology, hinge, and guanylate cyclase domains [45]) was attached to either the NPR-C lumenal and transmembrane domains (NPR-C/Acyto CTF13) (Figure 5B), or fused to the C terminus of the NPR-C CTF (NPR-C+A CTF13) (Figure 5C), the resulting constructs were not processed by γ-secretase (quantified in Figure 5I).
In contrast to the chimeras bearing the NPR-A C terminus, NPR-C CTF13 levels were increased with DAPT treatment (Figure 5D), as expected for this previously demonstrated substrate (Figure 2E and 2H). Addition of the lumenal NPR-A domain (NPR-A/C CTF13) (Figure 5E) or both the NPR-A lumenal and transmembrane domains (NPR-A/Ccyto CTF13) (Figure 5F) did not perturb the ability of the construct to be regulated by γ-secretase (quantified in Figure 5I). That these three CTFs bearing the NPR-C cytoplasmic domain are processed, but not the constructs bearing the NPR-A cytoplasmic domain, suggested an inhibitory role of the NPR-A cytoplasmic tail in γ-secretase processing. To test this hypothesis, we truncated the NPR-A C terminus by removing the three domains (kinase, hinge, and guanylate cyclase) to produce NPR-A ΔKG CTF13, which has a cytoplasmic tail of similar size to that of NPR-C. Treatment of cells expressing this construct with DAPT showed that NPR-A ΔKG CTF13 was regulated by γ-secretase (Figure 5G and 5I) to a similar extent as NPR-C CTF13. These data demonstrate that, while the NPR-A transmembrane and juxtamembrane domains are permissive as substrates, the C-terminal cytoplasmic tail of NPR-A is not permissive of γ-secretase processing. This inhibition is not simply due to the large size of the NPR-A cytoplasmic domain, however, because the γ-secretase substrate desmoglein-2 contains a cytoplasmic domain of similarly large size (Figure 3F).
To determine if the processing of a substrate more potently regulated by γ-secretase could be inhibited by the NPR-A cytoplasmic domain, we fused this domain to the C terminus of the vasorin CTF13 construct, producing Vasn+NPR-A CTF13 (Figure 5H). DAPT treatment of this fusion protein resulted in a small but significant 1.8-fold increase in CTF levels (Figure 5I), demonstrating its regulation by γ-secretase. Considering that vasorin CTF13 levels increase 10-fold upon γ-secretase inhibition (Figure 4G and 4L), we hypothesize that this observed reduction in cleavage efficiency of the vasorin transmembrane domain by γ-secretase is due to partial inhibition by the NPR-A cytoplasmic domain.
To exclude the possibility that the nonsubstrate chimeric CTFs we expressed were cleaved simply because they were in a subcellular location where no γ-secretase exists, we performed co-immunoprecipitation experiments. Previous reports indicate that substrate CTFs, and to a much lesser extent, full-length proteins prior to ectodomain shedding, associate with γ-secretase [46,47]. The initial binding by γ-secretase to a CTF may principally be regulated by nicastrin and only require that the CTF have a free N terminus [48]. We immunoprecipitated for the FLAG epitope at the C terminus of several of our stably transfected constructs (Figure 6A, bottom panel), then probed the immunoprecipitates for the co-precipitation of nicastrin (Figure 6A, top panel) and PS1 (Figure 6A, middle panel) to determine association with the γ-secretase complex. Both full-length vasorin and vasorin CTF13 were able to co-immunoprecipitate with the γ-secretase components (Figure 6A, lanes 3 and 4), as expected of these established γ-secretase substrates. By Western blot, the CTF derived from the full-length vasorin runs larger than vasorin CTF13 (Figure 6A, lower panel), suggesting that the sheddase for vasorin cleaves somewhat further than 13 amino acids from the membrane. We found that a slightly larger vasorin construct, vasorin CTF25, had a molecular weight more consistent with the natively produced CTF, and behaved similar to vasorin CTF13 (Figure S2B–S2D). Whereas full-length Itgβ1 was not able to significantly co-precipitate with nicastrin or PS, Itgβ1 CTF13 robustly associated with the γ-secretase components (lanes 5 and 6). Lysates of cells expressing Itgβ1 CTF13 have an immunoreactive band principally at the expected size, although immunoprecipitation also pulls down higher–molecular weight aggregates (Figure 6A, lower panel, lane 6). NPR-A CTF13, another nonsubstrate with a short ectodomain, was able to associate with γ-secretase to a much greater extent than full-length NPR-A (lanes 7 and 8). Co-immunoprecipitation of other substrates with γ-secretase was observed, and these results show a correlation between the amount of CTF present in the cellular lysates and the extent of association with nicastrin and PS1 (Figure S2A).
Considering that nonsubstrates are able to co-precipitate with γ-secretase, we sought to determine whether the nonsubstrate CTF chimeras can inhibit γ-secretase activity by comparing the levels of endogenous APP CTF in cells expressing substrates and nonsubstrates (Figure S3A). Under control and DAPT treatment conditions, we found no difference in APP CTFs between cells expressing substrates and nonsubstrates (Figure S3B), demonstrating that cellular expression of nonsubstrate CTFs does not interfere with endogenous γ-secretase processing. We also analyzed the processing of APP and Itgβ1 C100FLAG constructs in a purified in vitro γ-secretase activity assay. Under standard assay conditions, we detected cleavage of APP C100FLAG to generate an ICD, but we were unable to detect cleavage of Itgβ1 C100FLAG (Figure S3C). To examine the ability of Itgβ1 C100FLAG to inhibit the processing of APP, we co-incubated the two constructs together with γ-secretase and observed reduced APP ICD production (Figure S3D). Although it is unclear how this inhibition occurs, one plausible interpretation is that Itgβ1 is acting as a competitive inhibitor of APP's association with γ-secretase in this purified preparation. These data, taken together with the data in Figures 4 and 5, suggest that having a short ectodomain is sufficient to allow association with γ-secretase, but that permissive transmembrane and cytoplasmic domains are required for further processing by the protease. Furthermore, the ability of the nonsubstrate CTFs to associate with γ-secretase but not get cleaved provides further evidence for the existence of independent docking and catalytic sites within the γ-secretase complex [48,49].
Figure 6B shows partial amino acid sequences of the proteins identified in this proteomics screen, with the two nonsubstrates that we analyzed (Itgβ1 and NPR-A) listed at the bottom. Sequences are aligned along the transmembrane domains, and amino acids showing similarity down a column highlighted based on side chain polarity. In contrast to many other soluble proteases and even other intramembrane proteases [50], there is no apparent motif that predicts whether a protein is a γ-secretase substrate.
By performing a critical proteolytic event in several pathways of signal transduction, γ-secretase is essential for embryonic development, the maintenance of adult tissues, and the pathogenesis of certain diseases [51]. In this study, we have used an unbiased proteomic approach to identify proteins within the cell membrane that are regulated by γ-secretase. By quantitatively comparing the proteomes of cells treated with and without a potent γ-secretase inhibitor, we could identify accumulated substrates of the protease. Using tools such as small molecule inhibitors or gene silencing, the methods we applied here can, in principle, be exploited to determine the substrates regulated by any protease of interest. For enzymes similar to γ-secretase, where no substrate recognition sequence has been identified [52,53], quantitative proteomics may be the most effective unbiased method of endogenous substrate identification.
We found that among the thousands of proteins surveyed, only relatively few were processed by γ-secretase. In line with previously described substrates, all are type I transmembrane proteins whose ectodomain is shed prior to γ-secretase cleavage. Using a β-secretase inhibitor and a broad spectrum metalloprotease inhibitor, we were able to provide evidence for the proteases responsible for ectodomain shedding. For a few substrates, we found small but significant decreases in CTF levels by inhibiting α-secretase metalloproteases. That we found only a small decrease for a few substrates could have several explanations, such as the absence of a cognate ligand for the substrate to stimulate its ectodomain shedding, an incomplete inhibition of the responsible sheddase due to our obligatory use of a broad-spectrum sheddase inhibitor, or a generally low rate of ectodomain shedding of a particular substrate in the cultures under study. By choosing to focus our analysis on human HeLa cells, we limited the size and diversity of the proteome we examined, and thus it is not surprising that we did not identify all of the previously known γ-secretase substrates. Similar investigations of different cell types, or of various tissues from inhibitor-treated mice using other labeling methods [54], would greatly broaden the scale of analysis and likely reveal additional substrates. That only a select number of proteins are detected as γ-secretase substrates suggests that γ-secretase may not be an indiscriminate intramembrane protease, and that the γ processing of these substrates may have functional signaling implications. For some substrates, including Notch [55], ErbB4 [43], and N-cadherin [56], a functional γ-secretase cleavage has been clearly demonstrated, in that the respective ICDs are released from the membrane to modulate transcription. For other substrates, such as APP, the purpose of ICD production has been more elusive, but several functions have been proposed [57–59].
Several of the new substrates that we have identified could have important signaling functions regulated by γ-secretase. DNER, as a known Notch ligand and with its involvement in patterning cerebellar development, may signal antagonistically to Notch in a similar fashion as some other Notch ligands [60]. DG levels have been found to decrease in muscular dystrophy, and any deficit in signaling through the DG ICD, which is remarkably stable compared to the other ICDs we observed, may be an important and thus far unappreciated contributor to the disease's progression. Beyond promoting intercellular adhesion, desmoglein-2 has been implicated in signaling involving tumorigenesis, and expression of a recombinant cytoplasmic fragment of DSG2 has recently been reported to modulate apoptosis in intestinal epithelial cells [61]. Our finding that DSG2 can be shed by either β-secretase or an α-secretase–like metalloprotease followed by γ-secretase–mediated liberation of its ICD may identify the proteolytic pathway that generates this signaling fragment. NPR-C was previously considered an inactive clearance receptor, although some studies have found functional consequences of NPR-C signaling that are mediated by its short cytoplasmic tail [35]. Our demonstration that NPR-C is processed by γ-secretase to liberate this functional domain from the membrane may more clearly define the molecular mechanism behind NPR-C signaling, as well as raise the additional possibility of its modulating transcriptional activity by gaining access to the nucleus as an ICD. Although these proposed consequences of γ-secretase processing and ICD function for the newly identified substrates are speculative at this juncture, our description of the sequential proteolytic processing of these substrates recommends further functional studies.
The experiments with chimeric constructs provide insight into which features γ-secretase requires in recognizing and cleaving its substrates. First, the protein's ectodomain must be removed before cleavage, as previously reported [44]. For example, fusing the integrin ectodomain to vasorin's transmembrane and cytoplasmic domains precludes both sheddase and γ-secretase cleavage. However, when the large ectodomain is removed (producing the Int/Vas CTF13 constructs), these proteins are readily processed by γ-secretase. Although shortening of the ectodomain can enable binding to the γ-secretase complex (e.g., Intβ1 CTF13 and NPR-A CTF13; Figure 6A), this is not sufficient to promote proteolysis (Figures 4F and 5A). Our chimeric results suggest that the ectodomain contributes little to substrate specificity, although some reports suggest that the short ectodomain may modulate the avidity of γ-secretase processing [62,63].
Second, substrates must have both permissive transmembrane and cytoplasmic domains. Extensive mutagenesis experiments can now be performed to determine precisely what makes these domains permissive in vasorin but not in integrin, but in general, the primary sequence of proteins has thus far not been a reliable predictor of substrates of γ-secretase. Perhaps permissive domains allow the CTF to adopt a conformation amenable to γ cleavage, as is the case for substrates of the intramembrane protease rhomboid [64], but such a conformational change has yet to be examined experimentally.
Third, inhibitory domains may exist on a protein that preclude γ-secretase processing. We show this to be the case for NPR-A, whose large cytoplasmic domain reduces cleavage of the NPR-A CTF and also does so when fused to end of the NPR-C and vasorin CTFs (Figure 5). This inhibitory function is not due simply to the large size of the cytoplasmic domain; it may alternatively lie in protein folding that prevents proper entry of the transmembrane domain into the active site of γ-secretase, in a subcellular localization that protects the protein from recognition by γ-secretase, or in protein associations that anchor the cytoplasmic domain in a rigid conformation that γ-secretase does not efficiently recognize.
Finally, our data suggest that at least two pathways may exist that allow for the processing of CTFs. The first, as exemplified by vasorin, NPR-C, and other permissive CTF13 constructs, is the initial production of a CTF apart from γ-secretase that may later bind to the complex and undergo proteolysis. The second mechanism is suggested by the contrast between the full-length Vas/Int construct, which generates a CTF after ectodomain shedding in the secretory pathway and is processed by γ-secretase (Figure 4B), and Vas/Int CTF13, which shares the same size and sequence but is not processed by γ-secretase (Figure 4C). A possible explanation for this finding is that the occurrence of ectodomain shedding in situ flags the protein for γ-secretase cleavage by forcing a preferred conformation and/or by initiating a multi-protein interaction that passes the newly produced CTF directly to γ-secretase for processing. Further experimentation is now required to clarify these newly proposed mechanisms for substrate proteolysis by γ-secretase. This work should enhance our understanding of the many physiological processes regulated by this ubiquitous and conserved protease and at the same time provide insights into the proteolytic mechanism of γ-secretase that can be exploited for therapeutic advances.
HeLa cells were propagated for six doublings in DMEM lacking l-lysine and l-arginine (Invitrogen), and supplemented with 10% dialyzed fetal bovine serum (FBS) (Calbiochem), antibiotics, and either 12C14N arginine, 12C14N lysine (“light”), or 13C15N arginine, 13C15N lysine (“heavy”) (Cambridge Isotope Laboratories). Cells were then treated with the γ-secretase inhibitor DAPT [65] (light label condition) or DMSO (heavy label condition) as control for 16 h in labeling media containing 0.5% dialyzed FBS. After treatment, cells were suspended by trituration with PBS containing 5 mM EDTA and counted. Equal numbers of cells from each treatment were combined and subjected to hypotonic lysis and Dounce homogenization. Nuclei were removed from the homogenate by spinning at 1,000g for 10 min, and the remaining supernatant was spun at 125,000g for 1 h to pellet cellular membranes. The membrane fraction was washed with 100 mM Na2CO3 (pH 11.5), briefly sonicated, and spun again at 125,000g for 1 h to pellet membranes.
One hundred micrograms of the purified membrane proteins were run on an 8–16% Tris-Glycine SDS-PAGE gel, stained with Coomassie blue, divided into ten molecular weight gel slices, and subject to in-gel digestion with trypsin. Liquid chromatography tandem mass spectrometry (LC-MS/MS) was performed using an LTQ FT hybrid linear (2-D) ion trap-Fourier transform ion cyclotron resonance (FTICR) mass spectrometer (ThermoFisher) as previously described [66]. Resulting MS/MS spectra were matched to a composite target-decoy [67] human sequence database [68], by both SEQUEST and Mascot search engines. An in-house algorithm was used to select confident peptide identifications with an estimated false discovery rate less than 1%. Confident peptide identifications were then subjected to Vista, an automated software suite which measures the relative abundance of light and heavy isotopic peptide pairs [14,69]. This analysis yielded over 16,400 quantitative peptide comparisons, with an estimated false discovery rate of 10%.
Proteins were considered substrates when the following conditions were met: (1) peptides derived from a higher molecular weight gel band, consisting of full-length protein, have a Light:Total peptide ratio of 0.5; (2) peptides derived from a lower–molecular weight gel band, consistent with a shorter fragment after ectodomain shedding, have a Light:Total peptide ratio of 0.65 or greater; (3) each unique peptide is identified more than once; (4) multiple unique peptides from the protein are identified.
Human embryonic kidney (HEK) 293-FT (Invitrogen) and mouse embryonic fibroblasts derived from Presenilin1/2 null mice [21] were grown in Dulbecco's modified Eagle's medium containing 10% FBS, 2 mM l-glutamine, 100 μg/ml penicillin, and 100 μg/ml streptomycin. Transfections were performed with Fugene6 (Roche Applied Sciences). Stable cell lines were generated by transduction with lentivirus containing the cDNAs of interest, as previously described [70]. Cells were treated with the γ-secretase inhibitor DAPT (10 μM) for 6 or 16 h in Opti-MEM I (Invitrogen) to monitor CTF accumulation. HEK cells were treated for 6–16 h with the β-secretase inhibitor C3 (3 μM, BACE inhibitor IV) or the metalloprotease inhibitor GM6001 (15 μM), or for 6 h with the proteasome inhibitor epoxomicin (1 μM) or phorbol 12-myristate 13-acetate (PMA, 0.5 μM). All drugs were purchased from Calbiochem.
Full-length cDNAs were obtained from the National Institutes of Health Mammalian Gene Collection. The NPR-C and Itgβ1 cDNA were from mouse, and all other cDNAs were human. Expression constructs were C-terminally tagged with the FLAG epitope (DYKDDDDK) by inserting the sequence encoding FLAG into the 3′ primer before the stop codon. To generate artificial CTFs, the construct's signal peptide and putative CTF region were individually amplified by PCR with overlapping complimentary regions added to the primer. These two overlapping amplicons were then combined and PCR amplified with outside primers, as previously described [70]. Chimeric constructs were similarly produced using primers designed with overlapping regions. All expression constructs were verified by DNA sequencing.
Cells were lysed in 50 mM Tris-HCl (pH 7.4), 1% NP-40, protease inhibitor cocktail (Roche Applied Sciences), 2 mM 1,10-phenanthroline and 5 mM EDTA. Lysate was centrifuged at 1,000g for 10 min to remove nuclei. Protein concentrations were determined using a bicinchoninic acid-based assay (Pierce Biotechnology). Samples were then subjected to SDS-PAGE and Western blotting. APP was detected using the polyclonal antibody C9 (1:1,000) [71]; nicastrin with N1660 (1:2,500, Sigma); presenilin with MAB1563 (1:1,000, Chemicon); and FLAG tag with M2 (1:1,000, Sigma) or Rabbit anti-FLAG (F7425 1:1,000, Sigma). Western blots were probed with anti-mouse, anti-rabbit, or anti-rat secondary antibodies (1:10,000, Rockland Immunochemicals) and detected using the Odyssey infrared imaging system (LI-COR Biosciences). For co-immunoprecipitation experiments, cells were lysed in 1% CHAPSO with 25 mM Tris-HCl (pH 7.4), 100 mM NaCl, 2 mM EDTA and protease inhibitor cocktail. FLAG-tagged proteins were immunoprecipitated overnight with an M2 Affinity Gel (Sigma) and subjected to three washes with lysis buffer containing 0.5% CHAPSO before Western analysis. Immunoblots shown are representative of at least three experiments. Immunoreactive proteins were quantified using Odyssey Software v1.2 and the data were analyzed using a one-way analysis of variance with Tukey post-hoc comparison or a two-tailed Student t-test, where appropriate. Calculated comparisons of p < 0.05 were considered significant. All reported values represent the means ± standard error of the mean (SEM).
Purification of γ-secretase from the S-1 CHO cell line and preparation of recombinant substrates was performed as previously described [72]. Recombinant APP and Itgβ1 C100FLAG-tagged proteins consist of an N-terminal start codon (methionine) and a C-terminal FLAG tag joined to the 99-residue CTF, and were produced by bacterial expression and subsequent purification. APP and Itgβ1 have cytoplasmic domains of identical length. Purified γ-secretase was incubated with lipids (phosphatidylcholine and phosphatidylserine) and the C100FLAG substrate at 37 °C for 4 h. The reaction was assayed by Western blotting for the C-terminal FLAG tag. For co-incubation experiments where APP and Itgβ1 C100FLAG proteins were mixed, half the normal amount of each protein was combined so as to keep the total concentration of protein and detergent consistent.
|
10.1371/journal.pcbi.1002044 | Activated Membrane Patches Guide Chemotactic Cell Motility | Many eukaryotic cells are able to crawl on surfaces and guide their motility based on environmental cues. These cues are interpreted by signaling systems which couple to cell mechanics; indeed membrane protrusions in crawling cells are often accompanied by activated membrane patches, which are localized areas of increased concentration of one or more signaling components. To determine how these patches are related to cell motion, we examine the spatial localization of RasGTP in chemotaxing Dictyostelium discoideum cells under conditions where the vertical extent of the cell was restricted. Quantitative analyses of the data reveal a high degree of spatial correlation between patches of activated Ras and membrane protrusions. Based on these findings, we formulate a model for amoeboid cell motion that consists of two coupled modules. The first module utilizes a recently developed two-component reaction diffusion model that generates transient and localized areas of elevated concentration of one of the components along the membrane. The activated patches determine the location of membrane protrusions (and overall cell motion) that are computed in the second module, which also takes into account the cortical tension and the availability of protrusion resources. We show that our model is able to produce realistic amoeboid-like motion and that our numerical results are consistent with experimentally observed pseudopod dynamics. Specifically, we show that the commonly observed splitting of pseudopods can result directly from the dynamics of the signaling patches.
| Different types of cells are able to directionally migrate, responding to spatially-varying environmental cues. To do so, the cell needs to sense its environment, decide on the correct direction, and finally implement the needed mechanical changes in order to actually move. In this work we study the relation between the sensing-signaling system and the mechanical motion. We first show that membrane protrusions which drive the overall translocation occur exactly at the same locations at which membrane-bound signal-transduction effectors accumulate. These high concentration areas, also termed “patches”, exhibit interesting dynamics of disappearing and reappearing. Based on these findings, we develop a mathematical-computational model, in which membrane protrusions are driven by these membrane “patches”. These protrusions are then coupled to other cellular forces and the overall model predicts motion and its relationship to shape changes. Using our approach, we show that several observed features of cellular motility, for example the splitting of the cell tip, can be explained by the upstream signaling dynamics.
| Directional cellular migration is a widely observed phenomenon, ranging from mammalian cells to unicellular eukaryotes to bacteria. During development, as well as in mature organisms, cells respond to environmental cues and migrate to distant sites to perform different tasks, such as wound healing or immune response [1]. In other cases, cells respond to a nutrient gradient and migrate towards a food source [2], [3], or aggregate to form a multi-cellular slug [4], [5]. Directional motion according to external cues, known as chemotaxis, is typically controlled by signaling processes in the cell. Through signal transduction pathways, the external stimulation leads to internal symmetry breaking and to the formation of a distinct front and back. This sensing step is then coupled to cell mechanics, which is also governed by signaling processes which are highly conserved between different organisms [6].
In the last decade, many studies have been devoted to the characterization of different signaling components and systems in different organisms (see e.g. [7]–[9]). Other studies, both theoretical and experimental, have dealt with the biophysics of cellular motion including such aspects as actin polymerization, adhesion and myosin-based contraction [10]–[14]. However, an understanding of the coupling between the two systems – directional sensing and motility mechanics –is still incomplete, both from the experimental and the theoretical points of view. A modeling study of this coupling was undertaken in ref. [13], but from a perspective that does not build on observed correlations between these two parts of the overall chemotactic response. Yang et al. [15] used the level set method to link cell deformations with signaling events including PIP3 localization to calculate the pressure profile in a cell. However, their model was unable to predict experimentally observed cell shapes, probably because it did not take into account the complex signaling dynamics. In this paper we study how the signaling pattern and dynamics influence macroscopic features of cellular shape and motion, by using both experimental data and computational modeling. Specifically, we show that several experimental observations of cell motion can be explained by a better understanding of the spatio-temporal aspects of the aforementioned coupling.
In the social amoeba Dictyostelium discoideum, a large number of the signaling components have been identified through extensive genetic and biochemical investigations, along with their spatial intra-cellular distribution relative to an external gradient [7], [16], [17]. This distribution is usually non-uniform with several components located at the front while others are concentrated at the back of the cell [18], [19]. The cell motion is then accomplished by membrane protrusions at the front of the cell, along with retraction at the back. These protrusions are generated through the polymerization of actin filaments while the retraction is associated with cortical tension generated by actin-myosin interactions [20], [21]. For amoeboid cells, the protrusions take the form of pseudopods with finite life-times, leading to repeated cycles of extension and retraction.
One of the earliest measurable signaling events is the appearance of activated Ras, RasGTP, to the front of the cell [22]. This is then followed by the recruitment of other signaling molecules with a number of feedback loops [23]. Such experiments are typically carried out by exposing cells to a steep gradient originating from a pipette and do not address the subsequent motion of the cell. When exposed to a uniform stimulus, cells display a number of membrane “patches” in which the concentration of a signaling molecule is greatly increased. These patches have been implicated in the formation of pseudopods [24], [25] and RasGTP has been shown to co-localize with the site of F-actin polymerization in both chemotaxing cells and in cells undergoing random motility [22], [26]. Furthermore, it has been reported that RasGTP can drive localized actin polymerization via PI3K [27]. Recently, a number of features of chemotactic cell motion, including the rate of pseudopod formation, the distribution of de novo pseudopods and their persistence have been studied quantitatively [28]–[30]. This analysis revealed that the rate of formation of pseudopods is roughly independent of orientation of the cell with respect to the shallow gradient. Furthermore, it was argued that new pseudopods are not always located in the direction of the highest receptor occupancy, inconsistent with a deterministic “chemical compass” model [31], [32]. In fact, these experiments have been taken to imply the existence of a specialized tip splitting mechanism in which the location of a new pseudopod is highly correlated with the location of the current pseudopod from which it splits off.
Because the coupling of the directional sensing pathways to the motility machinery is currently not well understood, it has been difficult to develop detailed mathematical models that can simulate realistic cell motion. Most models to date have addressed distinct parts of motility, including retraction and protrusion [33], but are unable to describe the entire motility process; other models use ad-hoc rules to describe the motion [11], [12]. What has been lacking is a model that couples elements of the sensing machinery to cell motility.
Part of the challenge of developing models has been the lack of experimental data that reliably identify membrane protrusions mechanisms. Many experiments have been performed in assays where a thin horizontal subsection of the cell was visualized by confocal microscopy. Chemotaxing cells, however, can extend into the vertical direction. This vertical extent makes it difficult to quantify the correlation between the localization of signaling components and membrane extensions. In this paper, we examine the localization of RasGTP by GFP-tagged Ras binding domain (RBD). RBD-GFP intensity was measured along the membrane of Dictyostelium cells moving in a chemoattractant gradient and correlated with pseudopodal protrusions. This was done with cells restricted in the vertical direction such that fluorescent patches at the membrane could be visualized in a single confocal section and membrane protrusions could be quantitatively measured. In the first set of experiments, we used the well-established under-agar assay in which cells must lift a thin layer of agar as they move [34], while in the second set of experiments, we employed a microfluidic device in which the cells are constrained by the height of the chamber [35]. We used the results from these experiments to perform a quantitative analysis of the spatial correlation between signaling components and pseudopod extensions and found a strong spatial correlation between patches of RasGTP and membrane protrusions.
On the basis of these new experimental data, we develop a mathematical model for cell motion in which cell protrusions are driven by patches of an activator, qualitatively similar to the observed RasGTP patches at the front of chemotaxing cells. Our model incorporates a set of mechanisms that allow for the simulation of cell motion under a variety of experimental conditions and is studied here for the specific case of patch-driven chemotactic response in a static gradient. We show that our model produces realistic amoeboid-like motion and can demonstrate the effects of gradient steepness, cortical tension, and polarity on the cell shape and motion. Our model shows that the patch dynamics of membrane bound activators result in an apparent tip splitting behavior and that therefore an explicit splitting mechanism is not needed. Specifically, a patch in our model is stable and will not bifurcate into two or more spatially distinct patches before disappearance, as is the case in several physical systems [36]. Furthermore, we show that the apparent process of the cell “choosing” the better-oriented pseudopod [37] is simply an outcome of the disappearing-reappearing dynamics of the activator patches. Finally, we show that the results of automated pseudopod detection algorithms need to be carefully interpreted.
The Ras binding domain from human Raf1 binds strongly to RasG in the GTP bound form [22], [26], [27], [38]. We used RBD-GFP to track the localization of RasGTP at the cell membrane in chemotaxing Dictyostelium cells at 2 second intervals. Figure 1 shows several snapshots of these cells in both the under-agar assay (a) and the microfluidic assay (b). In both experimental setups, the vertical dimension of the cells was restricted (for more details see Methods), which ensured that most of the cell body remained in the focal plane of the microscope during its motion. In the microfluidic device, cells entered cross chambers only 2 µm high that connected parallel channels carrying buffer on one side of the cross chambers and 100 nM cAMP on the other. The length of the cross chambers varied from 650 µm to 100 µm thereby generating gradients of different steepness.
The cell speed in the under-agar assay, as well as the microfluidic devices, was found to be 8–10 µm/min. The chemotactic index (CI), defined as the ratio of the distance traveled in the direction of the gradient and the total distance, was 0.71 to 0.94 for cells in the microfluidic devices. It was not possible to compute a CI for the under-agar experiments since the precise direction of the gradients is not known.
We quantitatively compared the location of RBD-GFP patches and the location of membrane protrusions. Patches were detected using a global threshold for filtering background intensity and protrusions were detected by comparing the membrane location in successive frames (see Methods). The location of a patch in each frame can be defined by an angle θ between an arbitrary axis and the line connecting the center of the patch and the center of the cell. A similar angle φ can also be defined for the location of the pseudopod. An example of this analysis is presented in Figure S1 in the Supporting Information section.
To test the spatial correlation between the locations of patches and protrusions, we define a correlation function between a patch and a protrusion at frame i as(1)This correlation function takes on values between −1, corresponding to anti-correlated patch and pseudopod locations, and 1, corresponding to a patch location that coincides exactly with the pseudopod location. If patches and protrusions are completely uncorrelated this correlation function should average to zero (data not shown).
The correlation function of Eq. (1), for a particular cell in the microfluidic device, is shown in Figure 2 and remains close to the maximal value of 1 for most frames. We have analyzed three cells in the under agar experiment (305 frames, 297 of which showed both a patch and a pseudopod). We found an average correlation function of 0.83, 0.87 and 0.89 for co-localization of RasGTP patches and membrane protrusions. In the microfluidic device we analyzed eight cells, totaling 3421 frames of which 2289 showed both a patch and a pseudopod. Taking the data from all frames in both experiments that contain both a patch and a protrusion we found an average correlation function of 0.90 (±0.04). This correlation analysis implies a close relationship between protrusions and RasGTP patches: a new protrusion is accompanied by membrane-localized RasGTP accumulation in the same place in space. Furthermore, the fact that the correlation function for both assays is similar suggests that this relationship is independent of the experimental details. We have also tested five cells under uniform stimulation of cAMP (100 nM) and found an average correlation of 0.9 (±0.05). This indicates that the RasGTP-protrusion correlation is not specific to a gradient sensing process.
The measured strong correlation is consistent with a causal relationship in which a RasGTP patch almost always leads to a membrane protrusion. Previous experiments have demonstrated that Ras activation mediates leading edge formation, through activation of basal PI3K and other Ras effectors required for chemotaxis [22]. It was also shown that mutants with defective RasG exhibited a loss of directionality and severe loss of movement [22]. In this work, we focus on the spatial correlation, demonstrating that activated Ras localization and pseudopod formation occur at the very same location in the cell.
Based on the abovementioned results, we developed a computational motility model in which protrusions are generated by membrane patches of a putative chemical activator. The goal of the model is to allow for the study of the effects on cellular morphology of localized, transient protrusion forces, assumed to originate from the signaling system downstream of the patch dynamics. To do this requires embedding a patch generation mechanism into a full cellular mechanics simulation. In the absence of a complete understanding of all the relevant biophysical effects at the whole cell level, we opted for creating a relatively simple simulator, taking many experimental movies both from our own lab and from other groups (see [30], e.g.) as guidance. Later, we will discuss in detail which aspects of our results should be insensitive to some of the details of the mechanical model.
Following the aforementioned strategy, our motility model consists of two coupled modules: the first module contains a mechanism that creates transient localized patches while the second module describes the actual motion of the cell. We will consider a two-dimensional cell and will represent its membrane by a set of nodes. For the first module, we choose our recently developed excitable reaction-diffusion model which contains an activator field a and an inhibitor b [39]. Even though our model is not formulated at the level of specific biochemical components we can nonetheless use the activator a to mimic the observed behavior of activated Ras patches, so that their influence on downstream motility can be tested. The equations governing these fields can be written as(2)where Da and Db are the diffusivities of a and b, respectively, ε, β, and μ are constants and η is a noise term. This term is taken from a uniform distribution in the range [−1,1], but, to avoid overwhelming the system by simultaneous excitation of many coupled points, only a small fraction of the points (0.001–0.01%) are randomly given a non-zero noise term [39]. Such a noise pattern can be generated by feeding Gaussian white noise into a nonlinear excitable process (data not shown). Such processes have been directly demonstrated in genetic networks [40], [41]. In our previous work, we have shown that this model is excitable for a certain range of parameter values and that the inclusion of the noise term leads to the spontaneous formation of domains of high a. Due to the excitable nature of the model, these a-patches spontaneously disappear, followed by the appearance of new patches, similar to the observed RasGTP dynamics in our experiments and to the dynamics of PIP3 patches [24]; this has been discussed in detail elsewhere [39].
It should be noted that detailed and precise modeling of Ras dynamics is beyond the scope and purpose of this work. Once again, our goal is to test how the patch dynamics, and specifically its come-and-go nature, influence the macroscopic cell shape and motility. For this purpose, we only need a system that creates patches of one of the species, which can then be used as an activator for downstream processes. In fact, one can completely replace the patch dynamics by an artificial process which puts patches in by hand with the measured distributions, and recover all of our results (data not shown).
The excitability of the system is controlled by the parameter β : below β<0.6 the system is highly excitable while above this value the excitability is significantly reduced. Thus, varying this parameter along the cell boundary determines the rate of patch formation and choosing the front of the cell to be excitable while the back of the cell is unexcitable will lead to patch formation concentrated at the cell's front. Here, we do not explicitly concern ourselves with modeling the gradient sensing mechanism that detects the external chemical concentration field and determines β. Indeed, how a cell determines its front has been the subject of many theoretical studies [42], [43]. Here, we directly assume that front determination is accomplished through the formation of an internal compass. The direction of this compass is determined by the receptor occupancy and is therefore dependent on the external gradient direction. Specifically, we choose the internal compass direction, φint, to be the external direction φext plus some added noise:(3)The term ηφ represents all the possible fluctuations in the directional sensing process and is drawn from a Gaussian distribution with zero mean and width σ (see Figure 3). We assume that the width of the noise distribution is inversely proportional to the steepness of the gradient such that the directional sensing process is more accurate for steeper gradients (Figure 3) as is reflected in the increased chemotactic indices of cells responding to steeper gradients [44]. The front of the cell is then chosen to be the point on the membrane that is closest to the direction of the internal gradient φint. Once the angle is determined, β is chosen to be peaked around the front with a width that inversely depends on a (dimensionless) polarizability parameter p. This parameter p determines how abruptly β changes with the distance from the cell front and, hence, has an impact on the width of the excitable zone on the membrane. A high value of p corresponds to cells with a smaller width of the excitable zone and, thus, to more polarized cells, while a low value of p represents a larger excitable membrane zone and less polarized cells. The precise form of the excitability along the membrane is given in the Supporting Text S1.
The noise distribution width σ, the polarizability parameter p and the excitability β represent different aspects of the cellular response. The response of the cell depends on its polarization level: in relatively symmetric cells characteristic of early developmental stages, projections extend all along the cell's periphery, while in polarized (elongated) cells projections only extend at the cell's front [45]. This change is represented by p, which is an internal property of the cell and is hence independent of the external conditions. The polarization and gradient are connected to the patch formation mechanism through the parameter β (see also [39] and the Supporting Text S1 and Figure S2). As mentioned above, this parameterization is aimed at realistically describing the signaling system, so that the influence of different signaling behaviors on the cell motility can be tested.
The second module is responsible for cell motion through the definition of a force on each node, taken to be normal to the cell membrane:(4)In this equation, the first term results from the localization of a and couples the signaling module to the motility module. Specifically, motivated by our experimental results, this protrusion force is assumed to depend on the concentration of the activator a (describing RasGTP) in the first module. For simplicity, we have chosen a simple linear dependence as detailed in the Supporting Text S1. Using other forms, including those with a non-linear dependence, yielded essentially similar results.
In the actual cell, the relationship between the chemical driving and eventual actin polymerization leading to protrusion forces is rather complex. Under most conditions, our cell simulator is able to ignore all of these complications and get by with the simplest possible linear relationship. However, we show in Video S1 and in Figure S3a that for the case of driving the cell with two strong sources on opposite sides of the cell (see for example [30]) that this model is unable to capture the fact that pseudopods must eventually compete with each other and only one can win in the long run. We have therefore added one extra part to this patch chemical – protrusion force relationship, making it depend on a global resource G(t) which is consumed by the pseudopod construction process (see for example [45], where the authors state that cytoskeletal or membrane components are probably limited, causing the cell to occasionally “freeze”). Video S2 and Figure S3b show that, indeed, adding this effect yields the observed cell behavior. The details of how G is dynamically determined are discussed in the Supporting Text S1. For the case of chemotactic motion to a simple gradient, the case of primary interest here, this feature is relatively unimportant (see later).
The second term in the right side of Eq. (4) describes the cortical tension, which depends on the local curvature κ. γ represents the membrane rigidity, with higher values of γ corresponding to more rigid membranes. κ0 is the spontaneous curvature of the cell, which is the equilibrium curvature when the total force is zero, namely for a circular cell of radius R. Due to the differences in the acto-myosin cortex structure around the cell versus the protrusion area, we take γ to depend on position along the cell membrane. Recent experiments [46]–[48] have revealed that the tension is higher at the back, where presumably myosin bundles and crosslinks the cortical actin layer, as opposed to the front of the cell; thus we choose the back part of the cell, defined as the portions of the membrane for which a<0, to have a cortical tension (γ1) that is about twice as high as the cortical tension (γ2) in the front part of the cell where a>0. In addition, we have empirically discovered that in order to produce pseudopods with large aspect ratio, i.e. long and narrow, and to get the “valleys” between pseudopods to have a reasonable shape, we need to allow regions of the membrane with negative curvature to have a value γ3 that is smaller yet. A possible origin of this effect lies in that we are using a two-dimensional model to describe a three-dimensional cell (albeit moving within a limited three dimensional space). The tension force in 3D should of course be proportional to the total curvature and it might be the case that negative in-plane curvature tends to cancel the positive out-of-plane curvature, resulting in small net effect. In the Supporting Information (Text S1 and Figure S4) we show how this effect modifies cell shape dynamics and makes them more “biological”. It is important to note though that this additional assumption is not necessary in order to obtain the primary conclusions of the paper, which is the relation between signaling (RasGTP patches) and pseudopods and the implications for tip splitting.
The third term ensures that the cellular area A (which is the equivalent of the cellular volume in the 3D case) remains constant and can be viewed as an effective pressure. Finally, the last term represents an effective drag force, proportional to the local velocity v, and determines the time a pseudopod continues to move after the protrusion force has vanished. This term also yields a limit on the maximal speed, so that a constant force in one direction results in a constant speed rather than an unrealistic constant acceleration. A complete list of the parameter values can be found in the Supporting Text S1. The evolution of each node is found by solving(5)
The entire simulation is performed in the following sequential steps: First, the reaction-diffusion equations (3) are solved on the entire membrane to find the value of the activator a at each point. Second, the force on each node is computed using Eq. (4). Finally, the nodes are advanced simultaneously according to Eq. (5). The time scale in the simulations can be converted to physical units by comparing the simulation cell speed to the cell speed obtained in the experiments and by taking a cell length that is comparable to the experimental dimensions of a cell. The internal compass in our simulations is updated every 2 minutes and additional computational details, including a description of adding and removing nodes, can be found in the methods section. A schematic diagram of the model cell as well as movies of several simulations can be found in the Supporting Information. The cell shape and motion both seem qualitatively realistic, and specifically, the formation, retraction and bifurcation of pseudopods resemble those seen in real cells.
Snapshots of typical simulation runs are presented in Figure 4 where the cell contour is tracked over time for various parameter sets. All the simulations presented in Figure 4a–f were run for the same time period but note the 25% difference in y-axes, the distance traveled by the cell, in Figure 4a–d versus Figure 4e–f. The computational cell in Figure 4a (also shown in video S3 in the Supporting Information) functions as a reference cell and has a CI of 0.966. Decreasing the gradient steepness, through a larger value of the parameter σ as in Figure 4b (and Video S4), leads to a smaller value of the CI (0.778), which is consistent with experimental results [44]. Figure 4c (and Video S5) shows a cell with a smaller value of the internal polarizability parameter p that determines the width of the excitable region along the membrane. This cell has a reduced speed, is less elongated than the reference cell of Figure 4a, but has only a slightly reduced CI (0.931), which is consistent with our observations for less developed cells as well other experimental data [7]. In Figure 4d, we show a trajectory of a cell with high cortical tension, parameterized by γ1 and γ2. This cell exhibits fewer pseudopods but its speed is similar to that of Figure 4a and its CI is 0.954, which is very close to the CI of the reference cell. The model therefore predicts that the cortical tension does not strongly influence the CI of the cell, but does influence the frequency of pseudopod formation. In Figure 4e–f the magnitude of the friction parameter is varied, with low friction (Figure 4e) resulting in long pseudopods and increased cell speed and high friction (Figure 4f) resulting in shorter pseudopods and reduced cell speed.
Our model is able to capture several qualitative features of Dictyostelium motility. For example, the experiments show that some pseudopods are maintained while others are retracted (Figure 5a). Pseudopods that are aligned with the gradient were found to have a higher probability of being maintained, and vice versa [30]. Interestingly, these maintained pseudopods exhibit “come-and-go” RasGTP patch dynamics in which a patch appears, disappears, and then re-appears, all at the same location, as can be seen in experiment and is shown in Figure 5a. This come-and-go patch dynamics is also observed in the results from our numerical simulations, as shown in Figure 5b. Furthermore, new pseudopods in our experiments are often created close to a previous one (Figure 6a), consistent with previous experimental studies where it was characterized as tip splitting [29], [30]. Importantly, our simulations also exhibit this tip splitting (Figure 6b), even though our model does not include any specific splitting mechanism.
To further investigate this apparent tip splitting behavior, we generated numerical cell data and analyzed these data using the same software as in previous experimental studies ([28], [29], [49] see Methods for more details). To this end, the locations of the membrane nodes were recorded and the contour of the simulated cells was computed using Matlab (Mathworks, Natick, MA). This cell contour was then used to create a full “cell body” by interpolating the discrete node locations and identifying the points that are inside the closed contour. The movement and shape of the simulated cell were analyzed using Quimp3 [49] to extract pseudopod statistics. The results are shown in Figure 7a. The angles of new pseudopods show a clearly bimodal distribution similar to that obtained by Bosgraaf et al. [29], implying the presence of a tip splitting mechanism. However, the distribution of patches that drive the membrane protrusions of the simulated cell, for the same numerical data set, is unimodal with a maximum at an angle that corresponds to the gradient direction (Figure 7b). The equations preclude a bimodal distribution of angles of new pseudopods resulting from this distribution of patches, since every pseudopod results from a patch. The apparent bimodal distribution must therefore be generated by the pseudopod detection algorithm.
One important issue concerns the relative importance of the transient nature of the patch dynamics versus the global resource limitation in limiting the extensions of the pseudopods. Figure 8a shows G(t) for the simulation corresponding to the cell tracks shown in Fig. 4a. Clearly resource limitation is playing an important role and for this case the patch dynamics are mostly responsible for the initiation but not the cessation of protrusions. But, this is not necessary. In Fig. 8b we show a cell track example where we change the parameters of the chemical module to speed up patch dynamics (see details in the Supporting Text S1). As is seen in Figure 8c, G(t) oscillates but rarely dips down into the region where it limits protrusion; instead the patch dynamics is self-limiting. Altering the model in this manner does not change the aforementioned results regarding the response of the cell to varying the gradient strength and regarding the true source of apparent tip splitting seen in experimental studies.
Several recent studies have demonstrated that Ras activation is upstream of F-actin polymerization in a causal sequence leading to the formation of membrane protrusions and pseudopods [7], [22], [26], [27]. Our experimental assays on “2D” cells have been able to quantify the extent of spatial correlation between membrane areas of Ras activation (patches) and protrusions. We employed two different experimental assays to quantify this spatial correlation. In one set of experiments, we used the standard under-agar assay in which the vertical extent of cells was restricted by a thin layer of agar. We found a high spatial correlation between RasGTP patches and pseudopods in all analyzed cells, indicating that activated Ras and membrane protrusions occur at the same membrane location. However, drawbacks of this assay are that the vertical dimension is not known since cells are able to lift the agar to an unknown degree and that gradients are difficult to characterize. To overcome these shortcomings, we also performed experiments in microfluidic devices in which highly reproducible gradients with a well-defined direction and steepness are produced. Furthermore, the distance between the substratum and the roof of these devices is precisely specified (2 µm). We found that for cells in these devices the spatial correlation between patches and pseudopods was also large and comparable to the correlations found in the under-agar assay (Figure 2). This is also found to be the case for cells under uniform stimulation. Thus, the high spatial correlation between the patches and pseudopods appears to be insensitive to the details of the assay. We expect that this correlation is also large for cells that can extend freely in the vertical direction, however, this is difficult to determine since it requires a series of confocal scans in the vertical plane at each time point and significantly restricts the period of time a cell can be followed before suffering the effects of phototoxicity.
Establishing a high spatial correlation between the locations of active membrane regions and extending pseudopods is consistent with a causal relationship between the two and led us to create a model in which patches govern the location of membrane extensions. Our aim was to test how the dynamics of the signaling components influences the overall cell motility and shape dynamics. Our motility model addresses the two key ingredients, patch formation and pseudopod extensions, using two coupled modules that are responsible for obtaining realistic numerical cell shape and motion. First, the patch module is responsible for the creation of transient patches, as observed in the experiments. Second, the motility module incorporates a number of relevant forces acting on the cell membrane, including a term that couples the dynamic activator to the protrusive force. This modeling approach is distinct from previous attempts which mainly address specific stages of cell motility such as protrusion, adhesion or contraction [11], [33] or use a rule-based approach [12].
Our model, however, is still highly simplified. Since the biochemistry and specifically the exact reactions between the signaling components are still not fully known, the signaling module is mostly designed to replicate the experimental results so that the influence on the motility can be tested. We note that a recent paper by the Devreotes group introduces a very similar excitable medium approach [45]. Our motility module also simplifies a number of steps involved in generating membrane protrusions. For instance, the coupling between the patch module and the motility module, responsible for the protrusive force, is taken to be simply linear but may be more complex. Furthermore, the adhesion forces between the cell and the substratum are not explicitly modeled and are subsumed in the overall set of forces acting on the cell, as pushing forward is only possible in the presence of anchoring points. Also, a possible contribution from the bending energy is ignored. Despite these simplifications, our model is able to capture realistic cell behavior and shapes during chemotaxis (Figure 4), and provides insights into how dramatic changes in the cell shape can result from small changes in the signaling dynamics.
Our model contains a number of parameters that can be varied to mimic different experimental conditions. For example, the determination of the front of the cell in our model is a process that is subject to noise. This noise is taken from a distribution with width σ and the strength of the gradient can be adjusted by changing this width: a steep gradient corresponds to a narrow distribution (and small σ) while a shallow gradient corresponds to a wide distribution (and large σ). In agreement with experiments [44], [50], [51], we find that the CI is maximal for steep gradients and is reduced' for shallow gradients (Figure 4b).
Another parameter of the model, p, represents the cell's polarizability and controls the excitability change along the cell perimeter. As can be seen from Figure 4c, high values of this parameter lead to cells with a high CI, an elevated cell speed, and elongated cell shapes; this is the typical behavior of highly polarized Dictyostelium cells [7]. In contrast, low values of p result in rounder cells with a lower speed and lower CI, which is typical of cells in early developmental stages (see also Supporting Videos).
The cortical tension in our model is represented by the parameter γ, with high values of γ corresponding to a more rigid membrane. Not surprisingly, increasing the cortical tension leads to cell motion with fewer lateral pseudopods (Figure 4d). A direct comparison with experimental phenotypes is difficult since a quantification of the cortical tension in cells is problematic. However, it is commonly assumed that myosin is involved in establishing cortical rigor. Myosin mutants which have reduced cortical tension display more lateral pseudopods than wild-type cells and move more slowly [52], [53].
Recent studies of Andrew and Insall investigated chemotactic motion of Dictyostelium cells in the under-agar assay and presented evidence that new pseudopods were made in spatially restricted sites by splitting of the leading edge [30]. Furthermore, they found that pseudopods were generated at relatively constant intervals, independent of the orientation of the cell relative to the gradient, and that the survival and retraction of pseudopods were spatially controlled such that pseudopods aligned with the gradient were more likely to be maintained. They reasoned that their results contradict chemical compass models in which cells generate new pseudopods at the location of highest receptor occupancy (the needle of the compass) [31], [54]. They argued that cells guide their motion through a mechanism in which a new pseudopod splits off an existing one. A similar conclusion was reached by Bosgraaf and van Haastert, who analyzed a large number of pseudopodal extensions in chemotaxing Dictyostelium cells and found that the distributions of angles between the current and next pseudopod were bimodal with the peaks located at ±50° from the gradient direction [29].
Our model allows us to compare numerically obtained cell dynamics with these recent experimental observations. First of all, the reaction-diffusion model of the patch module is excitable and generates patches in a stochastic fashion. This guarantees that patches occur at rates that are set by the reaction-diffusion model and are independent of the cell's direction, consistent with experimental observations. Also, our model produces cell dynamics that resemble the tip splitting events observed in the experiments. In physical systems, tip splitting is a consequence of a spatial instability of the tip, resulting in the formation of multiple tips [36]. Such an instability, however, is not present in our model since an existing patch is stable, demonstrating that the observed events do not require an explicit tip splitting mechanism. The apparent tip splitting in our model is demonstrated in Figure 6b where a new patch appears close to the old one, leading to a new pseudopod that appears to split off from the old pseudopod.
The underlying patch dynamics can also explain the experimental observation that cells maintain pseudopods that are aligned with the direction of the gradient. As shown in Figure 5b, numerical patches can exhibit come-and-go dynamics characterized by the appearance of a patch, followed by its disappearance and re-appearance in roughly the same location. This repetitive patch formation at the same spatial location is more likely to occur in the direction of the gradient than away from the gradient. The accompanied membrane protrusion, however, does not necessarily exhibit this come-and-go dynamics, as protrusion initiation and cessation are smoothed and are not as abrupt as the upstream signaling. The time during which a pseudopod continues to move forward after the protrusive force has vanished is controlled in our model by the effective friction force parameter λ. This parameter represents the effective lag between the signal and its downstream response, for example due to the time needed for the process of actin polymerization. As a result, a series of consecutive but separate patches at the same location can lead to what looks like a “winning” single pseudopod.
It should be noted that in our model, the phenomenon of tip splitting results solely from the come-and-go dynamics of the patches, and is independent of other components of the model such as the global resource limitation, cortical tension and the specific form of the forces. All of these are needed for a realistic cell shape, but do not alter the main conclusions of our work, namely the effects of the signaling dynamics on the observed pseudopod behavior.
Pseudopods that are directed in the gradient direction and that have long apparent lifetimes can also underlie the experimentally observed bimodal distribution of pseudopod angles. Indeed, when we compared the pseudopodal angle distributions in numerical cell tracks using the automated software package Quimp3, we found a bimodal distribution even though the patch distribution exhibits a single peak in the direction of the gradient (Figure 7). Since our model does not contain an explicit tip splitting mechanism, and every pseudopod necessarily originates from a patch, this bimodality is purely an outcome of the algorithm, which detects a new pseudopod by identifying two spatially separated negative-curvature zones. The bimodal distribution produced by Quimp3 may result from undercounting pseudopods at zero angle and/or the elongated shape of cells.
In our model, the location of new pseudopods is determined by the location of patches, which are themselves controlled by the direction of the internal compass. The timescale for updating this internal compass is a parameter in our model and controls the persistence of the motion: a small timescale will lead to cells that change directions more often than cells with a larger timescale. Experimental values for this timescale, and how it depends on the external conditions, are presently unclear. The direction of the internal compass, and specifically its deviation from the external direction, is determined by the steepness of the gradient (Figure 3). For shallow gradients, the distribution of compass locations is wide while for steep gradients this distribution is narrow. As a direct consequence, we predict that the ratio between split and de novo pseudopods in shallow gradients is lower than in steep gradients. Experiments have only compared this ratio for cells in buffer and for cells in a gradient [29]. These experiments found, consistent with the above arguments, that the ratio is smaller for cells in buffer and extending this comparison for different gradient parameters would be interesting.
Our model contains a noisy internal compass with a direction that depends on the external gradient direction through our excitability parameter β and a noise level that is inversely proportional to the steepness of the gradient. In previous work, it was suggested that the generation of pseudopods at a constant rate and the generation of pseudopods in the “wrong” direction contradict the existence of such an internal compass [30]. However, our excitable reaction-diffusion system can produce patches at a constant rate. Furthermore, our results show that the internal compass model, even though it occasionally exhibits pseudopods directed in the wrong direction, is able to produce highly directed motion. Thus, our model is consistent with experimental results and indicates that cells might utilize an internal compass to direct their motion.
Since we want our cell simulator to behave in a robust manner even for more complex chemical driving fields, we have introduced several features of the motility module which do not appear to be essential for the case of primary interest here, namely motion in a stable, static gradient. Studies in which cells move in more complex environments, replete with obstacles and/or multiple sources, will be presented elsewhere; for those cases the global resource constraint is needed to ensure that eventually the cell moves in only one direction (see Supporting Figure S3) and the flexibility of negative curvature is needed (see Supporting Figure S4). We did not try to define a minimal model that would work only under more limited scenarios. Instead, our strategy was to embed patch dynamics in as realistic a motility module as we could infer from the data and then verify that our conclusions regarding cell shape, chemotactic index, and tip splitting were not affected by these more global considerations.
In conclusion, our model can capture several qualitative features of experimental cell motion. In particular, it is able to duplicate apparent tip splitting dynamics, apparent spatial control of pseudopod retraction, and the relatively constant rate of pseudopod formation. It is important to stress that these phenomena are produced without invoking a tip splitting mechanism suggesting that such a mechanism is not required in chemotaxing Dictyostelium cells. Furthermore, our model incorporates the notion of an internal compass which determines how the external gradient direction controls the locations of patches. Key in our model is the fact that our compass is subject to fluctuations. These fluctuations lead to a distribution of patches that is centered around the gradient direction but with a width that depends on the gradient strength. Thus, the needle of the compass is not necessarily pointing in the direction of the highest receptor occupancy at all times but fluctuates, leading to apparent tip splitting.
Based on the experimental results, our model connects the two major components in cellular chemotaxis, namely signaling and motility. We show that the dynamics of signaling molecules is related to the cell motility in a direct and localized manner, and this connection can explain a large amount of currently available data. This model was designed for the relatively simple system of Dictyostelium chemotaxis, but can also be extended to describe other types of cells such as immune cell migration, neuronal growth cone motility and cancer metastasis. We believe that highly interesting and valuable insights can be gained by focusing on the interplay between signaling and motility.
Microfluidic devices, originally designed and used to study gradient sensing in yeast [55], were modified to study cell migration in a “2D” environment by decreasing the height of the test chambers from 5 to 2 µm [35]. In brief, these devices consist of an array of parallel rectangular test chambers of various lengths between two flow channels that are 80 µm high. Continuous flow of buffer with zero or 100 nM cAMP in the flow channels creates stable linear gradients in the test chambers, with slopes determined by 100 nM/w, where w is the width of the chambers and varies from 100–650 µm.
Plasmid pDM115, a non-integrating vector containing the Ras binding domain of Raf1 tagged with GFP and driven by the actin15 promoter, was a gift from the van Haastert lab. Transformants of D. discoideum strain AX4 carrying this vector were selected for hygromycin or G418 resistance.
Exponentially growing cells were harvested from growth media by centrifugation, washed twice in KN2/Ca buffer (14.64 mM KH2PO4, 5.35 mM Na2HPO4, 100 µM CaCl2, pH 6.4), then resuspended at 5×106 cells/ml and shaken for 5 hrs with 50 nM pulses of cAMP every 6 minutes to induce development Chemotaxis under agar was performed as previously described by Andrew and Insall [30]. Exponentially growing cells were also pulse-developed prior to loading into the microfluidic flow channel carrying buffer without cAMP. Cells were given 10 minutes to settle onto the coverslip prior to establishment of the gradient. Cells were imaged as they migrated across the test chambers. Fluorescent images (488 nm excitation) were captured every 2 seconds with a 63× oil objective on a spinning-disk confocal Zeiss Axiovert microscope equipped with a Roper Quantum 512SC camera. Images were collected using Slidebook 5 (Intelligent Imaging Innovations, Inc.).
For each frame, the contour of the cell was extracted and areas of cytosolic high intensity fluorescence were filtered out. Membrane areas of high intensity were detected using a threshold algorithm. The threshold value was adjusted to the movie characteristics, and usually taken to be within 10% difference from the maximal intensity. Protrusions were determined using the difference in membrane location between consecutive frames. Negative protrusions, i.e. inward motion of the membrane or retraction of a pseudopod, were filtered out in this analysis. In the case of several patches and several protrusions, the high-intensity points were clustered using the dendrogram algorithm [56] based on their Cartesian distances in space, so that well-distinguished, separate patches were obtained. The protrusion points were also clustered, and then each protrusion cluster was paired with the nearest patch (see Supporting Text S1).
The cell membrane was parameterized by 100–200 nodes, conveniently stored as a double-linked list. The nodes represent the 1D membrane of the cell (see Figure S5 in the Supporting Information). To ensure sufficiently smooth variations of a along this membrane, we solved the reaction-diffusion equations (2) on a refined array of 5000 points. This was achieved by attaching a sub-array of points to each node in the membrane linked list.
The total number of membrane nodes is not constant and nodes are added and removed to keep the distance between them within a given range. When a pseudopod is extended, nodes are added at the tip where the membrane “stretches” and removed at the back of the cell. Care was taken such that the total amount of a and b remained constant during this reparametrization. Our default parameter set for the reaction-diffusion module and for the motility module is given in the Supporting Information.
The list of node locations was recorded every 500 iterations and used to construct the cell contour and the cell body. The cell was drawn using Matlab and the separate frames were constructed into a movie. This movie was later analyzed using Quimp3, an automated pseudopod-tracking algorithm [49].
|
10.1371/journal.pbio.1000147 | Plant Insecticide L-Canavanine Repels Drosophila via the Insect Orphan GPCR DmX | For all animals, the taste sense is crucial to detect and avoid ingesting toxic molecules. Many toxins are synthesized by plants as a defense mechanism against insect predation. One example of such a natural toxic molecule is l-canavanine, a nonprotein amino acid found in the seeds of many legumes. Whether and how insects are informed that some plants contain l-canavanine remains to be elucidated. In insects, the taste sense relies on gustatory receptors forming the gustatory receptor (Gr) family. Gr proteins display highly divergent sequences, suggesting that they could cover the entire range of tastants. However, one cannot exclude the possibility of evolutionarily independent taste receptors. Here, we show that l-canavanine is not only toxic, but is also a repellent for Drosophila. Using a pharmacogenetic approach, we find that flies sense food containing this poison by the DmX receptor. DmXR is an insect orphan G-protein–coupled receptor that has partially diverged in its ligand binding pocket from the metabotropic glutamate receptor family. Blockade of DmXR function with an antagonist lowers the repulsive effect of l-canavanine. In addition, disruption of the DmXR encoding gene, called mangetout (mtt), suppresses the l-canavanine repellent effect. To avoid the ingestion of l-canavanine, DmXR expression is required in bitter-sensitive gustatory receptor neurons, where it triggers the premature retraction of the proboscis, thus leading to the end of food searching. These findings show that the DmX receptor, which does not belong to the Gr family, fulfills a gustatory function necessary to avoid eating a natural toxin.
| Plants evolve to fend off the insects that attack them, often by synthesizing compounds toxic to insects. In turn, insects develop strategies to avoid these plants or resist their toxins. Some plant toxins are nonprotein amino acids. For example, seeds from numerous legumes contain high amounts of l-canavanine, a nonprotein amino acid that is structurally related to l-arginine and is highly toxic to most insects. How insects can detect l-canavanine remains to be elucidated. Using pharmacology, genetics, and behavioral approaches, we show that flies sense l-canavanine using the receptor DmX, an orphan G-protein–coupled receptor that has diverged in its ligand binding pocket from metabotropic glutamate receptors. Disruption of the DmXR gene, called mangetout (mtt), suppresses the l-canavanine repellent effect. DmXR is expressed and required in aversive gustatory receptor neurons, where it triggers the premature retraction of the proboscis, thus leading to the end of food searching. Our results indicate a mechanism by which some insects may detect and avoid a plant toxin.
| Taste is essential to distinguish between nutritious and toxic substances. To avoid eating toxins, animals are able to detect them by using a repertoire of taste receptors [1]. Although it is recognized that a bitter taste sensation is critical to avoid toxic substances [2],[3], the cellular and molecular mechanisms that have been established during evolution to detect a toxin are not well understood. In particular, how a receptor becomes tuned to a toxin is not well documented, mainly because the structure of its ligand binding pocket (LBP) and the evolutionary relationship with the ancestor receptor are not known.
In Drosophila, the family of gustatory receptors (Grs) is predicted to consist of 68 genes [4],[5]. This family of receptors, which consist of seven transmembrane domain proteins, is characterized by a very high level of amino acid divergence, showing as little as 8%–12% amino acid identity [5]. Such diversity suggests that the Gr family could cover the entire range of taste-receptive capability of the fly. Nevertheless, the extreme divergence within this family does not exclude the possibility of evolutionarily independent insect taste receptors not belonging to the Gr family. To date, only few receptors of the Gr family have been associated with a specific taste molecule: for example, the receptor for the sugar trehalose, called Gr5a [6], and the bitter compound caffeine coreceptors, called Gr66a and Gr93a [7],[8].
Plants synthesize many toxic molecules as defense mechanisms against predation [9],[10]. A number of such toxic compounds are nonprotein amino acids [11],[12]. The best-characterized example of nonprotein amino acid that plays a defensive role is l-canavanine (2-amino-4-guanidinooxybutyric acid) [13]–[15], which is massively accumulated in the seeds of many legumes (up to 143 mM in Medicago sativa [16]). l-Canavanine is a natural insecticide because it is structurally similar enough to l-arginine (Figure S1) to interfere with l-arginine metabolism and to be incorporated by arginyl-tRNA synthase in de novo proteins resulting in dysfunctional proteins [17]–[19]. Thus, these properties of l-canavanine render it a highly toxic secondary plant constituent [15]. To deal with this natural poison, some insects have generated several adaptive strategies. Indeed, the tobacco budworm Heliothis virescens uses detoxification [20] and the beetle Caryedes brasiliensis feeds exclusively on l-canavanine–containing seeds but catabolizes l-canavanine to l-canaline and urea [21]. However, these two mechanisms to circumvent the toxic properties of l-canavanine are specific to few insect species. Thus, the evidence for a protective function against predation for such nonprotein amino acids, i.e., whether and how insects are informed that plants contain l-canavanine, remains to be shown [15].
Amino acids are known to be the ligands of G-protein–coupled receptors (GPCRs) belonging to the family C [22],[23]. All members of this family display a common structural architecture characterized by a long N-terminal extracellular domain containing a bilobular LBP [24], a seven transmembrane domain, and an intracellular C-terminus. This family includes metabotropic glutamate receptors (mGluRs). In mammals and in insects, mGluRs, which are activated by the neurotransmitter glutamate, play different roles in the central nervous system [25],[26]. We have previously shown that one mGluR has diverged through evolution to give rise to the mX receptor, called DmXR in Drosophila [27]. Orthologs of DmXR are so far only found in insects [27]. DmXR differs from mGluRs in the distal part of the LBP, so that this receptor is an orphan receptor, which is not activated by glutamate [27]. However, we previously showed that the DmXR and mGluR LBPs share the crucial residues necessary to bind a ligand with amino acid structural properties [27]. To deorphanize the DmX receptor, we previously tested various molecules having such properties, including all the classical amino acids, and did not find any ligand [27].
Drosophilidae are saprophytic animals, and members of dipteran families such as Tephritidae or Scatophagidae are seed predators [28], so we asked whether l-canavanine could activate DmXR. Here, we show that l-canavanine is a ligand of DmXR in vitro. We then wondered whether insects could be informed that plants contain l-canavanine via the mX receptor. We have addressed this question by using Drosophila as an insect model. First, we confirmed that l-canavanine is highly toxic when ingested. We then tested whether Drosophila avoid eating food containing l-canavanine. We found that l-canavanine is recognized by flies and mediates a behavioral avoidance response via a chemosensory mechanism. Hence, l-canavanine is a repellent. We then analyzed the molecular and cellular bases of l-canavanine–induced repulsive behavior, using gustatory behavior, pharmacology, and genetic approaches. We found that l-canavanine is detected in vivo by the DmX receptor. To control the l-canavanine avoidance behavior, the DmX receptor is expressed and required in bitter-sensitive gustatory receptor neurons (GRNs). These findings show that the gustatory detection of a natural toxin relies on DmXR, a divergent mGluR not belonging to the Gr family.
To test whether l-canavanine could activate DmXR, we transiently expressed this receptor in human embryonic kidney (HEK) cells and assayed for l-canavanine–induced DmXR activation. We found that l-canavanine activated HEK cells expressing DmXR (Figure 1A and Figure 1B). In contrast, the close structural homolog l-arginine showed no agonist or antagonist effect on DmXR (Figure 1C). l-Canavanine did not activate HEK cells expressing the unique fly mGlu receptor, DmGluA (Figure 1A). We searched for an antagonist and found that N-methyl-l-arginine (NMA) inhibited DmXR activation by l-canavanine (Figure 1C and 1D). Previous sequence analysis, mutagenesis, and 3-D modeling studies had shown that the LBP of the DmXR is very homologous to the LBP of mGluRs [27]. The residues contacting the amino acid moiety of glutamate (the α-COO− and NH3+ groups) are conserved in DmXR (e.g., Thr-176), whereas the residues interacting with the γ-carboxylic group are not [27]. The Thr-176 residue is conserved in all mGluRs, and its mutation strongly decreases the affinity of these receptors for glutamate [24]. Similarly, l-canavanine did not activate Thr-176–mutated DmX receptor (DmXRT176A) transfected in HEK cells (Figure 1A), although this mutated receptor is actually localized at the plasma membrane (as shown in [27]). This indicates that the plant amino acid binds into the LBP. Altogether our data show that DmXR is a l-canavanine receptor because of a partial modification of its LBP from the LBP of mGluRs. These results suggest that Drosophila may be able to detect l-canavanine in vivo through the DmX receptor.
Since l-canavanine is described as a natural insecticide [15], we first examined whether ingested l-canavanine is also toxic for Drosophila melanogaster. We maintained 50 young wild-type (WT) flies on Drosophila medium containing 10 mM l-canavanine and compared their viability and their fecundity to flies maintained on medium without l-canavanine (n = 8). When flies fed on 10 mM l-canavanine, we did not observe massive mortality or dramatic decrease of the lifespan. However, all the offspring of flies constrained to eat 10 mM l-canavanine died during larval stages (number of offspring in control medium >1,000, number of offspring in 10 mM l-canavanine = 0). These results indicate that Drosophila is a l-canavanine–susceptible insect.
Because of its toxicity, we hypothesized that Drosophila may avoid eating l-canavanine if they have the choice. To test this, we performed a two-choice feeding preference test. This behavioral assay measures the consumption of a sucrose solution (5 mM) colored by two food dyes (blue or red) offered simultaneously to adult fly populations [29]. After 2 h in the dark, flies are counted on the color dye witnessed in their abdomen. In the control situation, WT flies preferred the blue solution, the preference index (PI) being 0.82±0.04 (Figure 2). We then added increasing concentrations of l-canavanine to the blue solution (1 mM to 40 mM). We found that l-canavanine inhibited the intake of the blue solution (Figure 2), leading to a symmetrical increase in the intake of the red solution (unpublished data). The l-canavanine effect increased with concentration, reaching a plateau at 30–40 mM. At these concentrations, the PI for the blue solution is 0.13±0.06 (Figure 2). A similar repulsive effect of l-canavanine was also visible when the drug was added to the red solution (unpublished data). To determine whether this repulsive effect was mediated by a chemosensory mechanism, we used flies carrying an adult-viable mutant allele of the pox-neuro (poxn) gene. In homozygous poxn flies, external chemosensilla are deleted or transformed into mechanosensilla [30],[31]. poxn flies fed equally on the two colored 5 mM sucrose solutions (PI for blue = 0.49±0.09) (Figure 2). When 40 mM l-canavanine was added to the blue solution, their feeding behavior did not differ (PI for blue = 0.51±0.09, Figure 2). Thus, these results show that l-canavanine is a repellent molecule and that Drosophila uses a chemosensory mechanism to detect this plant insecticide.
To study whether DmXR was involved in l-canavanine detection in vivo, we first used a pharmacological approach. We hindered DmXR function by using the NMA antagonist in the two-choice feeding test. When present in the medium, this drug should diminish the repulsive action of l-canavanine. We first tested whether NMA alone could influence the fly feeding behavior: WT flies were allowed to choose between a blue solution containing 30 mM NMA versus a red control solution, both containing 5 mM sucrose. We found that flies were insensitive to 30 mM NMA (PI for blue = 0.80±0.04 and 0.83±0.04 in control and NMA-fed flies, respectively, Figure 3A). However, significantly more flies fed on 20 mM l-canavanine in presence of 30 mM NMA (PI for blue = 0.68±0.07) than on 20 mM l-canavanine alone (PI for blue = 0.36±0.07) (Figure 3A). l-Arginine (30 mM), which is inactive on DmXR, had no effect on l-canavanine-induced repulsive behavior (PI for blue = 0.41±0.05, Figure 3A). Thus, blockade of DmXR with the antagonist NMA lowered the repellent effect of l-canavanine.
We then used a genetic approach, taking advantage of two fly lines in which the DmXR encoding gene that we called mangetout (mtt) is disrupted. The f06268 line carries a piggyBac transposon inserted into the mtt gene as determined by Exelixis sequence analysis [32], and the Df(2R)Exel7096 line completely removes the mtt locus [33] (Figure 4A). Both mutants are viable in homozygous conditions. We expected that the insertion of a transposon 35 bp downstream from the third exon of mtt would disrupt the transcription of the gene. Indeed, f06268 homozygous flies are mtt loss-of-function mutants since no RNA was detected by quantitative real-time reverse-transcriptase polymerase chain reaction (QRT-PCR) in adults (Figure 4B). In homozygous Df(2R)Exel7096 flies, mtt RNA was also not detectable by QRT-PCR (Figure 4B). As a control for genetic background, we used flies homozygous for the f01266 piggyBac transposon line [32] (Figure 4A), which express normal levels of mtt RNA (Figure 4B). In control two-choice feeding tests without l-canavanine, homozygous mttf06268, hemizygous mttf06268/Df(2R)Exel7096, and homozygous f01266 flies behaved as WT (Figure 3B). When 30 mM l-canavanine was added to the blue solution, mttf06268/mttf06268 and mttf06268/Df(2R)Exel7096 flies fed on this blue solution (PI for blue = 0.71±0.06 and 0.73±0.09, respectively), whereas WT and f01266 flies were repulsed (PI for blue = 0.16±0.06 and 0.23±0.08, respectively) (Figure 3B). Thus, mtt mutant flies are insensitive to 30 mM l-canavanine.
Both pharmacological and genetic data lead to the conclusion that DmXR is required for the detection of l-canavanine and that its activation hinders the feeding. Since DmXR may be the l-canavanine–sensitive receptor in chemosensory organs, we wondered whether this receptor was actually localized in these organs.
Drosophila, like other insects, base their feeding decisions on the presence or absence of specific volatile and nonvolatile chemicals present in the food. Volatile chemicals are in general detected by olfactory neurons, located mainly on the antenna, whereas nonvolatile chemicals like amino acids are detected by gustatory receptor neurons (GRNs). GRNs are present in taste sensilla localized in the legs, the labial palps (or labellum) found on the tip of the proboscis, and within the pharynx (called internal taste organs) [34]. As flies walk on their food sources, tarsal gustatory sensilla evaluate their chemical contents. If phagostimulants are present, the fly extends its proboscis, enabling labellum sensilla to have contact with the food. In the labellum, gustatory chemosensilla house two to four GRNs as well as a single mechanosensory neuron [35]–[37]. In each sensillum, the different subsets of specialized taste neurons are activated by specific classes of tastants, allowing Drosophila to detect sugars, bitter compounds, and water [4],[38].
To investigate whether mtt is expressed in gustatory sensilla, we first performed QRT-PCR experiments on dissected labellum and tarsi/tibiae of WT flies. As shown in Figure S2, mtt RNA is expressed in both organs bearing gustatory sensilla. To assess whether mtt would be present in GRNs, we compared the expression level of mtt RNA in WT to that found in poxn mutants, where chemosensory neurons are transformed into mechanosensory neurons. A significant decrease of the amount of mtt RNA was detected in the labellum and the tarsi/tibiae of poxn mutant compared to WT (Figure S2), suggesting that mtt is expressed in some GRNs. To visualize whether mtt is indeed expressed in gustatory chemosensilla, we then performed in situ hybridization experiments on labellum of WT flies. We found that mtt riboprobe hybridized to a single neuron-like cell within some chemosensory sensilla (Figure 5A–5D). These sensilla, which house five neurons, are clearly gustatory sensilla, because mechanosensory sensilla contain a single, mechanosensory neuron [39]. The in situ labeling appeared to be specific for mtt because chemosensory sensilla were not labeled in labellum from Df(2R)Exel7096 homozygous mtt mutant flies (Figure 5E). Altogether, QRT-PCR and in situ hybridization data indicate that mtt is expressed in only one GRN per labellar chemosensillum, consistent with a role of DmXR as a taste receptor.
Due to its very low level of expression, we were unable to use the in situ hybridization technique combined with immunocytochemistry to further analyze the mtt expression pattern. Hence, in order to investigate the nature of the GRNs expressing mtt, we took advantage of a GAL4 enhancer trap line, NP4288-GAL4, inserted 3.1 Kb upstream from the transcription start site of mtt (Figure 4A). A similar strategy had been undertaken for many other Grs, also reported to have a low level of expression [40]. The expression patterns of these receptors were analyzed using GAL4 transgenes containing taste receptor promoters [29],[40]–[42] or enhancer traps such as NP1017 [43]. These studies have shown that Gr66a-GAL4, Gr5a-GAL4, and NP1017-GAL4 lines drive specific expression in bitter-, sugar-, and water-sensitive GRNs, respectively [29],[42],[43]. When NP4288-GAL4 was crossed with a green fluorescent protein (GFP) reporter line, we observed GFP-positive neurons in taste organs. In the labellum, chemosensory sensilla contained one GFP-positive neuron (Figure 5F–5I), in accordance with the in situ hybridization data. We observed around 28 GFP-positive neurons per labial palp, and two in each tarsus of the legs (Figure 5J and 5K, and Table 1). In addition, four GFP-positive neurons were present in the labral sense organs (LSO) and in the ventral cibarial sense organs (VCSO) (unpublished data), which are bilaterally symmetrical internal taste organs located in the pharynx [44]. Interestingly, Gr66a-GAL4, but not Gr5a-GAL4 or NP1017-GAL4, drives also expression in the LSO and VCSO (Table 1), suggesting that NP4288 is expressed in bitter-sensitive GRNs. To determine whether NP4288-GAL4 and Gr66a-GAL4 are indeed coexpressed, we analyzed transgenic fly lines expressing UAS-nlsGFP under the control of both NP4288 and Gr66a GAL4 drivers, and then counted and compared the number of GFP-positive neurons to that of flies containing each driver alone. In flies that express either the NP4288-GAL4 or Gr66a-GAL4 driver, an average of 28.4 and 26.6 neurons were observed per labial palp, respectively (Figure S3). In flies that express both drivers, an average of 28.8 neurons were detected per palp (Figure S3). Furthermore, we also observed coexpression of both drivers in the LSO and in the foreleg tarsi (table in Figure S3). This indicates that most, if not all, GRNs that express Gr66a also express NP4288, suggesting that this transgene reflects a role for DmXR in bitter-sensitive GRNs. Altogether, these results show that mtt is expressed in one GRN per sensilla that likely correspond to the bitter-sensitive GRNs.
To investigate whether GRNs are sensitive to l-canavanine, we examined a direct behavioral measure of leg GRN stimulation by the proboscis extension reflex (PER) paradigm: when the leg tarsi encounter an attractive sugar solution, the proboscis often extends [42],[43],[45]. If a toxic or bitter compound is added to the sugar solution, the PER is inhibited [42],[43],[45]. This assay enables the application of the drugs only on the legs, which carry solely taste sensilla. In addition, we took care that the proboscis never touched the drugs when it extended, so that we were sure that there was no ingestion of drugs (and consequently, no central effect of these drugs). Using the classical PER paradigm (5-s stimulation by touching the leg tarsi either with a 100 mM sucrose solution or with a 100 mM sucrose+40 mM l-canavanine solution), we found that the occurrence of PER was not affected by l-canavanine in WT or mtt mutant flies (Figure 6). However, after the PER, WT flies usually sustain their proboscis extension to search for food when their legs are maintained in contact with the sugar solution as shown in Video S1. This sustained phase was strongly affected by l-canavanine, since significantly more flies prematurely retracted their proboscis (78±5% of proboscis retraction [PR]) compared to 25±4% of PR for the sucrose solution, Figure 6). This PR phenotype occurred generally just after the proboscis extension (Video S1). We then tested whether the l-canavanine–induced PR phenotype requires DmXR and found that this phenotype disappeared in the mtt loss-of-function mutants (around 25% of PR, Figure 6). Altogether, these data indicate that DmXR is required in leg GRNs for the l-canavanine detection.
To determine in which GRNs mtt is required, we established transgenic flies carrying a mtt RNA interference (RNAi) construct under the control of UAS sequence [46]. We first expressed mtt RNAi with NP4288-GAL4 in heterozygous mttf06268 flies and tested the effect of l-canavanine by the PER/PR behavioral assay. The occurrence of PER was not affected by l-canavanine in controls and mtt-knockdown flies (Figure S4A). However, RNAi knockdown of mtt suppressed the l-canavanine–induced premature PR phenotype (Figure 7A) in a comparable manner to that observed in mtt mutant flies. This indicates that mtt is expressed in NP4288-GAL4–positive cells, which overlap Gr66a-GRNs in taste organs. However, NP4288-GAl4 drives also expression in cells during development and in the adult brain (unpublished data), precluding us from concluding that mtt is required only in GR66a-GRNs for l-canavanine detection. Thus, we next used the Gr66a-GAL4 driver to specifically express the mtt RNAi in Gr66a-GRNs of heterozygous mttf06268 flies. As already observed with NP4288-GAL4–driven knockdown of mtt, the occurrence of PER was unmodified in this genetic condition (Figure S4A). Importantly, Gr66a-GAL4–induced knockdown of mtt significantly reduced the occurrence of l-canavanine–induced premature PR (Figure 7A). This clearly indicates a requirement of mtt in Gr66a-GRNs for l-canavanine sensitivity.
To clearly demonstrate that l-canavanine detection required the presence of DmXR only in Gr66a GRNs, we performed rescue experiments by targeting mtt expression in distinct types of GRNs of mtt homozygous mutants by using the GAL4/UAS system [46]. Indeed, several types of GRNs are present in the tarsi, such as sucrose-, bitter-, and water-sensitive GRNs (Figure S5). Expression of mtt in the different subsets of GRNs did not affect the percentage of PER (Figure S4). We then analyzed the PR and found that expression of mtt in sugar or water GRNs did not rescue the mutant phenotype (Figure 7B). In contrast, expression of mtt in Gr66a-GRNs rescued l-canavanine sensitivity (Figure 7B). Thus, it is the stimulation of DmXR, in Gr66a-GRNs, which is responsible for l-canavanine–induced PR. To further verify that only Gr66a-GRNs are necessary for l-canavanine sensitivity, we expressed the hid and rpr proapoptotic genes [47] in these neurons, or inhibited their neurotransmitter release with the tetanus toxin transgene [48]. In both cases, the PER was not affected (Figure S6), but the l-canavanine–induced PR was lost (Figure S6). Altogether, these results demonstrate that DmXR is expressed and required only in Gr66a-GRNs for l-canavanine detection.
The sites of taste reception are localized to the dendrites of GRNs [36], which are bipolar neurons containing a single dendrite and a single axon. To confirm that DmXR was actually the l-canavanine taste receptor and not a regulatory receptor modulating Gr66a-GRN synaptic transmission, we performed two kinds of experiments. First, we expressed a HA-tagged receptor in Gr66a-GRNs to determine its subcellular localization in the labellum. As shown in Figure 8A, the receptor was highly concentrated at the dendrite and not detected in the GRN axon in accordance with a gustatory function. Second, we tested whether DmXR could modify Gr66a activation by other repellents than l-canavanine. It has been shown that Gr66a-GRNs are required for caffeine-aversive behavior, Gr66a being a gustatory receptor for caffeine [8]. As already published, we could find that caffeine inhibited sucrose-induced PER (Figure 8B). However, when the PER occurred, we noticed that there was a high rate of premature PR. The caffeine-induced PR phenotype occurred generally just after the PER, similar to what was observed with l-canavanine (Video S2). We then tested the caffeine-induced phenotypes in mtt mutants and did not find changes in caffeine-induced PER inhibition and caffeine-induced PR (Figure 8B). These data clearly demonstrate that DmXR acts as a l-canavanine gustatory receptor in Gr66a-GRNs.
The ability to avoid ingestion of toxic plants compounds is crucial for insect survival. However, before the current study, only two receptors, Gr66a and Gr93a, which are essential for the caffeine response, were associated with a specific bitter tastant [7],[8]. The nonprotein amino acid l-canavanine is known to be toxic to insects, when ingested ([18] and this study). Here, our results show that l-canavanine is detected as a repulsive molecule. With a pharmacogenetic approach, we have shown that Drosophila uses a taste detection mechanism mediated by the orphan GPCR, DmXR, which is activated by l-canavanine to trigger this avoidance behavior. This process occurs in bitter-sensitive GRNs where this receptor is expressed.
By using the two-choice feeding test, the repulsive effect of l-canavanine was clearly demonstrated. Contrary to the known repellents (caffeine, quinine [8],[49]), l-canavanine does not affect the PER. However, l-canavanine triggers the retraction of the proboscis following its initial extension impairing the food intake. Indeed, after the PER, WT flies usually sustain their proboscis extension to search for food when their legs are maintained in contact with the sugar solution. This sustained phase is strongly affected by l-canavanine, since significantly more flies prematurely retracted their proboscis at this stage. This inhibition of sustained proboscis extension is not specific for l-canavanine, but is also observed, with the same rate, in the presence of caffeine. Hence, caffeine induces a fast response, which is the PER inhibition, and a slow response, which is the PR, whereas l-canavanine only induces the slow response. One open question is to understand the molecular mechanisms responsible for such a difference between caffeine and l-canavanine in the response dynamics, knowing that both drugs act on the same cell type (Gr66a-GRNs). GPCR-induced signal transduction pathways rely on the activation of intracellular heterotrimeric G-proteins [50]–[52]. To date, two G-proteins have been implicated in the taste pathway, and both are required for sugar perception [53],[54]. However, a direct evidence for a coupling between Grs and G-proteins has not been demonstrated [38]. As we have shown in our cell transfection assay ([27] and this study), DmXR is a genuine GPCR. It is thought that the Gr66a receptor, like other Grs, is a putative GPCR [8],[38]. However, we do not know to which type of G-protein these two receptors are coupled in vivo. A first explanation about the different dynamics induced by l-canavanine and caffeine could be that that DmXR and Gr66a are coupled to distinct G-proteins or that both receptors are coupled to the same G-protein, but with different efficiencies. A more speculative explanation may be due to the different structural features of the two receptors, taking into account that Grs are structurally related to the olfactory receptors in Drosophila [40]. As was recently shown, Drosophila olfactory receptors may act as ligand-gated channels instead of being coupled to a G-protein [55],[56]. Thus, Gr66a receptor may also be a ligand-gated channel. Because of the absence of any intermediate, changes in membrane excitability would be more rapid in presence of caffeine compared to l-canavanine. This may explain the difference in the response dynamics between these two drugs.
This study shows that DmXR is expressed and required in bitter-sensitive leg GRNs. However, DmXR is also known to be expressed in the adult brain, in agreement with our observations that NP4288-GAL4 is expressed in this tissue (unpublished data). This suggests the existence of an unknown endogenous ligand, different from l-canavanine, triggering DmXR activation in the brain. To exclude any action of l-canavanine in the brain, we took care that flies avoided ingesting the drug solutions by applying them only on legs during the PER analysis. In addition, we used GRN-restricted drivers, allowing us to specifically analyze the peripheral function of DmXR. Finally, the absence of any defects in the caffeine-induced response of mtt mutants flies excludes a role of DmXR in second and higher order neurons involved in the control of the studied gustatory behavior. As we observed mtt expression in the labellum and in internal taste organs (LSO and VCSO), it is very likely that these taste sensilla also play a role in the l-canavanine–induced aversive behavior. In agreement with this, we observed that flies did not drink a l-canavanine/sucrose–containing solution when directly applied on the labellum (unpublished data), confirming the presence of l-canavanine–sensitive GRNs. So, we assume that DmXR is a l-canavanine–tuned gustatory receptor in all these taste organs.
Surprisingly, it is not a Gr member that has been selected to detect l-canavanine, despite the very high sequence diversity of this family. Indeed, DmXR belongs to the mGluR GPCR subfamily because of its close sequence relationship [27]. DmXR and mGluR LBP sequences and 3-D model comparisons have shown that DmXR has diverged only in the LBP part interacting with the γ-carboxylic group of glutamate [27]. These modifications have targeted and changed two residues that are conserved in all mGluRs and are crucial for glutamate-induced activation [24],[27]. Our study shows that these structural changes are correlated with the ligand selectivity of the receptor. Indeed, DmXR has a divergent LBP so that glutamate is no more an agonist but l-canavanine is. Conversely, the Drosophila mGlu ortholog receptor, DmGluA, is not activated by l-canavanine. This suggests that the original conformation of the mGluR LBP was more adapted to diverge and to recognize l-canavanine than the one of the Grs. In addition, to give rise to this new type of gustatory receptor, appropriate GRN expression has also been added during the evolution of the DmXR function since mGluR expression is mainly found in the central nervous system [25]. Thus, our study suggests that other GPCRs, different from Grs, may have evolved in insects to recognize specific tastants.
Finally, the insect-borne diseases are largely increasing, partly due to the development of insecticide resistance. Thus, it becomes urgent to identify insect-specific targets for the design of new drugs against insects. Our work illustrates that the pharmacological and functional characterization of the insect-specific GPCRs, which likely control insect-specific physiological processes, is a way to discover new protection or fighting strategies against harmful insects.
l-glutamate, l-canavanine, l-arginine, and γ-N-methyl-l-arginine (NMA) were from Sigma. Human embryonic kidney (HEK 293) cells were cultured and transiently transfected by electroporation as previously described [27]. Carrier plasmid DNA (pRK5) (14 μg), plasmid DNA containing HA-DmXR WT, HA-DmXRT176A mutant (4 µg), DmGluRA (2 µg), and plasmid DNA containing Gαqi9 (2 µg) (to enable the artificial coupling of DmXR and DmGluRA to phospholipase C, [57]) were used for the transfection of 107 cells. Determination of inositol phosphate (IP) accumulation in transfected cells was performed as previously described [27].
Drosophila stocks were raised on standard fly food medium at 25°C on a 12-h light/dark cycle. WT Canton S flies were used as control flies in all behavioral assays. For experiments using pox-neuro (poxn) adult mutant flies, homozygous flies carrying the poxn70−23 mutant allele were used [30]. mttf06268 and f01266 lines carry PBac(WH) transposon and are described in [32]. The Df(2R)Exel7096 line carries a small deficiency that completely removes the mtt locus and some adjacent genes (CG8697 to CG2397) [33]. The p(UAS-mtt) transgene construct was generated by cloning the hemagglutinin N-terminally tagged full coding sequence of DmXR (HA-mtt) [27] into the pUAST transformation vector and injected into w1118 embryos. Several insertions lines were obtained. After QRT-PCR analysis (unpublished data), the w1118;UAS-mttC5 line was chosen for this study. The mtt RNAi line was obtained after amplifying DmXR cDNA sequence with the sense primer 5′-ACT ACT TCT AGA GGC GAT GTG GCA ACA G-3′ and the antisense primer 5′-CCG GGC TCT AGA ATA AGT TTG TTT GCA G-3′. This sequence was digested with the XbaI restriction enzyme and subcloned into AvrII-digested pWIZ. This new construct was then digested with the NheI restriction enzyme and ligated with the same XbaI-digested PCR product. A clone with the second insert oriented opposite to the first was then selected and used for injection of w1118 embryos. Several insertion lines were obtained. After QRT-PCR analysis (unpublished data), the w1118;UAS-mtt RNAi1 line was chosen for this study. The Gr66a-GAL4 (II chromosome) and Gr5a-GAL4 (II chromosome) promoter GAL4 lines are generous gifts from H. Amrein (Duke University, United States). The NP1017-GAL4 (X chromosome) enhancer trap line was kindly provided by T. Tanimura (Kyushu University, Japan). The NP4288 enhancer trap was obtained from the GETDB Stock Center (Kyoto, Japan) [58]. The UAS-hid:UAS-rpr line was a gift from J. R. Martin (Paris Sud University, France). The UAS-TeT line was kindly provided by C. J. O'Kane (Cambridge University, England). The w;UAS-mCD8-GFP and the w;UAS-nlsGFP were obtained from the Bloomington Stock Center.
Total RNAs were extracted from whole adult flies (for the analysis of mtt mutants) or dissected labella, tarsi, and tibiae (for the analysis of mtt expression) by using Trizol (Sigma). cDNAs were generated from 1 µg of total RNAs treated with DNase I (Ambion) by using random decamers (Ambion) and Moloney murine leukemia virus reverse transcriptase (Invitrogen). Real-time PCR was done using Applied Biosystems SYBR Green PCR mix according to the manufacturer's instructions. PCR was done as follows: 10 min at 95°C followed by 40 cycles: 15 s at 95°C, 60 s at 60°C. Housekeeping genes used to normalize DmXR expression levels were RpL13, Tbp, and Pgk. Sequences of the primers are RpL13 5′-AGGAGGCGCAAGAACAAATC and 5′-CTTGCTGCGGTACTCCTTGAG, Tbp 5′-CGTCGCTCCGCCAATTC and 5′-TTCTTCGCCTGCACTTCCA, Pgk 5′-TCCTGAAGGTCCTCAACAACATG and 5′-TCCACCAGTTTCTCGACGATCT, and DmXR 5′-CGAATGCAACTGGTTCCTTCTC and 5′-TGAGGAAGTACTCCTCGAAC.
Labella were dissected from flies and collected in 4% paraformaldehyde in PBS with 0.05% Triton X-100 on ice. After fixation overnight at 4°C, samples were washed 6×10 min in PTX (PBS, 2% Triton X-100) at room temperature. Prehybridization was then done for 2 h at 55°C in hybridization buffer (HB) (50% formamide, 5× SSC, 0.5 mg/ml yeast tRNA, 0.1 mg/ml Salmon Sperm DNA, 0.05 mg/ml heparin, 0.3% Triton X-100). Hybridization was performed overnight at 55°C with digoxigenin-labeled antisense mtt riboprobe derived from mtt cDNA and prepared according to the manufacturer's instructions (Roche). Washes were performed at 58°C in HB followed by washes in HB/PTX mix (3/1, 1/1, and 1/3, respectively). After blocking in 0.5% Blocking Reagent (Roche) in PTX, samples were then incubated with anti-Dig-AP (Roche) overnight at 4°C. Samples were then washed 6×10 min in PTX. NBT/BCIP mix (Roche) was used to visualize the digoxigenin-labeled probe. Samples were mounted in 90% glycerol.
To visualize HA-DmXR protein expression, we dissected the labella from Gr66a-GAL4;UAS-HAmtt/+flies from adult head, fixed them overnight in 4% paraformaldehyde in 1× phosphate-buffered saline (PBS), 0.3% Triton X-100. Immunostaining was performed in 1× PBS 3% Triton X-100 and 0.5% Blocking Reagent (Roche). The following antibodies were used: monoclonal rat anti-HA (Roche; 1∶200) and Cy3-conjugated donkey anti-rat (Jackson ImmunoResearch; 1∶500). Samples were mounted in Vectashield. Images were acquired using a Leica microscope and CoolSNAP camera.
|
10.1371/journal.pbio.1000599 | Fatty Acid Desaturation Links Germ Cell Loss to Longevity Through NHR-80/HNF4 in C. elegans | Preventing germline stem cell proliferation extends lifespan in nematodes and flies. So far, studies on germline-longevity signaling have focused on daf-16/FOXO and daf-12/VDR. Here, we report on NHR-80/HNF4, a nuclear receptor that specifically mediates longevity induced by depletion of the germ line through a mechanism that implicates fatty acid monodesaturation.
nhr-80/HNF4 is induced in animals lacking a germ line and is specifically required for their extended longevity. Overexpressing nhr-80/HNF4 increases the lifespan of germline-less animals. This lifespan extension can occur in the absence of daf-16/FOXO but requires the presence of the nuclear receptor DAF-12/VDR. We show that the fatty acid desaturase, FAT-6/SCD1, is a key target of NHR-80/HNF4 and promotes germline-longevity by desaturating stearic acid to oleic acid (OA). We find that NHR-80/HNF4 and OA must work in concert to promote longevity.
Taken together, our data indicate that the NHR-80 pathway participates in the mechanism of longevity extension through depletion of the germ line. We identify fat-6 and OA as essential downstream elements although other targets must also be present. Thus, NHR-80 links fatty acid desaturation to lifespan extension through germline ablation in a daf-16/FOXO independent manner.
| Reproduction and aging are two processes that seem to be closely intertwined. Experiments in Caenorhabditis elegans and Drosophila have shown that depletion of the germ line increases lifespan and that this process depends on insulin and lipophilic-hormone signaling. Recently, it was demonstrated that when germline stem cells (GSCs) cease to proliferate, fat metabolism is altered and this affects longevity. In this study, we have identified a nuclear hormone receptor, NHR-80, that mediates longevity through depletion of the germ line by promoting fatty acid desaturation. The nhr-80 gene is up-regulated at the mRNA and protein levels in germline-less animals, leading to the transcription of the gene, fat-6, and the production of oleic acid (OA). Our experiments also show that the NHR-80/FAT-6/OA pathway does not require the presence of DAF-16 but instead, depends fully on the presence of DAF-12, a steroid receptor that affects lifespan. We provide evidence that other NHR-80 targets must be present concomitantly. Our results reinforce the notion that fat metabolism is profoundly altered in response to GSC proliferation, and the data contribute to a better understanding of the molecular relationship between reproduction, fat metabolism, and aging.
| Removing the germ line of Caenorhabditis elegans extends its lifespan by approximately 60% [1]. Eliminating germ cells also increases the lifespan of Drosophila, suggesting that a conserved mechanism links the germ line to longevity [2]. In C. elegans, removal of the germ line can be achieved either by laser ablation of germline precursor cells at early developmental stages or through mutations that impair the proliferation of germline stem cells (GSCs) [3]. The glp-1(e2141ts) and mes-1(bn7) alleles deplete the germ line by either blocking proliferative signals for GSC or inhibiting cell division in the P lineage at early embryonic stages. As a result, animals carrying these alleles are long-lived [3]. Longevity is not merely caused by sterility because animals lacking both germ cells and the somatic gonad are sterile but not long lived. The germ line and the somatic gonad have been suggested to have opposite effects on longevity [1], but the molecular basis of germline-mediated longevity remains poorly understood.
Hsin and Kenyon showed that the presence of daf-16/FOXO or daf-12/VDR is required for extending the lifespan of animals whose germ line had been ablated [1]. DAF-16/FOXO is a forkhead transcription factor that translocates into intestinal nuclei and promotes transcription when GSCs stop proliferating [4]. daf-16/FOXO is a key downstream component of the insulin/IGF1 signaling (IIS) pathway that also regulates longevity [5],[6]. Although inhibiting GSC proliferation and down-regulating the activity of the IIS pathway both result in lifespan extension and translocation of DAF-16 into intestinal nuclei, several experiments show that these manipulations are not equivalent. First, GSC removal extends the lifespan of daf-2 mutants that are already long-lived due to a constitutively down-regulated IIS pathway (hypomorphic allele of the sole insulin receptor) [1]. Nuclear translocation of DAF-16 requires the intestinal protein KRI-1 (an ankyrin-repeat protein) when it is provoked by a GSC proliferation arrest, but not by daf-2 mutations [4]. Finally, the transcription elongation factor, TCER-1, promotes the transcriptional activity of daf-16/FOXO when GSCs stop dividing but not in long-lived IIS mutants [7]. Taken together, these data indicate that IIS and GSCs affect longevity though distinct mechanisms although they are both mediated by DAF-16.
The second pathway required for germline longevity is the DAF-12 lipophilic-hormone signaling pathway. In response to the loss of germ cells, the cytochrome P450, DAF-9 [8], and the Rieske protein, DAF-36 [9], use cholesterol to produce a steroid hormone (dafachronic acids) that activates the nuclear hormone receptor, DAF-12/VDR [10]. DAF-12/VDR is homologous to the vertebrate vitamin D receptor, and its presence in its activated form is required to extend lifespan though depletion of the germ line [1]. The interactions of KRI-1/DAF-16/TCER-1 and the DAF-9/DAF-36/DAF-12 pathways are still unclear.
Similar to KRI-1, DAF-9 and DAF-12 facilitate the nuclear translocation of DAF-16 triggered by germline removal [4], suggesting that the lipophilic-hormone signaling pathway may act upstream of DAF-16. However, recent work showed that DAF-12 and DAF-16 also function separately. First, germline-less animals in which DAF-16 is forced into intestinal nuclei still require daf-12 to be long-lived [4]. Second DAF-12 and DAF-16 promote the expression of different gene sets [11]. The Kenyon lab showed that sod-3 and cdr-6 are DAF-16 and DAF-12 targets, respectively [11]. The K04A8.5 lipase is induced in glp-1(e2141ts) mutant animals in a daf-16 dependent manner but is not affected by daf-12 [12]. Since the K04A8.5 lipase is also required for lifespan extension, these results suggest that the KRI-1/DAF-16/K04A8.5 pathway can promote longevity independently of daf-12 and that DAF-16 dependent transcription does not strictly require daf-12. Finally, it has also been shown that the DAF-12 lipophilic-hormone signaling pathway can mediate longevity in response to the somatic reproductive tissues, but this also requires the presence of DAF-16 [11]. Thus, it is still unclear whether DAF-12 can promote longevity in the absence of daf-16.
In the present study, we searched for new nuclear receptors that can mediate longevity of C. elegans through depletion of the germ line using an RNAi-based genetic screen. We report that nhr-80/HNF4 is required for extending lifespan through germline removal, although it does not affect the lifespan of wild type animals. We show that NHR-80 is specific to this pathway since other longevity paradigms are not affected by a loss-of-function mutation of nhr-80. Moreover, the levels of NHR-80 increase in intestinal cells when germ cells are depleted. This increase is physiologically relevant because (1) overexpressing nhr-80 further extends the lifespan of germline-less animals and (2) germline ablation leads to the nhr-80 dependent up-regulation of the stearoyl-CoA desaturase (SCD), fat-6, that produces oleic acid (OA) from stearic acid and (3) increased fatty acid desaturation and OA production are necessary to extend the lifespan of germline-less animals. A link between fat metabolism and germline-mediated longevity has already been reported by the Ruvkun lab. In this recent report, the authors reported that the triacylglyceride (TAG) lipase (K04A8.5) is required for germline longevity [12]. Both fat-6 and K04A8.5 are induced in germline-less animals, and their inactivation by RNAi fully suppresses lifespan extension by depletion of the germ line. However, in contrast to the K04A8.5 lipase that acts downstream of the KRI-1/DAF-16 pathway and independently of the DAF-9/DAF-36/DAF-12 lipophilic-hormone signaling pathway, we show that the NHR-80/FAT-6/OA pathway does not require the presence of daf-16 but necessitates the presence of daf-12. Taken together, our data and that of Wang et al. are consistent with the conclusion that the lifespan benefits triggered by inhibiting GSC proliferation require a important modification of the metabolism of fat since the NHR-80/FAT-6/OA and the KRI-1/DAF-16/K04A8.4 pathways are activated independently to promote longevity though the activation of fat remodeling enzymes.
To identify new genes required for lifespan extension triggered by germline ablation, we screened for genes encoding for nuclear receptors (NHRs) using RNAi by feeding. We sought genes whose inactivation could suppress the lifespan of glp-1(e2141ts) mutant C. elegans without affecting the lifespan of wild type animals. Therefore, we compared the proportion of dead animals in glp-1(e2141ts) mutant and wild type animals after 20 d of RNAi treatment (starting at day 1 of adulthood). At 20 d, 50% to 60% of glp-1(e2141ts) mutants were alive, compared to less than 30% of wild type animals. In our screen, successful candidates lowered the survival of glp-1(e2141ts) mutants significantly but did not affect that of wild type animals at 20 d of adulthood. Of the 195 NHRs present in the Ahringer library (70% of all NHRs present in C. elegans), only one, nhr-80, in addition to our positive control (daf-12), reduced the lifespan of glp-1(e2141ts) mutants without affecting wild type lifespan. In wild type fertile animals, NHR-80 promotes fatty acid (FA) desaturation without affecting lifespan [13].
C. elegans carrying the nhr-80(tm1011) allele behave similarly to wild type animals subjected to nhr-80 RNAi, suggesting that it is a loss-of-function mutation [13]. Although the nhr-80(tm1011) allele does not affect the lifespan of wild type animals (Figure 1A, Table S1), it fully suppresses that of glp-1(e2141ts) mutants (Figure 1A, Table S1), without restoring germline development (unpublished data). To ensure that down-regulation of nhr-80 did not merely suppress the glp-1(e2141ts) allele, we examined whether nhr-80 RNAi could also suppress another surrogate for germline-mediated longevity, mes-1(bn7) mutants. The lifespan extension observed in sterile mes-1(bn7) mutants is suppressed by nhr-80 RNAi (Figure 1B, Table S1). This suggests that NHR-80 is required for extending longevity through depletion of the germ line.
To determine whether nhr-80 is specific to germline-mediated longevity, we knocked out nhr-80 in several other longevity paradigms [5],[14],[15]. For example, daf-2(e1370) mutant animals are long-lived due to a reduction-of-function mutation in the insulin receptor and are not affected by nhr-80 RNAi treatment (Figure 2A, Table S1). Similarly, dietary restriction (DR) and cyc-1 RNAi, which reduces mitochondrial function [14], extend longevity to a comparable extent in nhr-80(tm1011) mutants and in wild type animals, suggesting that these pathways are not affected by NHR-80 (Figure 2C and D, Table S1). We conclude that NHR-80 specifically promotes germline longevity.
We analyzed the localization of NHR-80, using a functional protein fused to a GFP tag (see Materials and Methods and Figure 3B). In contrast to previous reports [16], we found that NHR-80 is localized in the nucleus and that it is expressed in the intestine and in neurons (some head and tail neurons, as well as the ventral cord; Figure 3A and B). This discrepancy is likely due to the fact that we fused a GFP tag to the full-length NHR-80 sequence driven by its own promoter, while Miyabayashi et al. fused the tag to the nhr-80 promoter and may have missed the nuclear import signal [16]. We found that NHR-80 nuclear localization is constitutive and independent of the presence of the germ line (i.e., glp-1(e2141ts);nhr-80(tm1011);NHR-80::GFP mutant animals at restrictive versus permissive temperature) (Figure 3A and B). However, the intensity of the NHR-80::GFP in intestinal cells is increased by 60% when glp-1(e2141ts);nhr-80(tm1011) mutant animals are shifted to the restrictive temperature at the L1 stage while no changes are noted in neuronal cells (Figure 3C and D). When glp-1(e2141ts) mutant animals were shifted to restrictive temperature at later stages (L4 and day 1 of adulthood), the induction of intestinal NHR-80 still occurs, but less dramatically (L4 and day 1 of adulthood; Figure S1). Taken together with previous data showing that late shifts extend lifespan to a lesser extent [3], our data suggest that NHR-80 induction correlates with the extent to which lifespan is extended. To confirm this increase, we measured the overall expression level of nhr-80 by qRT-PCR and found that it is induced 5.6-fold in glp-1(e2141ts) mutant animals (Figure 3E). Thus, in glp-1(e2141ts) mutant animals, overall nhr-80 mRNA levels and the intensity of NHR-80::GFP in the intestine are increased. This strongly suggests that NHR-80 promotes longevity in the intestine. Supporting this notion, we found that nhr-80 RNAi also suppresses the lifespan of glp-1(e2141ts) mutant animals, although neurons are refractory to RNAi (Figure S2, Table S1).
Because nhr-80 is a positive longevity regulator that is induced in glp-1(e2141ts) mutant animals, we examined whether overexpressing nhr-80 could recapitulate germline-mediated longevity in a wild type context. Surprisingly, the nhr-80 transgene, which fully restores the longevity of glp-1(e2141ts);nhr-80(tm1011) mutant animals (Figure S3, Table S1), fails to extend the lifespan of wild type animals (Figure 4A, Table S1) but increases the mean lifespan of glp-1(e2141ts) mutant animals by 80% (Figure 4B, Table S1). It is remarkable that a loss-of-function mutation and an overexpression of nhr-80 have opposite effects on lifespan in the absence of proliferating GSCs only (Figures 1A and 4B, Table S1). The mechanism through which this is achieved remains to be determined. Our data do not allow discrimination between activation by binding of NHR-80 to a ligand, post-translational modifications, or interaction with a partner. To gain further insights into NHR-80 functions, we examined the interaction of nhr-80 with other longevity determinants of germline-less animals.
To examine mechanisms through which NHR-80 may promote the longevity of germline-depleted animals, we first tested whether nhr-80 longevity function depends on DAF-16. To address this question, we assessed whether NHR-80 may promote DAF-16 nuclear localization, similar to KRI-1 (orthologous to the human disease gene KRIT1; [4]). We found that down-regulating nhr-80 by injection of double-stranded RNA does not affect the localization of DAF-16 (Figure 5A). Next, we tested whether daf-16 was required for the increase of nhr-80 mRNA levels triggered by the depletion of the germ line or the longevity extension provoked by nhr-80 overexpression. We found that, in daf-16(mu86);glp-1(e2141ts) mutant animals, the nhr-80 mRNA levels are increased relative to wild type (Figure 5B; 3.7-fold; p = 0.002) and that overexpressing nhr-80 increases lifespan by 40% (Figure 5C, Table S1), while nhr-80 RNAi decreases lifespan by 58% (Figure S4, Table S1). However, the transcriptional induction of nhr-80 and the lifespan extension triggered by nhr-80 overexpression are decreased in the absence of daf-16 relative to glp-1(e2141ts) mutants, suggesting that DAF-16 can modulate the transcriptional induction of nhr-80. Thus, DAF-16 is not strictly required for nhr-80 function, but DAF-16 can modulate nhr-80 mRNA levels.
We observed that nhr-80 overexpression fails to extend the lifespan of glp-1(e2141ts);daf-12(rh61rh411) mutants (Figure 5D, Table S1) but extends the lifespan of glp-1(e2141ts);daf-9(rh50) mutants (Figure 5E, Table S1). Thus, DAF-12 is required for NHR-80 mediated longevity, but not DAF-9. This suggests that NHR-80 functions independently of the DAF-9 derived ligand, Δ7-dafachronic acid [10]. To confirm this idea, we tested whether nhr-80 overexpression could also extend lifespan of glp-1(e2141ts);daf-9(rh50) mutants in the presence of Δ7-dafachronic acid. We found that this was the case: in the presence of Δ7-dafachronic acid, nhr-80 overexpression extends the lifespan of glp-1(e2141ts);daf-9(rh50) double mutants to the same extent (Figure S5, Table S1). These data confirm that NHR-80 genetically interacts with DAF-12 in a daf-9 and Δ7-dafachronic acid independent manner.
To further explore how daf-12 and nhr-80 may interact, we measured the nhr-80 mRNA levels in glp-1(e2141ts);daf-12(rh61rh411) mutant animals and found that nhr-80 mRNA levels are induced in these animals relative to wild type (Figure 5F), suggesting that DAF-12 is not strictly required for the nhr-80 transcriptional induction. However, similar to DAF-16, DAF-12 can modulate nhr-80 mRNA levels (Figure 5F). This may be explained by the presence of two distant DAF-12 binding sites on the NHR-80 promoter (Figure S6). We also measured the daf-12 mRNA levels to test whether DAF-12 could be a target of NHR-80 and found that they are not affected in a glp-1(e2141ts);nhr-80(tm1011) context (Figure 5G). Taken together, our data indicate that nhr-80 and daf-12 are not exclusive transcriptional targets of one another. The simplest way to explain this interaction is that DAF-12 and NHR-80 function in concert to promote longevity. The finding that nhr-80 is functional in glp-1(e2141ts);daf-9(rh50) mutants shows that DAF-12 can promote longevity independently of daf-9 in a germline-less animals.
Next, we examined the role of NHR-80 targets. Known transcriptional targets of NHR-80 in a wild type context include three Δ9-desaturases involved in lipid metabolism: fat-5, fat-6, and fat-7 [13]. fat-5 is a palmitoyl-CoA- Δ9-desaturase (PCD) that converts palmitic acid to palmitoleic acid while fat-6 and fat-7 are stearoyl-CoA- Δ9-desaturases (SCD) that convert stearic acid to OA [13]. We measured mRNA levels of these three targets and of the lipase K04A8.5, another enzyme involved in lipid metabolism and germline-mediated longevity [12]. We found that, in glp-1(e2141ts) mutants, fat-5, fat-6, and K04A8.5 are strongly induced, while fat-7 is repressed (Figure 6A, 6B, 6C, and 6D). In glp-1(e2141ts);nhr-80(tm1011) mutants, the induction of fat-6 is abolished (p≤0.001), fat-5 is suppressed to a lesser extent (p = 0.002) and K04A8.5 and fat-7 are unaffected (Figure 6A, 6B, 6C, and 6D). Because lipid desaturases have been shown to be transcriptional targets of DAF-16 in daf-2 mutants [17], we also measured the mRNA levels of fat-5, fat-6, fat-7, and K04A8.5 in daf-16(mu86);glp-1(e2141ts) mutant animals. In this background, the induction of fat-5 and K04A8.5 no longer occurs (p = 0.005; Figure 6A and 6D), but that of fat-6 or fat-7 was not affected (p = 0.3; Figure 6B and 6C). Thus, our data indicate that fat-6 transcriptional up-regulation depends on NHR-80, while that of fat-5 depends on both NHR-80 and DAF-16. As previously reported, K04A8.5 is a DAF-16 target [12], but not of NHR-80 (Figure 6B). This suggests that, in the glp-1(e2141ts) context, fat-6 is a NHR-80 target. To verify whether the nhr-80 dependent up-regulation of fat-6 results in an increased production of OA, we measured the OA concentration in fertile and germline-less animals. We found that the levels of OA as well as the stearic/oleic acid ratio are specifically increased in germline-less animals (Figure S7). This confirms the up-regulation of the FAT-6/OA pathway in glp-1(e2141ts) mutants.
To investigate the relevance of each Δ9-desaturase to glp-1(e2141ts) longevity, we deactivated all of the fat genes individually in a glp-1(e2141ts) background. We found that deletion of the fat genes independently in glp-1(e2141ts) mutant animals does not affect longevity (Figure S8, Table S1). However, using qRT-PCR, we found that, similar to other reports in wild type animals [18], compensatory mechanisms occur between fat-6 and fat-7 in germline-depleted animals and fat-7 mRNA levels are strongly up-regulated in glp-1(e2141ts);fat-6(tm331) mutant animals (p≤0.001; Figure S9). To bypass this compensatory mechanism, we generated the triple mutants glp-1(e2141ts);fat-6(tm331);fat-7(wa36) (SCD activity is fully abolished), glp-1(e2141ts);fat-6(tm331);fat-5(tm420), and glp-1(e2141ts);fat-7(wa36);fat-5(tm420). Concomitant deletion of fat-5 with one of the SCDs (fat-6 or fat-7) either slightly increases or decreases the lifespan of glp-1(e2141ts) mutant animals, while it does not affect the lifespan of wild type animals (27 and 20 d for glp-1(e2141ts);fat-7(wa36);fat-5(tm420) and glp-1(e2141ts);fat-6(tm331);fat-5(tm420) mutants, respectively, versus 26 d for glp-1(e2141ts) mutants; Figure S10, Table S1). In contrast, deleting two SCDs (fat-6 and fat-7) sharply decreases the longevity of glp-1(e2141ts) mutant animals without affecting that of wild type animals (MLS = 14 d for glp-1(e2141ts);fat-6(tm331);fat-7(wa36) mutants; Figure 7A and B, Table S1). Importantly, the addition of OA, the product of the reaction catalyzed by FAT-6/FAT-7, during adulthood in glp-1(e2141ts);fat-6(tm331);fat-7(wa36) mutants restores the lifespan of these animals (MLS = 26 d for supplemented mutants and 25 d for glp-1(e2141ts) mutants), while it does not affect the lifespan of either fat-6(tm331);fat-7(wa36) mutant or wild type animals (Figure 7A and B, Table S1). This suggests a key role for OA or one of its metabolites in linking germ cell loss to longevity. Next, we showed that blocking further processing of OA by desaturation to poly-unsaturated fatty acids using fat-2 RNAi does not affect the lifespan of glp-1(e2141ts) mutants (Figure S11A, Table S1). Finally, we note that OA fails to extend the lifespan of glp-1(e2141ts) mutant animals. This is surprising because overexpressing nhr-80 significantly increases the lifespan of these animals (Figure 7A, Table S1). Our data therefore indicate that NHR-80 promotes longevity through both the FAT-6/OA branch and other critical targets. We conclude that SCD activity and OA itself is required, but not sufficient, for germline-mediated longevity.
The finding that OA is produced in response to the NHR-80/FAT-6 pathway provides an additional possibility to test the interaction of the FAT-6/OA branch with the other main germline longevity pathways: the KRI-1/DAF-16/K04A8.5 and the DAF-9/DAF-12 lipophilic-hormone pathways. Since our results demonstrated that KRI-1/DAF-16/K04A8.5 acts independently from the NHR-80/FAT-6 pathway, we verified whether OA could extend the lifespan of daf-16(mu86);glp-1(e2141ts) mutant animals and found that it fails to do so (Figure S12, Table S1). This confirms that addition of exogenous OA is not equivalent to nhr-80 overexpression, which extends the lifespan of the daf-16(mu86);glp-1(e2141ts) double mutants (Figure 5C, Table S1) and indicates that FAT-6 is not the only NHR-80 target that promotes longevity. It is consistent with the idea that OA is already highly produced in glp-1(e2141ts) and daf-16(mu86);glp-1(e2141ts) mutants because the NHR-80/FAT-6/OA is already effective.
Next, we tested the effect of OA on the lifespan of glp-1(e2141ts);daf-12(rh61rh411) and glp-1(e2141ts);daf-9(rh50) mutants and found that it fails to extend lifespan in these two backgrounds (Figure S12, Table S1). Furthermore, we found that fat-6 mRNA levels are still induced in glp-1(e2141ts);daf-12(rh61rh411) mutant relative to wild type animals (Figure S13). Taken together, our data exclude the possibility that the FAT-6/OA branch is a common target of DAF-12 and NHR-80. Thus, we conclude that DAF-12 does not interact with NHR-80 by co-promoting FAT-6.
Because our data suggest that OA and overexpressing nhr-80 are distinct interventions, we next verified whether NHR-80 and OA must work in concert to mediate germline longevity. To address this, we first investigated whether OA could extend lifespan in the absence of nhr-80. Convincingly, the lifespan of glp-1(e2141ts);nhr-80(tm1011) mutant animals is not affected by OA (Figure 8A, Table S1). Conversely, we overexpressed nhr-80 in animals that could no longer produce OA (glp-1(e2141ts);fat-6(tm331);fat-7(wa36)) and found that it fails to increase the lifespan of these animals (Figure 8B, Table S1). In this experiment, we noted that overexpressing nhr-80 does not increase the lifespan of glp-1(e2141ts);fat-6(tm331);fat-7(wa36) mutants complemented with OA. This is surprising because nhr-80 overexpression increases the lifespan of glp-1(e2141)ts mutants (Figure 4B, Table S1). The reason for this is unclear but may be explained by the fact that OA is not as efficient when it is externally provided since it is prone to oxidation. Finally, we showed that the effect of OA on the lifespan of glp-1(e2141ts);fat-6(tm331);fat-7(wa36) mutant animals is no longer significant when nhr-80 is knocked down by RNAi (Figure 8C, Table S1). Taken together, our data suggest that OA and NHR-80 promote longevity in concert. This can be simply explained by the assumption that the OA producing pathway is not the only longevity-promoting branch downstream of NHR-80 (Figure 9).
In the present study, we show that, when germ cells are removed, nhr-80 mRNA and protein levels increase. This promotes the mono-desaturation of stearic acid to OA by inducing the transcription of the stearoyl-CoA-desaturase, fat-6/SCD1. This cascade is physiologically relevant to longevity since both nhr-80 and the SCD activity are required to augment the lifespan of germline-depleted animals. Furthermore, the lack of SCD activity can be bypassed by addition of exogenous OA in the medium, confirming the pivotal role of this metabolite.
Our data also indicate that the FAT-6/OA branch is required, but not sufficient, to promote longevity in response to depletion of the germ line downstream of NHR-80. This is evidenced by the fact that overexpressing nhr-80 extends the lifespan of both glp-1(e2141ts) and daf-16(mu86);glp-1(e2141ts) mutants but OA does not, suggesting that these two interventions are not equivalent (Figures 4B, 5C, 7A, and S13A, Table S1). Supporting this view, we showed that OA and NHR-80 must act in concert to support the lifespan extension conferred by germ cell loss. Indeed, OA does not increase the lifespan of glp-1(e2141ts);nhr-80(tm1011) mutants and overexpressing nhr-80 is inefficient when the SCD genes are deleted (OA producing genes). Moreover, while OA restores the lifespan of glp-1(e2141ts);fat-6(tm331);fat-7(wa36) mutant animals, it essentially fails to do so when nhr-80 is deactivated by RNAi (Figure 8C, Table S1). Thus, our data support the notion that the OA production pathway is not the only longevity-promoting branch downstream of NHR-80. Finding other lifespan promoting NHR-80 targets will be an important goal in the future.
Our data are also compatible with the non-exclusive hypothesis that OA may act as a NHR-80 ligand. Although we do not provide biochemical evidence for this interaction in C. elegans, several articles have shown that long chain free fatty acids act as a ligand for the NHR-80 homolog in Drosophila and mammals, HNF4 [19]–[23]. It will be interesting to critically test this possibility in the future by either performing structural studies or transactivation assays to test the binding of OA to the NHR-80 ligand binding domain.
Our results suggest that lipid metabolism and, in particular, fatty acid desaturation links signals from the germ line to longevity. This confirms previous findings suggesting a link between fat metabolism and longevity in germline-less animals through the K04A8.5 lipase [12]. However, results presented here argue that the NHR-80/FAT-6/OA and the KRI-1/DAF-16/K04A8.5 pathways can act independently. First, we observed that the increase in nhr-80 mRNA levels observed in glp-1(e2141ts) mutant animals still occurs in the absence of daf-16 (Figure 5B). Second, the translocation of DAF-16 into intestinal nuclei occurs in the absence of nhr-80 (Figure 5A). Third, we show that NHR-80 and DAF-16 have distinct transcriptional targets, although some overlap between the two transcription factors exists. The transcription of fat-6 is elevated in glp-1(e2141ts) mutant animals in a NHR-80 dependent manner (Figure 6B) and K04A8.5 mRNA levels are increased in a DAF-16 dependent way (Figure 6D; [12]). fat-5 is also increased, but its induction relies on both DAF-16 and NHR-80 (Figure 6A). Fourth, the overexpression of nhr-80 increases the lifespan of daf-16(mu86);glp-1(e2141ts) mutant animals (Figure 5C, Table S1), demonstrating that NHR-80 signaling does not require the presence of daf-16. Finally, OA does not increase the lifespan of daf-16(mu86);glp-1(e2141ts) mutants, suggesting that, similar to glp-1(e2141ts) mutants, the SCD activity is already elevated in these animals (Figure S12A, Table S1).
Thus, our data confirm that germline ablation leads to an alteration of fat metabolism that is required for extending lifespan but challenge the view that this is centered on insulin signaling. Germ cell removal extends longevity in response to two independent fat modifying pathways: the NHR-80/FAT-6/OA and the KRI-1/DAF-16/K04A8.5 pathways.
DAF-12 is also required for longevity extension by depletion of the germ line [1], and it seems to act upstream of DAF-16 since its presence is partially required for DAF-16 translocation into intestinal nuclei in response to germline ablation [4]. However, recent observations suggest that DAF-12 can also act in parallel to DAF-16 [11],[12]. Our data show that the NHR-80/FAT-6/OA and DAF-12 act in concert, independently of DAF-16. Indeed, overexpressing nhr-80 and providing exogenous OA do not affect the lifespan of glp-1(e2141ts);daf-12(rh61rh411) mutant animals (Figures 5D, S13B, Table S1), indicating that the NHR-80/FAT-6/OA pathway requires the presence of DAF-12. The expression levels of daf-12 are not affected in glp-1(e2141ts);nhr-80(tm1011) mutants and the induction of nhr-80 mRNA levels occurs in the absence of daf-12, but it is slightly decreased relative to glp-1(e2141ts) mutants, suggesting that, similar to DAF-16, DAF-12 modulates the NHR-80 response (Figure 5F). Our data therefore indicate that the DAF-12/NHR-80 interaction is not strictly transcriptional. Rather, we propose that DAF-12 and NHR-80′s targets act together to promote longevity or that NHR-80 and DAF-12 share some critical targets for longevity. Since fat-6 mRNA levels are normally induced in glp-1(e2141ts);daf-12(rh61rh411) double mutants relative to wild type, we can exclude the possibility that the FAT-6/OA branch acts downstream of DAF-12. Thus, other targets must be involved. We propose a model where DAF-12 and NHR-80 targets cooperate to promote longevity in concert with the FAT-6/OA branch. However, it is also possible that DAF-12 interacts directly with NHR-80. Whether DAF-12 and NHR-80 only cooperate through their critical targets or whether they physically interact remains to be determined.
We were surprised to find that overexpressing nhr-80 extends the lifespan of glp-1(e2141ts);daf-9(rh50) double mutants in the presence or in the absence of Δ7 dafachronic acid, the DAF-12 ligand produced by DAF-9 ([10]; Figures 5E and S5, Table S1). This suggests that NHR-80 interacts with DAF-12 independently of DAF-9 and Δ7 dafachronic acid and explains why treating wild type animals overexpressing nhr-80 with Δ7 dafachronic acid (DAF-12 ligand) fails to recapitulate germline longevity (unpublished data). We note that the Kenyon lab had already suggested that DAF-12 might function in a DAF-9 independent manner [4]. Indeed, mutations causing DAF-16 to be constitutively nuclear extends the lifespan of germline-less animals lacking daf-9, but not of animals lacking daf-12 [4].
It is possible that DAF-12 can also be activated by other ligands or that it can interact with NHR-80 under its unliganded form. DAF-12 might also be activated by an unknown cofactor or by a heterodimeric partner nuclear receptor to interact with NHR-80. The finding that the overexpression of nhr-80 fails to extend the lifespan of wild type animals (where DAF-12 is not activated; Figure 4A, Table S1) suggests indeed that other ligand(s) or co-activators may only be present in germline-less animals. However, we cannot exclude that other, yet unidentified, modulators preclude lifespan extension through nhr-80 overexpression. In the future, it will be interesting to explore these non-exclusive possibilities.
Similar to daf-2 mutant animals, germline-depleted animals store more fat than wild type animals [24]. This is not a systematic trait of long-lived animals since diet-restricted animals store less fat. Thus, higher fat content is not a general cause for life extension. However, it is still not clear whether high fat content extends lifespan when the germ line is depleted. First, the high fat content phenotype [24] is hard to reconcile with the finding that the K04A8.5 lipase is induced in these animals [12]. Although RNAi against the K04A8.5 lipase increases Nile Red staining in glp-1(e2141ts) mutant animals, these data are difficult to interpret since it was shown that Nile Red does not stain fat reliably [24]. It is possible that the K04A8.5 lipase changes the composition of fat rather than altering overall fat content by degrading a subset of TAGs only. Second, the nhr-80(tm1011) allele does not affect the overall fat content of glp-1(e2141ts) mutants (unpublished observation). This indicates that fat content is not the cause for longevity extension in germline-depleted animals since glp-1(e2141ts);nhr-80(tm1011) mutants are short-lived. Our data suggest that fatty acid desaturation, and therefore, fat composition, is altered in germline-less animals. This modification of fat composition does not directly impact fat content but correlates with longevity. Indeed, our data clearly establish that OA production is not sufficient to extend lifespan, but it is strictly required. Further work should aim at understanding all aspects of fat metabolism that matter to germline-mediated longevity.
Although nhr-80 is one of 269 genes encoding nuclear hormone receptors in C. elegans that are homologous to the mammalian HNF4 gene [25], there are interesting clues that suggest that the mechanism we describe in this work may be conserved. First, we report that NHR-80 and OA must act in concert to promote longevity. Although we failed to provide clear evidence that OA activates NHR-80, structural and biochemical data showing a direct interaction between long chain fatty acids and HNF4 in Drosophila and mice [19]–[23] suggest that a similar regulation may also occur in C. elegans and that OA may be an NHR-80 ligand. Second, we show that fat-6 is a target of NHR-80. Strikingly, SCD1, the mammalian homolog of FAT-6, is strongly up-regulated in response to ovariectomy in mice [26] and is a known target for several nuclear receptors in mammals. Taken together, these data suggest that our work may be relevant in mammalian systems. Further experiments are underway to firmly identify the NHR-80 analog.
N2 Bristol was used as the wild-type strain. Nematodes were grown and maintained under standard conditions [27]. HGA8011, HGA8013, and BX156 were grown in the presence of Oleic acid until day 1 of adulthood to bypass any developmental delays. C. elegans strains (i.e., genotype, origin, strain name) are listed below:
nhr-80(tm1011)III, CGC*, BX165
fat-5(tm420)V, CGC, BX107
fat-6(tm331)IV, CGC, BX106
fat-7(wa36)V, CGC, BX153
fat-6(tm331)IV; fat-7(wa36)V, CGC, BX156
fat-6(tm331)IV; fat-5(tm420)V, CGC, BX160
fat-7(wa36)V; fat-5(tm420)V, CGC, BX110
daf-12(m20)X, CGC, DR20
daf-16 (mu86)I, CGC, AD105
mes-1(bn7)X, CGC, SS149
glp-1(e2141ts)III, Gift from Kenyon Lab, CF1903
glp-1(e2141ts)III; daf-12(rh61rh411)X, Gift from Kenyon Lab, CF1658
glp-1(e2141ts)III; daf-9(rh50)X, Gift from Kenyon Lab, CF1916
daf-16(mu86)I; glp-1(e2141ts)III; muIs109, Gift from Kenyon Lab, CF1935
daf-16(mu86)I; glp-1(e2141ts)III, Gift from Kenyon Lab, CF1880
glp-1(e2141ts)III; nhr-80(tm1011)III, Made in our Lab, HGA8000
N2; lynEx**, made in the laboratory, HGA8001
glp-1(e2141ts)III; lynEx, made in the laboratory, HGA8002
glp-1(e2141ts)III; nhr-80(tm1011)III; lynEx, made in the laboratory, HGA8003
daf-16(mu86)I; glp-1(e2141ts)III; lynEx, made in the laboratory, HGA8004
glp-1(e2141ts)III;daf-12(rh61rh411)X; lynEx, made in the laboratory, HGA8005
glp-1(e2141ts)III; fat-5(tm420)V, made in the laboratory, HGA8006
glp-1(e2141ts)III; fat-6(tm331)IV, made in the laboratory, HGA8007
glp-1(e2141ts)III; fat-7(wa36)V, made in the laboratory, HGA8008
glp-1(e2141ts)III; fat-7(wa36)V; fat-5(tm420)V, made in the laboratory, HGA8009
glp-1(e2141ts)III; fat-6(tm331)IV; fat-5(tm420)V, made in the laboratory, HGA8010
glp-1(e2141ts)III; fat-6(tm331)IV; fat-7(wa36)V, made in the laboratory, HGA8011
glp-1(e2141ts)III; daf-9(rh50)X; LynEx, made in the laboratory, HGA8012
glp-1(e2141ts)III; fat-6(tm331)IV; fat-7(wa36)V;LynEx, made in the laboratory, HGA8013
daf-2(e1370)III, CGC, CB1370
* CGC = Caenorhabditis Genetics Center
** lynEx = [(pJG01(nhr-80p::nhr-80::gfp) and co-injection marker myo-2p::DsRed]
Double and triple mutant strains were generated using standard genetic procedures. The glp-1(e2141ts) mutation was assayed by testing sterility and lack of germ line at the restrictive temperature (25°C). Presence of the nhr-80(tm1011), fat-5(tm420), fat-6(tm331), or fat-7(wa36) mutations were assayed by PCR using allele specific primers. To generate nhr-80p::nhr-80::gfp expressing animals (HGA8001; HGA8002; HGA8003; HGA8004; HGA8005; HGA8012; and HGA8013), pJG01 was injected as described [28] at 50 ng/µL. Co-injection marker myo-2p::DsRed was injected at 20 ng/µL. Transgenes are called lynEx. The myo-2p::DsRed marker and pPD95.75 (empty vector) have no effect on life span (unpublished data).
RNAi clones from the Ahringer's Library were grown overnight at 37°C in LB containing Ampicillin (50 µg/mL) and Tetracyclin (12.5 µg/mL). Each RNAi clone was spread on NGM plates supplemented with carbenicillin (25 µg/mL). RNAi expression was induced by the addition of 1 mM IPTG on top of seeded bacteria. About 150 glp-1(e2141ts) (CF1903) and wild type L1 larvae were added on each RNAi plate and incubated at 25°C until day 1 of adulthood. At day 20 the proportion of worm alive was visually inspected. Clones that led to a majority of dead worms at this time for both wild type and glp-1(e2141ts) mutant animals were selected for further analysis.
The plasmid pJG01(nhr-80p::nhr-80::gfp) containing nhr-80 tagged with GFP driven by the endogenous nhr-80 promoter was constructed by amplifying genomic DNA from 1.4 kb upstream from the start codon until the end of the nhr-80 coding sequence without the stop codon. The 4kb XmaI/KpnI nhr-80 fragment generated was inserted upstream and in frame of the GFP sequence in the worm expression vector pPD95.75 (Fire Lab Vector Kit–Addgene). Essential parts of the plasmid pJG01 were sequenced.
Primer sequences:
Nhr80-XmaI-F: 5′-GGGGTGCCCCCGGGGGGATCGAGACACTTTTCTTACTCCTC-3′
Nhr80-KpnI-R: 5′-AACGGGGTACCCCGTTTTTCAAGCTTTGCCTGACCCA-3′
Lifespan assays were conducted according to standard protocols [29]. All assays were performed at 20°C, starting from day 1 of adulthood. For glp-1(e2141ts) mutants, lifespan assays (and associated controls), animals were grown at 25°C from the L1 stage until the L4 stage to prevent germ cell proliferation. The rest of the assay was performed at 20°C. For cyc-1 RNAi experiments (and associated controls), larvae were left at 15°C for 24 h, shifted to 25°C for 24 h, and shifted back to 20°C at the L4 stage to avoid dauer formation or other larval arrest. All strains containing the glp-1(e2141ts) allele used in lifespan assays were completely sterile. Unless mentioned otherwise, lifespan assays of fertile strains were conducted on plates supplemented with 15 µM 5-Fluoro-Uracil in order to prevent progeny from hatching. Worms crawling off the plate, exploding, bagging, or contaminated were excluded. Plotting and statistical analysis were done using the Biopylife software. Biopylife was designed by students from INSA de Lyon using the following free softwares: R, MySQL, Python, MySQLdb, Qt4, PyQt4, and MacTeX. Biopylife allows easy plotting of lifespan assays and determines mean, maximal lifespan, and p values using log-rank (Mantel-Cox) statistics.
NGM medium were prepared with the addition of 100 µM oleic acid (NuChek Prep) right before pooring plates.
Δ7 dafachronic acid was added on top of seeded bacteria to make a final NGM concentration of 100 nM. Worms were transferred on fresh plates every other day.
DR was performed through bacterial deprivation from the fifth day of adulthood [30].
For each gene, analyses were performed on triplicate of at least five independent extracts. For each analysis, we present two statistical tests: the parametric unpaired two-tailed t student test and the non-parametric Wilcoxon rank test. While the former test assumes a Gaussian distribution of the samples, the latter does not. Asterisk indicates the p value of the Wilcoxon rank-sum test as follows: *p<0.1, **p<0.05, ***p<0.01. Standard deviations are displayed as error bars.
DAF-16 nuclear localization assays were performed as previously described [4]. Briefly, daf-16(mu86);glp-1(e2141ts);muIs109 [Pdaf-16::gfp::daf-16] mutant animals were injected with water (negative control), kri-1 RNAi (control), or nhr-80 RNAi at 1 mg/mL and recovered at 15°C. The next day, animals were shifted to 20°C to lay eggs for several hours on OP50 plates. To obtain F1 animals with a glp-1(e2141ts) phenotype, eggs were shifted to 25°C, 24 h after being laid and shifted back to 20°C at the L4 stage. To obtain F1 animals without any glp-1(e2141ts) phenotype, eggs were left at 20°C during the whole course of the experiment. On day 1 of adulthood, F1 animals were assayed for DAF-16 nuclear localization in intestinal cells using a fluorescent microscope. Animals were scored as having nuclear-localized DAF-16 if the majority of intestinal cells displayed a distinct concentration of GFP in the nucleus. Twenty-five to 50 animals were analyzed for each condition.
The MetaMorph software was used for image processing and ImageJ was used for fluorescence quantification. Fluorescence microscopy was performed using an axioplan microscope (Zeiss), with a cooled charge-coupled device camera. GFP images of sterile and fertile worms (i.e., grown until L4 at permissive or restrictive temperature and then switched at 20°C) were collected at day 1 of adulthood. Regions of Interest (ROI) were manually designed within either the two first intestinal nuclei and head neuronal nuclei; the software (ImageJ) calculated the mean intensity value for pixel intensity within the ROIs. At least 10 animals were analyzed for each condition. Averages and errors are presented in our graphs.
Each condition was analyzed in three independent extracts. For each extract, 4,000 to 5,000 adult worms at day 1 of adulthood grown at 25°C on ht115 were washed three times in M9 buffer. Worms were then homogenized in 2 ml of methanol/ 5 mM EGTA (2∶1 v/v) with FAST-PREP (MP Biochemicals). 100 ml were evaporated and the dry pellets were dissolved in 0.2 ml of NaOH (0.1 M) overnight and proteins were measured with the Bio-Rad assay. Lipids corresponding to the total homogenate were extracted according to Bligh and Dyer [31] in chloroform/methanol/water (2.5/2.5/2.1, v/v/v), in the presence of the internal standards glyceryl triheptadecanoate (5 g). The lipid extracts were hydrolysed in KOH (0.5 M in methanol) at 50°C for 30 min, and transmethylated in boron trifluoride methanol solution 14% (SIGMA, 1 ml) and hexane (1 ml) at 80°C for 1 h. After addition of water (1 ml) to the crude extract, FAMEs were extracted with hexane (3 ml), evaporated to dryness, and then dissolved in ethyl acetate (180 ml). FAME (1 ml) were analyzed by gas-liquid chromatography [32] on a Clarus 600 Perkin Elmer system using a Famewax RESTEK fused silica capillary columns (30 m×0.32 mm i.d., 0.25 m film thickness). Oven temperature was programmed from 110°C to 220°°C at a rate of 2°C per min and the carrier gas was hydrogen (0.5 bar). The injector and the detector were at 225°C and 245°C, respectively.
|
10.1371/journal.pmed.1002309 | Contribution of systemic and somatic factors to clinical response and resistance to PD-L1 blockade in urothelial cancer: An exploratory multi-omic analysis | Inhibition of programmed death-ligand 1 (PD-L1) with atezolizumab can induce durable clinical benefit (DCB) in patients with metastatic urothelial cancers, including complete remissions in patients with chemotherapy refractory disease. Although mutation load and PD-L1 immune cell (IC) staining have been associated with response, they lack sufficient sensitivity and specificity for clinical use. Thus, there is a need to evaluate the peripheral blood immune environment and to conduct detailed analyses of mutation load, predicted neoantigens, and immune cellular infiltration in tumors to enhance our understanding of the biologic underpinnings of response and resistance.
The goals of this study were to (1) evaluate the association of mutation load and predicted neoantigen load with therapeutic benefit and (2) determine whether intratumoral and peripheral blood T cell receptor (TCR) clonality inform clinical outcomes in urothelial carcinoma treated with atezolizumab. We hypothesized that an elevated mutation load in combination with T cell clonal dominance among intratumoral lymphocytes prior to treatment or among peripheral T cells after treatment would be associated with effective tumor control upon treatment with anti-PD-L1 therapy. We performed whole exome sequencing (WES), RNA sequencing (RNA-seq), and T cell receptor sequencing (TCR-seq) of pretreatment tumor samples as well as TCR-seq of matched, serially collected peripheral blood, collected before and after treatment with atezolizumab. These parameters were assessed for correlation with DCB (defined as progression-free survival [PFS] >6 months), PFS, and overall survival (OS), both alone and in the context of clinical and intratumoral parameters known to be predictive of survival in this disease state.
Patients with DCB displayed a higher proportion of tumor-infiltrating T lymphocytes (TIL) (n = 24, Mann-Whitney p = 0.047). Pretreatment peripheral blood TCR clonality below the median was associated with improved PFS (n = 29, log-rank p = 0.048) and OS (n = 29, log-rank p = 0.011). Patients with DCB also demonstrated more substantial expansion of tumor-associated TCR clones in the peripheral blood 3 weeks after starting treatment (n = 22, Mann-Whitney p = 0.022). The combination of high pretreatment peripheral blood TCR clonality with elevated PD-L1 IC staining in tumor tissue was strongly associated with poor clinical outcomes (n = 10, hazard ratio (HR) (mean) = 89.88, HR (median) = 23.41, 95% CI [2.43, 506.94], p(HR > 1) = 0.0014). Marked variations in mutation loads were seen with different somatic variant calling methodologies, which, in turn, impacted associations with clinical outcomes. Missense mutation load, predicted neoantigen load, and expressed neoantigen load did not demonstrate significant association with DCB (n = 25, Mann-Whitney p = 0.22, n = 25, Mann-Whitney p = 0.55, and n = 25, Mann-Whitney p = 0.29, respectively). Instead, we found evidence of time-varying effects of somatic mutation load on PFS in this cohort (n = 25, p = 0.044). A limitation of our study is its small sample size (n = 29), a subset of the patients treated on IMvigor 210 (NCT02108652). Given the number of exploratory analyses performed, we intend for these results to be hypothesis-generating.
These results demonstrate the complex nature of immune response to checkpoint blockade and the compelling need for greater interrogation and data integration of both host and tumor factors. Incorporating these variables in prospective studies will facilitate identification and treatment of resistant patients.
| A new type of cancer treatment called checkpoint blockade therapy activates the immune system to fight cancer.
When these therapies work, patients with advanced disease can experience long-lasting disease control or even cures.
However, most patients will not experience these benefits, and it is crucial to identify these patients in advance so that we can develop better treatments for them.
In this study, we studied 29 patients with advanced bladder cancers treated with a checkpoint blockade drug called atezolizumab.
We examined features of the tumor and the immune system, as well as clinical features.
We found that these features were related to each other, and to the success of therapy, in various ways.
Patients who had a diverse repertoire of T cells in their blood tended to survive longer. Patients who had poor clinical prognostic factors, like having cancer that had traveled to their liver, tended to have worse survival.
This study demonstrates that we need to take the tumor, immune system, and clinical picture into account if we are to improve the efficacy of immune-mobilizing therapies in cancer.
Some patients may be too sick to benefit from checkpoint blockade therapy, despite, in some cases, having biomarkers in their tumors that would predict benefit.
| Atezolizumab has demonstrated responses in 15%–25% of patients with advanced urothelial carcinoma and improved survival compared to historical expectations [1,2]. Similar to predictive factor analyses in melanoma, colon cancer, and non-small cell lung cancer studies with other checkpoint blockade agents, Rosenberg and colleagues reported a statistically significant association between mutation load and response to atezolizumab in urothelial cancer patients [2]. However, mutation load in the atezolizumab study was predicted based on an estimate using a targeted panel and not with whole exome sequencing (WES). Similar to findings from prior studies, the association between this predicted mutation load and outcomes in patients with urothelial cancer was not dichotomous; there were tumors from patients with elevated mutation load that did not respond to therapy, and vice versa. Additionally, positive programmed death-ligand 1 (PD-L1) staining of infiltrating immune cells by immunohistochemistry was associated with, but poorly predicted, response. A statistical model suggested that both PD-L1 staining and mutation load impacted the likelihood of response. However, the authors did not recommend their clinical use.
Collectively, studies to date imply that a combination of immune parameters are necessary to gain further precision in determining the likelihood of benefit from these immunotherapies and that a single biologic marker will be insufficient. There have been few attempts to integrate molecular and immunologic data from patients treated with checkpoint blockade and their tumors. Consequently, we performed whole exome, RNA, and T cell receptor (TCR) sequencing (TCR-seq) of tumor samples from patients treated with atezolizumab, as well as TCR-seq of matched, serially collected peripheral blood.
All research involving human participants was approved by the authors' Institutional Review Board (MSKCC IRB), and all clinical investigation was conducted according to the principles expressed in the Declaration of Helsinki. Written informed consent was obtained from the participants.
The analyses that were and were not included in the prespecified analysis plan are detailed in S2 Text.
All patients had locally advanced or metastatic urothelial carcinoma and were treated at Memorial Sloan Kettering Cancer Center (n = 29) on protocol NCT02108652 [2]. All patients initiated therapy in 2014, were treated with atezolizumab 1,200 mg IV every 3 weeks, and provided written consent according to Institutional Review Board-approved protocols permitting tissue and blood collection, sequencing, and correlative studies. Patient tumor samples were assessed prospectively and centrally (by HistoGeneX, Brussels, Belgium) for PD-L1 expression by immunohistochemistry with the SP142 assay (Ventana, AZ, USA) [1]. The PD-L1 tumor-infiltrating immune cell (IC) status was defined by the percentage of PD-L1–positive ICs in the tumor microenvironment: IC0 (<1%), IC1 (≥1% but <5%), and IC2/3 (≥5%), as defined in the original study. Six patients had multiple samples evaluated for PD-L1 IC status; for 4 of these patients, the sample used for PD-L1 IC status in this analysis was the same as the sample that was whole exome sequenced. For the remaining 2, the PD-L1 status of 1 patient’s tumor (1994) used a separate primary tumor sample that also agreed in status with a metastatic sample; the other patient’s tumor (6229) used a metastatic sample site that agreed in status with another metastatic sample site. Smoking status was evaluated using previously completed self-reported smoking questionnaires or review of medical records. One patient was excluded because the patient did not consent to correlative studies beyond PD-L1 testing that was performed as part of the clinical trial.
All tumor tissue used for sequencing was obtained prior to dosing with atezolizumab. Tumor samples used for whole exome sequencing were all formalin-fixed paraffin-embedded (FFPE). The presence of tumor tissue in the sequenced samples was confirmed by examination of a representative hematoxylin and eosin-stained slide by a genitourinary pathologist (H.A.). Peripheral blood mononuclear cells (PBMCs) were isolated and stored as previously described [3]. PBMCs were collected pretreatment and during treatment.
Tumor responses to atezolizumab were evaluated by CT scan every 9 weeks for the first 12 months following day 1 of cycle 1. After 12 months, tumor assessments were performed every 12 weeks. The response evaluation criteria in solid tumors (RECIST) version 1.1 was used to define objective clinical responses by the institutional radiologist.
Tumor samples from patients 0522 and 6800 were excluded from tumor TCR analyses after failing quality control. Patients 0979, 7592, and 8214 did not have available tumors for TCRβ sequencing and were therefore excluded from tumor TCR analyses as well. Genomic DNA was purified from total PBMCs and tumor samples using the Qiagen DNeasy Blood extraction kit. The TCRβ CDR3 regions were amplified and sequenced using immunoSEQ® (Adaptive Biotechnologies, Seattle, WA), as previously described [4]. In brief, bias-controlled V and J gene primers were used to amplify rearranged V(D)J segments for high-throughput sequencing at approximately 20X coverage. After correcting sequencing errors via a clustering algorithm, CDR3 segments were annotated according to the International ImMunoGeneTics Collaboration [5] to identify the V, D, and J genes that contributed to each rearrangement. A mixture of synthetic TCR analogs in each PCR was used to estimate the absolute template abundance (i.e., the number of cells bearing each unique TCR sequence) from sequencing data, as previously described [6]. The estimated tumor-infiltrating T lymphocytes (TIL) content was calculated as previously described [6–8]. To determine TIL content in FFPE samples as a T cell fraction, we amplified several housekeeping genes and quantitated their template counts to determine the amount of DNA usable for TCRβ sequencing. ImmunoSEQ then amplifies and sequences the molecules with rearranged TCRβ chains. Because the immunoSEQ assay aligns sequences to the IMGT database, sequences are annotated as complete VDJ rearrangements or nonproductive rearrangements (a stop codon or out of frame CDR3 region was generated during VDJ recombination in 1 of the alleles); all downstream analysis in this work proceeded with complete, productive sequences. To estimate the number of starting templates that were in the sample, the number of sequence reads for each TCRβ sequence is measured. Synthetic control templates were also spiked into each sample, thereby enabling quantitation of input TCRβ templates from the read counts. For each sample, Shannon entropy was also calculated on the clonal abundance of all productive TCR sequences in the data set. Shannon entropy was normalized to the range by dividing Shannon entropy by the logarithm of the number of unique productive TCR sequences in the data set. This normalized entropy value was then inverted to produce the clonality metric. Those T cell clones whose frequencies differed between samples from a given subject taken at different time points, or between cell populations (e.g., between total PBMCs and tumor), were computationally identified as previously described [9]. The input data consisted of the absolute abundance for each TCR clone in each sample. Fisher’s exact test was used to compute a p-value for each clone across the 2 samples against the null hypothesis that the population abundance of the clone is identical in the 2 samples. We corrected for multiple testing to control the false discovery rate (FDR) using the Benjamini-Hochberg procedure and employed a significance threshold of 0.01 on adjusted p-values.
Twenty-six FFPE-derived tumor and frozen PBMC-derived normal paired samples were sequenced by exome hybrid capture, using the IDT xGen Whole Exome Panel (https://www.idtdna.com/pages/products/nextgen/target-capture/xgen-lockdown-panels/xgen-exome-panel) and standard protocols. Briefly, each sample was used to create a barcoded Illumina library, tumor samples were pooled at optimal multiplex to create an equimolar pool into the hybrid capture reaction, which was performed according to the manufacturer’s suggested protocol. Similarly, normal samples were pooled and introduced to the hybrid capture reaction. Following the recovery of captured library fragments, PCR amplification was performed, the resulting pools of fragments were quantitated using qPCR (Kapa Bio) and sequenced in separate lanes by paired-end 150 bp reads, using the Illumina HiSeq 4000. Whole exome sequencing results for 1 sample (for patient 4072) were excluded after failing to meet coverage requirements.
DNA sequencing data for the tumor and normal samples were aligned to the GRCh37 reference using bwa-mem (v. 0.7.10) with default settings. The resulting BAMs were processed through Picard MarkDuplicates and the GATK (v. 3.5–0) pipeline, including Base Quality Score Recalibration and Indel Realignment. Single nucleotide variants (SNVs) were called from Mutect (v. 1.1.6) and Strelka (v. 1.0.14) with default settings. Variants from either call were included and the variants calls were further filtered to those with depth (in normal and tumor samples) ≥7 reads, >10% tumor variant allele frequency (VAF), and ≤3% normal VAF [10]. Mutations per megabase was computed by normalizing the number of mutations by the number of exonic loci with ≥7 reads in normal and tumor samples, calculated using Pageant (https://github.com/hammerlab/pageant).
Variants were annotated as missense variants by Varcode (v. 0.5.10, https://github.com/hammerlab/varcode) and PyEnsembl (v. 1.0.3, https://github.com/hammerlab/pyensembl) using Ensembl Release 75 and annotated as deleterious using PolyPhen (v. 2.2.2). DNA damage response (DDR) genes were gathered from [11,12].
Mutational signatures were inferred from the somatic mutation calls using deconstructSigs (v 1.6.0).
Human leukocyte antigen (HLA) types for each patient were computed from the normal sequencing data using OptiType (v. 1.0.0).
RNA was extracted from 26 FFPE tumor samples and evaluated for quality and quantity using the Agilent RNA pico chip. Each sample was prepared for sequencing by constructing an Illumina Tru-Seq Stranded RNA kit, according to the manufacturer’s protocol. The resulting libraries were amplified by PCR, quantitated, pooled, and processed through a hybrid capture intermediate using the IDT xGen Exome reagent (as above). The captured fragments were quantitated, diluted, and were sequenced using 2 x 150 bp paired-end reads on the Illumina HiSeq 4000.
The RNA sequencing (RNA-seq) data were aligned to the GRCh37 reference in Ensembl Release 75 using STAR (v. 2.4.1d), and transcript quantification was performed using kallisto (v. 0.42.3). The STAR alignment was only used for identifying variant-supporting reads in the RNA. For gene-level analysis, the transcript quantifications were aggregated to the gene level using tximport (http://f1000research.com/articles/4-1521/v1).
Expressed mutations and neoantigens were computed using Isovar (v. 0.0.6, https://github.com/hammerlab/isovar), based on the RNA reads overlapping each mutation. Sleuth (v. 0.28.1) was used for differential expression analysis, and Gene Set Enrichment Analysis (GSEA) was used for pathway enrichment analysis. ESTIMATE was used to quantify immune and stromal scores from RNA-seq data.
Neoantigens were computed from all nonsynonymous mutations using Topiary (v. 0.1.0, https://github.com/hammerlab/topiary) and NetMHCCons (v. 1.1) with HLA alleles calculated by OptiType. As with expressed mutations, expressed neoantigens were those supported in the RNA with at least 3 uniquely-mapped reads matching the cDNA sequence.
All statistical analysis was performed in Python and R (v. 3.3.1). Cohorts (v. 0.4.0, https://github.com/hammerlab/cohorts) and Biokepi (https://github.com/hammerlab/biokepi) were used to orchestrate the analysis. The Mann Whitney and Fisher’s Exact test were performed using the Python scientific computing library, SciPy (v. 0.18.1). Kaplan-Meier curves were computed with Lifelines (v. 0.9.1.0). Survival and logistic regression models were estimated using PyStan (v. 2.12.0.0), and the Stan statistical computing software (v. 2.12.0). Survival analyses utilized a proportional hazards piecewise exponential model with a random walk baseline hazard. The analysis for presence of a time-varying covariate effect was performed in R using survival (v. 2.39.5) to look for the association of scaled Schoenfeld residuals with log(time), whereas the estimation of the time-varying covariate effect was performed using Stan. This analysis estimated the covariate effect at each timepoint with a random-walk prior. In some cases, alternative specifications of models written in Stan were interrogated as sensitivity analyses; see the project’s GitHub repository (https://github.com/hammerlab/multi-omic-urothelial-anti-pdl1) and S3 Text for details.
All analysis code is available at https://github.com/hammerlab/multi-omic-urothelial-anti-pdl1 for open access by readers.
Mutation calls, TCR-seq, and RNA-seq data are available at http://doi.org/10.5281/zenodo.546110. Additional data are available at https://github.com/hammerlab/multi-omic-urothelial-anti-pdl1.
Twenty-nine patients with metastatic urothelial cancer from a single institution, treated with atezolizumab, as part of a single-arm phase II study (IMvigor 210, NCT 02108652), were included in the analyses. The patients displayed characteristics typical of the metastatic urothelial cancer population studied in IMvigor 210: 25 of 29 were males with urothelial cancers of bladder origin, and 18 of 29 had a reported prior smoking history (Table 1). Patients had an Eastern Cooperative Oncology Group (ECOG) performance status of 0 or 1 and had 0 to 3 prior regimens of chemotherapy. Of this group, 25 patients had sufficient tumor tissue for WES, 26 for RNA-seq, and 24 for TCR-seq. Twenty-nine had a pretreatment peripheral blood sample on which TCR-seq could be performed; 24 had 1 pretreatment and at least 1 posttreatment peripheral blood collection.
The importance of T cells to the anti-tumor response has long been known [15]; the relevance of intratumoral and peripheral TCR clonality to the anti-tumor response is an area of active study. A single previous study of melanoma patients treated with anti-PD-1 therapy demonstrated that patients whose tumors featured both high levels of tumor-infiltrating T lymphocytes (TIL) along with high TIL clonality were more likely to experience radiographic response to therapy [8]. A separate study examined the peripheral TCR repertoire in anti-cytotoxic T-lymphocyte-associated protein 4 (CTLA-4)–treated patients with prostate cancer or melanoma and found that clonotype stability was associated with response [16]. To our knowledge, no prior study has reported both intratumoral and peripheral TCR clonality in a single population treated with checkpoint blockade therapy.
We performed TCR-seq of tumors and PBMCs at serial time points in our cohort. Due to limitations in sample availability (Methods), this analysis included tumors from 24 patients and peripheral blood from 29 patients, including pretreatment samples in all patients, and between 1 and 8 total time points. A median of 141,255 (range 43,052–335,089) T cells were analyzed per peripheral blood sample, including 82,636 (range 24,095–207,860) unique TCRs, with a median clonality of 0.080 (range 0.014–0.37) and a median T cell proportion of 0.31 (range 0.082–0.64). In the tumors, the corresponding values included 1,402 (range 63–133,167) T cells, 1,086 (range 67–56,273) unique TCRs, clonality of 0.096 (range 0.033–0.34), and T cell proportion of 0.097 (range 0.0098–0.33).
In our patient group, we first asked whether there was an association between outcome and either TIL clonality or TIL proportion, or with clonality in the peripheral blood. Consistent with the data from Tumeh and colleagues [8], tumors from patients who experienced a durable clinical benefit (DCB) exhibited a higher TIL proportion than those patients who experienced progressive disease, with a median of 0.21 (range 0.049–0.33) in tumors from patients who had progression-free survival (PFS) greater than 6 months versus 0.069 (range 0.0098–0.24) in tumors from patients who did not (n = 24, Mann-Whitney p = 0.047, Fig 1A). The consistency of results is notable given that Tumeh and colleagues used a different methodology (IHC) to assess TIL proportion than was used in our study. However, TIL proportion was not associated with continuous PFS (n = 24, log-rank p = 0.32) or OS (n = 24, log-rank p = 0.26) when split by median proportion. TIL clonality alone was not significantly associated with DCB (n = 24, Mann-Whitney p = 0.10, Fig 1A), continuous PFS (split by median clonality, n = 24, log-rank p = 0.51), or continuous OS (split by median clonality, n = 24, log-rank p = 0.47). Tumors with less than the median TIL proportion or TIL clonality, considered jointly as 1 feature, were less likely to display DCB (Fig 1B, 25% of patients with DCB versus 81% of patients without DCB, n = 24, Fisher's Exact p = 0.021). Considering TIL proportion and TIL clonality separately, when using median as a threshold, did not result in a significant difference in terms of DCB in either case (S1A Fig). It remains unclear whether TIL clonality adds to TIL proportion in its association with DCB in this study (TIL proportion and TIL clonality versus TIL proportion alone, n = 24, log-likelihood p = 0.100).
We next examined pretreatment peripheral blood clonality and its relationship to DCB. Because a diverse TCR repertoire in circulation may increase the likelihood that a tumor-specific T cell population is present, we hypothesized that TCR clonality would be inversely associated with response. We found that low pretreatment peripheral TCR clonality was associated with improved PFS (split by median clonality, n = 29, log-rank p = 0.048), overall survival (OS, split by median clonality, n = 29, log-rank p = 0.011), and OS greater than 12 months (DCB-OS, n = 29, Mann-Whitney p = 0.0061; Fig 1C, 1D and 1E), although not with DCB (Fig 1F, n = 29, Mann-Whitney p = 0.25).
Finally, we explored the relationship between intratumoral and peripheral TCR clonality. Individual T cell clones present in tumors can be tracked in the peripheral blood during treatment (examples in S1B Fig). Expansion of tumor-associated TCRs occurred in the peripheral blood in all patients (Fig 1G). However, a more pronounced expansion of intratumoral TCR clones was observed in DCB patients at 3 weeks after initiation of treatment (second dose of therapy) (n = 22, Mann-Whitney p = 0.022, Fig 1H) that was not significant at 6 weeks after therapy initiation (n = 20, Mann-Whitney p = 0.17, S1C Fig). Interestingly, all patients with low pretreatment peripheral TCR clonality and high TIL clonality survived greater than 12 months (DCB-OS, S1D Fig).
To further examine intratumoral factors associated with therapeutic efficacy, we performed WES on 25 FFPE archived tumor samples. Mean target coverage was 129 (range 44–194) in tumors and 73 (range 59–91) in normal tissue. SNVs were identified and annotated as silent, missense, or nonsense mutations (Fig 2A). There was no significant association between median missense mutation load and DCB (median mutations per megabase 3.24 [range 0.038–11.46] in patients with DCB compared to 0.45 [range 0.019–9.90] in those without DCB, n = 25, Mann-Whitney p = 0.22, Fig 2B). There was also no significant association between missense mutation load and DCB-OS (n = 25, Mann-Whitney p = 0.37, S2A Fig). In a survival analysis for time to disease progression or mortality, the estimated hazard ratio associated with increase in missense SNV count per megabase was 0.92 (95% CI [0.78, 1.09]). These results are not surprising given that the present sample size (n = 25) is underpowered to detect an effect of magnitude similar to that observed by Rosenberg and colleagues [2] (power = 0.2, assuming median of 12.4 versus 6.4 mutations per megabase among patients with DCB versus non-DCB response).
When filtering to expressed mutations, we found a median of 0.79 (range 0.00–3.36) expressed mutations per megabase for patients with DCB and a median of 0.16 (range 0.00–3.34) expressed mutations per megabase for patients without DCB (n = 25, Mann-Whitney p = 0.26, S2B Fig). Consistent with the known importance of specific variant calling pipelines to output [17,18], we found that different filtering techniques impacted the association with DCB (S1 Table). Missense mutation load, when counting only mutations that were removed after postprocessing (via Base Quality Score Recalibration [BQSR] and depth/VAF filtering), was predictive of response (n = 25, Mann-Whitney p = 0.0078).
One hypothesis for explaining the association between mutation load and outcome to treatment with checkpoint blockade is the generation of neoantigens, altered peptides presented by the major histocompatibility complex that are capable of eliciting an antitumor T cell response and are more common with increased mutation load. After performing in silico HLA typing (Methods), we examined predicted neoantigens that were 8 to 11 amino acids in length, resulting from the missense mutations of patients treated with atezolizumab. There was no significant association between predicted neoantigens per megabase and either DCB or DCB-OS. Patients with DCB had a median 4.58 (range 0.037–39.48) predicted neoantigens per megabase, while patients without DCB had 1.35 (range 0.00–20.22) (n = 25, Mann-Whitney p = 0.55, S2C Fig and S2D Fig). Filtering of predicted neoantigens to focus only on those expressed in RNA (Methods) also demonstrated no significant association between expressed predicted neoantigens and clinical benefit with atezolizumab (n = 25, Mann-Whitney p = 0.29, Fig 2C and S2B Fig). Here, too, we acknowledge the limitations in statistical power to detect associations due to the sample size of our study.
Given that the mutation load and outcomes were weakly associated in the complete IMvigor210 dataset and not statistically significantly associated in this cohort, we embarked upon an exploration of additional factors, including tumor microenvironmental and systemic measures, which may modify the importance of this variable or independently affect outcomes.
To this end, we examined the time-varying association between mutation load and PFS to see if mutation load had a differential association with early hazards in contrast to late hazards. We found evidence of time-varying effects of somatic mutation load on PFS in this cohort (n = 25, p = 0.044, see Methods). We estimated these effects to consist of a stronger association of somatic mutation load with reduced mortality or disease progression more than 3 months after treatment (n = 11, hazard ratio (HR) = 0.69, 95% CI [0.38, 0.99]) as compared to that during the first 3 months (n = 25, HR = 0.91, 95% CI [0.75, 1.07]; Fig 3A). This effect estimate yielded a p-value for interaction of 0.1, which does not contradict the test for presence of the effect (n = 25, p = 0.044) because that test is better powered. When a similar analysis was performed for time-varying association with OS, the evidence in support of the existence of time-varying effects was similar (n = 25, p = 0.082; notable, despite p > 0.05, given the PFS results above). In terms of the estimate of these effects, patients who survived longer than 3 months exhibited a stronger association between the number of somatic mutations per megabase and a lower risk of subsequent mortality (n = 11, HR = 0.80, 95% CI [0.60, 1.00]) as compared to those who survived less than or equal to 3 months (n = 25, HR = 1.02, 95% CI [0.79, 1.22], Fig 3B). This suggests that the time-varying effect is not likely an artifact of differential association with survival versus progression. Looking at the Kaplan-Meier estimates of PFS among patients with mutation load per megabase above and below the median value of 1.03, it is apparent that there is very little separation of these 2 populations until approximately 3 months after treatment and that there is a high frequency of both progression and mortality events at this time (Fig 3C). While we report these results using a prespecified threshold of 3 months, in a nonparametric analysis we found that the reduction in risk associated with somatic mutation load increased steadily over time without the emergence of a clear inflection point (S3C Fig; see S3 Text and S3D Fig for further model interrogation). That said, we note that the 95% confidence intervals around the hazard ratios prior to 3 months include 1 (all p-values > 0.05; aggregate p(HR > 1) = 0.21), while those after 3 months are significantly less than 1 (all p-values < 0.05; aggregate p(HR > 1) = 0.020; p = 0.07 for interaction comparing aggregate HRs). This suggests that our threshold of 3 months, which was selected based on clinical experience, may be a convenient summary of these results.
These data suggest that in patients with rapidly progressive disease, factors other than mutation load likely determine their outcome. This observation is not surprising in that clinical factor analysis of this disease state has identified a heterogeneous population of patients, with 5 clinical factors distinguishing those likely to experience a rapid and early death from those more likely to survive longer [13]. We hypothesized that such patients might simply be too clinically and systemically unwell to mount the necessary immune response, despite some of them harboring tumor biomarkers thought to confer a likelihood of DCB, including elevated mutation load. When we examined the 5-factor score in this subset relative to the rest of the dataset, we found that, indeed, patients who survived less than or equal to 3 months exhibited a significantly higher 5-factor score (3.00 (range 2.00–4.00) in contrast to 1.50 (range 0.00–4.00) in patients who survived longer than 3 months (n = 26, Mann-Whitney p = 0.018, S3A Fig). Patients surviving less than 3 months were much more likely to have liver metastases: 100% in patients surviving less than or equal to 3 months and 22% in patients surviving longer than 3 months (n = 29, Fisher's Exact p = 0.00097, S3B Fig). There were no significant differences in these patients with respect to Bacillus Calmette–Guérin (BCG) exposure (n = 29, Fisher's Exact p = 0.20), missense SNV load (n = 25, Mann-Whitney p = 0.26), and pretreatment peripheral TCR clonality (n = 29, Mann-Whitney p = 0.12). Keeping in mind the limited sample size of this cohort, these data suggest that there is a subset of nearly end-stage patients with cancer in whom clinical variables may negate immunological response, despite the presence of 1 or more favorable tumor-associated biomarkers. The inclusion of these clinical variables is warranted in future studies.
Several studies have suggested that an “inflamed” tumor microenvironment, tumor, or IC PD-L1 expression increase the likelihood of response to checkpoint blockade. As seen in the published IMVigor 210 cohort, PD-L1 IC expression was significantly associated with DCB in this subset (n = 29, Spearman rho = 0.48 p = 0.0083, S4A Fig). We quantified immune infiltration from RNA-seq using ESTIMATE [19]. The immune score, while associated with the TIL proportion estimated through TCR-seq (S4B Fig), was estimated to be 764.37 (range −1195.08 to 1509.65) in patients with DCB and 263.49 (range −1100.78 to 1734.28) in patients without DCB but was not significantly different (n = 26, Mann-Whitney p = 0.33, S4C Fig). When we performed GSEA using the Hallmark Geneset [20], we did not observe any differentially expressed gene sets between tumors from patients with DCB versus no DCB. Furthermore, RNA expression of PD-L1 did not correlate with reported IC PD-L1 staining level (n = 26, Spearman rho = 0.045 p = 0.83, S4D Fig). We did not observe a difference in tumor MHC class I expression according to DCB (S4E Fig, HLA-A: n = 26, Mann-Whitney p = 0.26, HLA-B: n = 26, Mann-Whitney p = 0.36, HLA-C: n = 26, Mann-Whitney p = 0.24).
Given that such agnostic approaches did not reveal a clear association between tumor microenvironment factors and response, we pursued a hypothesis-driven approach examining the genes that show up-regulation at the cell surface during T cell exhaustion. When categorized by DCB, there was no significant difference in expression of such genes, including CTLA-4, TIGIT, HAVCR2 (TIM-3), or LAG-3 [21]. When grouped by PD-L1 staining, we found low expression of all markers in the PD-L1 low group (IC0), as expected. However, in the PD-L1 high group (IC2), HAVCR2 exhibited significantly higher expression in tumors from patients who experienced DCB than in those who did not (S4F Fig). Interestingly, of the 4 IC2 tumors among HAVCR2-high patients (HAVCR2 expression greater than the median), 2 had missense SNV loads at or below the median (2 and 57); the other 2 had 180 and 412 SNVs. Additionally, although Rosenberg and colleagues [2] found that luminal cluster II showed a significantly higher response rate among the 4 subtypes of RNA expression from The Cancer Genome Atlas (TCGA), no significant association was found here between the 4 clusters and DCB (n = 20, Fisher's Exact p = 0.36) (S4G Fig) nor between the luminal/basal subcategorization and DCB (n = 20, Fisher's Exact p = 1.00), possibly due to sample size.
Unanswered questions that arise from the many studies of biomarker correlates of checkpoint blockade response are whether measures such as mutation load, PD-L1 staining, and others reflect the same “tumor state” or if each confers an independent effect on outcome?
When examined in conjunction with mutation load, the greater the expression of PD-L1, the more negative the association of mutation load with hazard (i.e., higher mutation load was associated with longer survival). Among patients with tumors showing little-to-no expression of PD-L1 (IC0 rated), each unit increase in missense SNV count per megabase was associated with a negligible change in hazard (n = 4, HR = 1.43, 95% CI [0.75, 2.98]). Among patients with tumors expressing PD-L1 at moderate or high levels (IC1 or IC2 staining), missense SNV count per megabase was associated with lower risk for disease progression or mortality (among IC1: n = 11, HR = 0.75, 95% CI [0.47, 1.14]; among IC2: n = 10, HR = 0.73, 95% CI [0.48, 1.06]). Although our limited sample size precludes making an assertion that mutation load is associated with survival in any particular subgroup (e.g., when looking among IC1 and IC2 tumors alone), our data do support the presence of an interaction among these variables (p = 0.046 for interaction; S5A Fig). Given the plausibility of the finding that somatic mutation load may correlate better with survival among patients with an inflamed tumor microenvironment, the addition of somatic mutation load to PD-L1 IC staining warrants further study.
We found a similar, albeit weaker, interaction effect when looking at the association of somatic mutation load (missense SNV count per megabase) and PFS, according to the presence/absence of liver metastasis prior to treatment administration (p = 0.14 for interaction). Among patients without liver metastasis, somatic mutation load was associated with a lower risk for disease progression or mortality (n = 16, HR = 0.73, 95% CI [0.50, 1.02], S5B Fig) than patients with liver metastasis (n = 9, HR = 0.96, 95% CI [0.66, 1.37], S5B Fig).
To our surprise, although both PD-L1 staining and mutation load were each associated with response in the original study [2], these variables did not correlate with each other (Fig 4A). Furthermore, pretreatment peripheral TCR clonality did not correlate with mutation load (Fig 4B). The lack of association between these variables suggests that each might have an independent or semi-independent role in determining the likelihood of response to therapy. TCR clonality and infiltration did, however, correlate with PD-L1 IC score: those tumors with higher clonality or higher infiltration also featured higher PD-L1 staining (p = 0.02 and p = 0.01, respectively, Fig 4C and 4D).
In an analysis to see whether the association between pretreatment peripheral TCR clonality and PFS varied by PD-L1 IC score, we found some evidence of an interaction (p = 0.015 for interaction; Fig 4E). Among patients with low levels of PD-L1 expression, there was little association between pretreatment peripheral TCR clonality and PFS (among IC0: n = 4, HR [mean] = 1.86, HR [median] = 1.55, 95% CI [0.50, 4.99], p(HR > 1) = 0.21; among IC1: n = 11, HR [mean] = 0.69, HR [median] = 0.58, 95% CI [0.15, 1.84], p(HR < 1) = 0.19). Among patients with high levels of PD-L1 expression, by comparison, we observed almost complete separation of PFS according to pretreatment peripheral TCR clonality (among IC2: n = 10, HR [mean] = 89.88, HR [median] = 23.41, 95% CI [2.43, 506.94], p(HR>1) = 0.0014; Fig 4E). Similar results were seen in analyses with respect to OS and in a logistic regression analysis for DCB (S5C Fig, S5D Fig, S5E Fig).
To resolve the hypothesis that those patients with low peripheral TCR clonality simply were healthier, we examined the association between 5-factor score and pretreatment peripheral TCR clonality and did not find such an association (n = 26, Spearman rho = 0.25 p = 0.22, S5F Fig).
In a multivariate survival model for time to disease progression or mortality, which allows the effect of each biomarker to vary according to intratumoral PD-L1 IC score, we find that the correlation of each intratumoral, peripheral, or clinical biomarker with disease progression or mortality is relatively independent of the others (Fig 4F, S5G Fig). Perhaps with the notable exception of the association between liver metastatic status and time to progression or survival, the correlation of each intratumoral or peripheral biomarker with outcome is strongest in the group with the highest levels of IC PD-L1 expression (S2 Table).
The treatment of previously incurable metastatic solid tumors with checkpoint blockade agents has led to dramatic success in a minority of patients, a finding that has generated substantial excitement in the field, with associated correlative studies and drug development. Here, we undertook the in-depth characterization of tumors and peripheral blood from 29 patients treated on IMvigor 210, a Phase II study in which 310 patients were treated with the anti-PD-L1 agent atezolizumab. In this cohort of patients, we illustrate the importance of host immune factors, including intratumoral and peripheral TCR clonality, infiltration, and expansion, to clinical outcomes. We did not find a significant association between mutation or expressed neoantigen load and PFS or DCB (defined as PFS > 6 months). However, we did observe a time-dependent relationship between mutation load and outcome, wherein a relationship between mutation burden and outcome could only be detected in those patients surviving greater than 3 months. This analysis implies that patients who experienced rapid progression may display systemic indicators of immune deficiency despite elevated mutation load in the tumors. Calculation of the hazard ratios for each measured biomarker and clinical factor underscores the concept that a complex interaction of both host and tumor variables determines whether a patient will experience clinical benefit from anti-PD-L1 therapy. Although the overall study found significant associations between mutation load as measured by the Foundation Medicine targeted sequencing panel and radiographic response [2], there was no statistically significant association between mutation load and DCB or survival in the patient subset studied here, despite the similarity of our study population to the parent study. This contrast may be due to a combination of factors. First, though statistically significant, the association in the overall study was not categorical: as in other studies of mutation load, this factor alone was not predictive of response. Second, we have less power to detect this association in our smaller subset compared with the larger studied cohort. Third, standardized definitions and calculations of mutation load do not exist as of yet; each published study has used differing methodologies [2,10,22,23]. Indeed, in this study, depending on the method used, the association between mutation load and clinical outcomes varied from p < 0.08 to p > 0.4 (area under the curve [AUC] values and p-values in S1 Table). To illustrate the fickle nature of defining mutation load, counting only the mutations excluded by BQSR, as opposed to only those remaining after BQSR, showed a significant association with DCB. Together, these findings underscore the need for improved and standardized mutation calling methods. The weak association of mutation load with DCB and the lack of such standardization render this biomarker unfit for application to individual patients at present. Furthermore, if validated in another dataset, this analysis implies that a clinical and immunological state may exist in patients with advanced cancer, such that patients with very rapidly progressing disease and expected death in less than 3 months do not respond despite the presence of positive intratumoral biomarkers.
In an attempt to deepen our understanding of the biology of response and resistance, we studied additional factors. We found that even in this small dataset, TCR clonality below the median in the peripheral blood prior to treatment, expansion of tumor-associated TCR in the periphery 3 weeks after initiating treatment, and higher TIL proportion are all associated with clinical benefit. These data suggest that TCR-seq provided additional insights into response and resistance beyond mutation load and PD-L1 staining. With respect to biomarker development, our study implies that noninvasive metrics such as pretreatment peripheral TCR clonality and known prognostic features such as the presence of liver metastases may be worthy of further study in urothelial cancer patients treated with PD-L1 blockade.
From a mechanistic perspective, these findings imply an important relationship between circulating and intratumoral immunity upon PD-L1 blockade. We hypothesize that low TCR clonality in the peripheral blood prior to treatment increases the likelihood that a patient harbors 1 or several clones capable of tumor recognition, whether of neoantigens or tumor-associated antigens. The expansion of tumor-associated TCRs in the peripheral blood underscores the continuity of the tumor and blood compartments, and suggests that the activity of PD-L1 blockade may involve circulating T cells more than was previously thought. Indeed, this raises the possibility that anti-tumor T cells may home from the periphery to the tumor before later recirculating.
Finally, though limited in power by the small sample size, we attempted to integrate the importance of the studied variables. This analysis demonstrated both hypothesized and unexpected interactions. For example, while mutation load seemed to be associated with outcome more significantly in PD-L1 IC1 and IC2 tumors, high PD-L1 IC staining in the setting of high peripheral TCR clonality was associated with a substantial hazard for poor outcome. Given the significance of PD-L1 expression in mediating response to anti-PD-L1 therapy, the presence of these interactions may argue in their favor as predictive rather than prognostic biomarkers. Further analysis is required to elucidate the role of these biomarkers in mediating response to checkpoint blockade.
This study has several limitations. The patients under study were treated at a single institution and represent a small subset of the overall study, limiting statistical power. As a single-arm Phase II study, there is no control arm for comparison. Tumor samples were FFPE and were not collected immediately prior to treatment initiation. Only 1 sample per patient was utilized, which does not necessarily capture the heterogeneity of each tumor. Finally, a number of analyses were performed on a small number of patients without an independent validation cohort; although most analyses were prespecified, there is a risk of Type II error, and no adjustments were made for multiple testing.
In conclusion, we have demonstrated the potential value of pursuing an integrated study of somatic, immune, and clinical features in order to elucidate the biological mechanisms underlying response to checkpoint blockade and ultimately improve our ability to practice precision medicine. We hope this work will motivate further multi-omics studies of response to checkpoint blockade.
|
10.1371/journal.pbio.1002207 | CDK8-Cyclin C Mediates Nutritional Regulation of Developmental Transitions through the Ecdysone Receptor in Drosophila | The steroid hormone ecdysone and its receptor (EcR) play critical roles in orchestrating developmental transitions in arthropods. However, the mechanism by which EcR integrates nutritional and developmental cues to correctly activate transcription remains poorly understood. Here, we show that EcR-dependent transcription, and thus, developmental timing in Drosophila, is regulated by CDK8 and its regulatory partner Cyclin C (CycC), and the level of CDK8 is affected by nutrient availability. We observed that cdk8 and cycC mutants resemble EcR mutants and EcR-target genes are systematically down-regulated in both mutants. Indeed, the ability of the EcR-Ultraspiracle (USP) heterodimer to bind to polytene chromosomes and the promoters of EcR target genes is also diminished. Mass spectrometry analysis of proteins that co-immunoprecipitate with EcR and USP identified multiple Mediator subunits, including CDK8 and CycC. Consistently, CDK8-CycC interacts with EcR-USP in vivo; in particular, CDK8 and Med14 can directly interact with the AF1 domain of EcR. These results suggest that CDK8-CycC may serve as transcriptional cofactors for EcR-dependent transcription. During the larval–pupal transition, the levels of CDK8 protein positively correlate with EcR and USP levels, but inversely correlate with the activity of sterol regulatory element binding protein (SREBP), the master regulator of intracellular lipid homeostasis. Likewise, starvation of early third instar larvae precociously increases the levels of CDK8, EcR and USP, yet down-regulates SREBP activity. Conversely, refeeding the starved larvae strongly reduces CDK8 levels but increases SREBP activity. Importantly, these changes correlate with the timing for the larval–pupal transition. Taken together, these results suggest that CDK8-CycC links nutrient intake to developmental transitions (EcR activity) and fat metabolism (SREBP activity) during the larval–pupal transition.
| Arthropods are estimated to account for over 80% of animal species on earth. Characterized by their rigid exoskeletons, juvenile arthropods must periodically shed their thick outer cuticles by molting in order to grow. The steroid hormone ecdysone plays an essential role in regulating the timing of developmental transitions, but exactly how ecdysone and its receptor EcR activates transcription correctly after integrating nutritional and developmental cues remains unknown. Our developmental genetic analyses of two Drosophila mutants, cdk8 and cycC, show that they are lethal during the prepupal stage, with aberrant accumulation of fat and a severely delayed larval–pupal transition. As we have reported previously, CDK8-CycC inhibits fat accumulation by directly inactivating SREBP, a master transcription factor that controls the expression of lipogenic genes, which explains the abnormal fat accumulation in the cdk8 and cycC mutants. We find that CDK8 and CycC are required for EcR to bind to its target genes, serving as transcriptional cofactors for EcR-dependent gene expression. The expression of EcR target genes is compromised in cdk8 and cycC mutants and underpins the retarded pupariation phenotype. Starvation of feeding larvae precociously up-regulates CDK8 and EcR, prematurely down-regulates SREBP activity, and leads to early pupariation, whereas re-feeding starved larvae has opposite effects. Taken together, these results suggest that CDK8 and CycC play important roles in coordinating nutrition intake with fat metabolism by directly inhibiting SREBP-dependent gene expression and regulating developmental timing by activating EcR-dependent transcription in Drosophila.
| In animals, the amount of juvenile growth is controlled by the coordinated timing of maturation and growth rate, which are strongly influenced by the environmental factors such as nutrient availability [1,2]. This is particularly evident in arthropods, such as insects, arachnids and crustaceans, which account for over 80% of all described animal species on earth. Characterized by their jointed limbs and exoskeletons, juvenile arthropods have to replace their rigid cuticles periodically by molting. In insects, the larval–larval and larval–pupal transitions are controlled by the interplay between juvenile hormone (JH) and steroid hormone ecdysone [3–7]. Drosophila has been a powerful system for deciphering the conserved mechanisms that regulate hormone signaling, sugar and lipid homeostasis, and the molecular mechanisms underlying the nutritional regulation of development [1,2,8–11]. In Drosophila, all growth occurs during the larval stage when larvae constantly feed, and as a result their body mass increases approximately 200-fold within 4 d, largely due to de novo lipogenesis [12]. At the end of the third instar, pulses of ecdysone, combined with a low level of JH, trigger the larval–pupal transition and metamorphosis [3,6,13]. During this transition, feeding is inhibited, and after pupariation, feeding is impossible, thus the larval–pupal transition marks when energy metabolism is switched from energy storage by lipogenesis in larvae to energy utilization by lipolysis in pupae.
The molecular mechanisms of ecdysone-regulated metamorphosis and developmental timing have been studied extensively in Drosophila [3,5,14,15]. Ecdysone binds to the Ecdysone Receptor (EcR), which heterodimerizes with Ultraspiracle (USP), an ortholog of the vertebrate Retinoid X Receptor (RXR) [16–21]. By activating the expression of genes whose products are required for metamorphosis, ecdysone and EcR-USP are essential for the reorganization of flies’ body plans before emerging from pupal cases as adults. Despite the tremendous progress in our understanding of the physiological and developmental effects of EcR-USP signaling, the molecular mechanism of how the EcR-USP transcription factor interacts with the general transcription machinery of RNA polymerase II (Pol II) and stimulates its target gene expression remains mysterious. EcR is colocalized with Pol II in Bradysia hygida and Chironomus tentans [22,23]. Although a number of proteins, such as Alien, Bonus, Diabetes and Obesity Regulated (dDOR), dDEK, Hsc70, Hsp90, Rigor mortis (Rig), Smrter (Smr), Taiman, and Trithorax-related (TRR), have been identified as regulators or cofactors of EcR-mediated gene expression [13,24–32], it is unknown how these proteins communicate with the general transcription machinery and whether additional cofactors are involved in EcR-mediated gene expression. In addition, it remains poorly understood how EcR activates transcription correctly after integrating nutritional and developmental cues.
The multisubunit Mediator complex serves as a molecular bridge between transcriptional factors and the core transcriptional machinery, and is thought to regulate most (if not all) of Pol II-dependent transcription [33–40]. Biochemical analyses have identified two major forms of the Mediator complexes: the large and the small Mediator complexes. In addition to a separable “CDK8 submodule”, the large Mediator complex contains all but one (MED26) of the subunits of the small Mediator complex [36,38,41]. The CDK8 submodule is composed of MED12, MED13, CDK8, and CycC. CDK8 is the only enzymatic subunit of the Mediator complex, and CDK8 can both activate and repress transcription depending on the transcription factors with which it interacts [37,42]. Amplification and mutation of genes encoding CDK8, CycC, and other subunits of Mediator complex have been identified in a variety of human cancers [43,44], however, the function and regulation of CDK8-CycC in non-disease conditions remain poorly understood. CDK8 and CycC are highly conserved in eukaryotes [45], thus analysis of the functional regulation of CDK8-CycC in Drosophila is a viable approach to understand their activities.
Previously, we have shown that CDK8-CycC negatively regulates the stability of sterol regulatory element-binding proteins (SREBPs) by directly phosphorylating a conserved threonine residue [46]. We now report that CDK8-CycC also regulates developmental timing in Drosophila by linking nutrient intake with EcR-activated gene expression. We show that homozygous cdk8 or cycC mutants resemble EcR mutants in both pupal morphology and retarded developmental transitions. Despite the elevation of both EcR and USP proteins in cdk8 or cycC mutants, genome-wide gene expression profiling analyses reveal systematic down-regulation of EcR-target genes, suggesting the CDK8-CycC defect lies between the receptor complex and transcriptional activation. CDK8-CycC is required for EcR-USP transcription factor binding to EcR target genes. Mass spectrometry analysis for proteins that co-immunoprecipitate with EcR and USP has identified multiple Mediator subunits, including CDK8 and CycC, and our yeast two-hybrid assays have revealed that CDK8 and Med14 can directly interact with the EcR-AF1 domain. Furthermore, the dynamic changes of CDK8, EcR, USP, and SREBP correlated with the fundamental roles of SREBP in regulating lipogenesis and EcR-USP in regulating metamorphosis during the larval–pupal transition. Importantly, we show that starving the early third instar larvae causes precocious increase of CDK8, EcR and USP proteins, as well as premature inactivation of SREBP; whereas refeeding of the starved larvae reduces CDK8, EcR, and USP proteins, but potently stimulates SREBP activity. These results suggest a dual role of CDK8-CycC, linking nutrient intake to de novo lipogenesis (by inhibiting SREBP) and developmental signaling (by regulating EcR-dependent transcription) during the larval–pupal transition.
The Drosophila cdk8 and cycC genes were originally identified based on the function and sequence conservation to their yeast and human orthologs [47–49]. cdk8K185 and cycCY5 are null alleles that delete part of cdk8 (882 bp) and all of cycC (2,733 bp), respectively, and the homozygous mutants are both prepupal lethal [50]. Mutant animals are able to develop to prepupae, likely due to maternally loaded CDK8 and CycC mRNAs and proteins, because embryos derived from the cycCY5 germline clones are smaller and are embryonic lethal without proper denticle formation (S1 Fig). In contrast to the wild-type pupae (Fig 1A), 96% of cdk8K185 (Fig 1B) and 97% of cycCY5 (Fig 1C) homozygous mutants fail to evert their anterior spiracles (quantified in S2A Fig), and prepupae of both mutants are partially separated from their pupal cases (arrows in Fig 1B and 1C). In addition, pupariation is delayed by about 2 to 3 d in the cdk8 and cycC mutants (Fig 1G and 1H).
To investigate the effects of CDK8-CycC on developmental timing, we first analyzed the cdk8-cycC double mutant animals by genetically combining the cdk8K185 and cycCY5 null alleles in the same organism. The phenotypes in the cdk8-cycC double mutant animals were similar to cdk8 or cycC single mutants, including pupal morphology (Figs 1D and S2A), delayed pupariation (Fig 1G and 1H), and prepupal lethality. The levels of cdk8 and cycC mRNA (S2B Fig) and their protein products (S2C Fig) are diminished in cdk8 or cycC single and double mutant larvae when assayed at the third instar larval stage (L3). The protein level of CycC is significantly reduced in cdk8 and cycC mutants, but the level of CDK8 is not affected in cycC mutants (S2C Fig), thus the stability of CycC is dependent on CDK8 but not vice versa.
To validate that the loss of CDK8-CycC causes the defects in pupal morphology and development, we tested whether the mutant phenotypes could be rescued by expression of wild-type CDK8 or CycC. Since CDK8 and CycC form the CDK8 submodule with MED12 and MED13 in a 1:1:1:1 stoichiometry [51], proper dosage of these four subunits is critical for the formation and function of a viable CDK8 sub-module. To ensure proper expression levels and patterns, we generated transgenic flies using genomic fragments of cdk8 and cycC loci with EGFP tags at their C-termini (S3 Fig). The X-ray crystal structure of human CDK8-CycC complex demonstrates that the C-termini of CDK8 and CycC are not involved in their interaction [52], thus epitope tags fused to C-termini were expected to avoid functional disruption of the CDK8-CycC complex. These constructs were transposed to chromosome 2; the transgenic flies are referred to as “cdk8+-EGFP” or “cycC+-EGFP” for simplicity. We genetically combined these transgenes with cdk8 or cycC null alleles, thus CDK8 or CycC proteins were tagged with EGFP in the rescued animals (“w1118; cdk8+-EGFP; cdk8K185” for cdk8-rescued animals, and “w1118; cycC+-EGFP; cycCY5” for cycC-rescued animals). The genotypes of the rescued adult animals were validated by PCR analysis (S3 Fig). Importantly, these transgenic lines rescue both the pupal morphology (Figs 1E and 1F, and S2A) and developmental timing (Fig 1G and 1H). The rescued animals are no longer prepupal lethal, and they emerge as adult flies. These observations indicate that CDK8 and CycC are required for proper developmental transitions in Drosophila.
The phenotypes of cdk8 and cycC mutants (Fig 1) are reminiscent of loss-of-function alleles of EcR-B1, the major EcR isoform that controls the larval-to-pupal transition [53]. The EcR gene encodes three isoforms (EcR-A,-B1, and-B2) that are expressed in tissue- and developmental stage-specific manners [21,53,54]. To test the possibility that CDK8-CycC and EcR-B1 regulate similar molecular events that control the developmental transitions, we first examined whether the expression of EcR target genes was affected in cdk8 or cycC mutants. By mining the microarray data that we published previously [46], we analyzed the mRNA levels of 67 genes whose products are related to the ecdysone and JH activities as reported in the literature (S1 Table) [3,6]. In cdk8 or cycC mutants, the mRNA levels of 33 of these genes are significantly decreased whereas mRNA levels of 10 of these genes are increased more than 1.5-fold compared to the control. Most of the down-regulated genes are EcR-activated genes, while most of the up-regulated genes respond to JH activity (Fig 2A and S1 Table).
We used qRT-PCR assays to verify the levels of several well-characterized direct target genes of EcR, such as broad, E74, E75, E78, Hsp27 (Heat shock protein 27), ImpE2 (Ecdysone-inducible gene E2), Sgs1 (Salivary gland secretion 1), and Sgs5 [3,6]. As shown in Fig 2B, the expression of these EcR-activated genes was significantly reduced in L3 wandering cdk8 and cycC mutants. There was a small reduction of EcR mRNA levels, but a mild increase of usp mRNA levels, in the cdk8 and cycC mutants (Fig 2C). EcR normally represses the expression of the mid-prepupal gene βFtz-F1 during the larval stage [15,55]; however, the expression of βFtz-F1 was dramatically increased in the cdk8 and cycC mutant larvae (Fig 2C), suggesting that the function of EcR is disrupted in the cdk8 and cycC mutants. Likewise, the levels of jheh1 (JH-epoxide hydrolase) and JhI-26 (JH-inducible protein 26) were significantly increased in cdk8 and cycC mutants (Fig 2C). JHEH1 is involved in the catabolic processing of JH, while the expression of JhI-26 is induced by either methoprene or JH III [56]. Therefore, the expression of both EcR and JH-regulated genes was deregulated in cdk8 and cycC mutants, consistent with the developmental retardation phenotype (Fig 1G and 1H).
To test whether ecdysone-induced EcR target gene expression was generally compromised in cdk8 or cycC mutants, we analyzed the effect of cdk8 or cycC mutation on the expression of the multimerized hsp27 EcRE (ecdysone response element)-lacZ reporter [21,54]. In response to the treatment of 20-hydroxyecdysone (20E), the most biologically potent EcR ligand, β-galactosidase activity was induced in the control salivary gland cells as expected (Fig 2D versus 2D’). However, this response was significantly compromised in salivary glands from the cdk8 (Fig 2E’) and cycC (Fig 2F’) homozygous mutants; the glands from the same animals (Fig 2E and 2F, respectively) were used as the controls. These results were consistent with reduced expression of EcR target genes in cdk8 or cycC mutants (Fig 2A and 2B).
Both the ecdysone ligand and the EcR-USP transcription factor complex are required for the expression of EcR target genes. 20E directly binds to the ligand-binding domain (LBD) of EcR, which then activates EcR target gene expression [3,6,13]. Therefore, down-regulated expression of EcR target genes in cdk8 and cycC mutants may be due to defective biosynthesis of 20E, or defects in EcR-activated transcription. To test whether the biosynthesis of 20E is defective in cdk8 and cycC mutants, we analyzed the expression of enzymes that are required for the biosynthesis of 20E, such as nvd (neverland, encoding an oxygenase-like protein) and a family of cytochrome P450 enzymes including dib (disembodied), phm (phantom), sad (shadow), shd (shade), spo (spook), and spok (spookier), collectively known as the Halloween genes [57–59]. The expressions of sad and spok were decreased in the cdk8 and cycC mutants, but the expression of nvd and other Halloween genes were not significantly affected (Fig 3A). In addition, we analyzed the mRNA levels of cyp18a, which encodes a cytochrome P450 enzyme involved in degradation of 20E [60], and a few genes encoding factors involved in regulating the expression of the Halloween genes, such as mld, kni, and vvl [61]. As shown in Fig 3B, no obvious changes of these genes were observed in both cdk8 and cycC mutants. Nevertheless, reduction of sad and spok in cdk8 and cycC mutants indicates that the biosynthesis of ecdysone may be defective in the cdk8 and cycC mutants.
Next, we measured the levels of ecdysteroids in cdk8 and cycC mutants from early L3 larval stage to white prepupal (WPP) stage. Compared to the control, the levels of ecdysteroids are significantly lower in cdk8 mutant animals during the wandering L3 and WPP stages than the control, while the levels of ecdysteroids are lower in cycC mutants only during the late wandering stage (Fig 3C). Nevertheless, the levels of ecdysteroids are continuously increased from early L3 to WPP stage in both cdk8 and cycC mutants, indicating that the biosynthesis of ecdysteroids is compromised, but not completely abolished, in these mutants.
To further determine whether the developmental retardation in cdk8 and cycC mutants is caused by impaired 20E biosynthesis, we fed the homozygous mutants with fly food supplemented with 200 μM of 20E, which is an established approach used to examine whether developmental defects are caused by mutations that disrupt biosynthesis of ecdysteroid [62–64]. However, the defective prepupal morphology of the cdk8 and cycC mutants was not rescued (Fig 3D). In contrast, food supplement of 20E rescued animals with prothoracic gland (PG) cells ablated by PG-specific expression of reaper gene (Fig 3E), which triggers apoptosis, or animals with spok specifically depleted in the PG (Fig 3F) using a PG-specific driver (phm-Gal4). Therefore, supplement of 20E to larvae depleted in factors required for ecdysone biosynthesis rescues their developmental delay [61]. Consistent with reduced ecdysteroids level in cdk8 mutants (Fig 3C), feeding cdk8 mutant with 200 μM of 20E in food had a mild effect on the time from egg deposition to pupariation compared to the control (Fig 3G). However, feeding cycC mutants with 20E had no effect on their developmental delay, also consistent with the weaker effect of cycC mutation on ecdysteroids levels (Fig 3C and 3G). Importantly, the larval–pupal transition is still significantly retarded in cdk8 or cycC mutants, even when fed with 20E (Fig 3G). Other concentrations of 20E in food, ranging from 2 μM to 2mM, also failed to rescue the developmental defects of the cdk8 and cycC mutants (S4 Fig). These results suggest that defective biosynthesis of 20E alone is not sufficient to explain the developmental defects of cdk8 or cycC mutants, which is consistent with the strongly compromised effects of 20E on EcRE-lacZ expression in salivary gland cells in cdk8 and cycC mutants (Fig 2E and 2E’ and 2F and 2F’). Considering that CDK8 and CycC function as subunits of the transcription cofactor Mediator complex, which is known to regulate the transcriptional activity of several nuclear hormone receptors in mammals [65,66], the most likely scenario is that the cdk8 and cycC mutants are defective in the regulation of EcR-dependent gene expression in peripheral tissues, in addition to impairing ecdysone biosynthesis in the PG.
To understand how loss of CDK8 or CycC reduced EcR-target gene expression, we tested whether the protein levels of EcR and USP were affected in cdk8 and cycC mutants. Since the major defects occurred during the larval–pupal transition, we first analyzed the expression of EcR, USP, CDK8, and CycC in wild-type larval and pupal extracts from the early L3 larvae (84hr AEL [after egg laying]), the late L3 wandering larvae (112hr AEL), at pupariation (0 hr APF [after puparium formation]), and pupae (72 hr APF). The level of EcR-B1 (105 kDa) was low in the early L3 stage, but was significantly increased from the wandering to 72 hr APF stage (Fig 4A), which lags the temporal expression profile of EcR-B mRNA [54,67]. The monoclonal antibody against USP (AB11) recognizes two forms of USP: the 54 kDa full-length USP protein and the 48 kDa truncated USP that lacks the most N-terminal portion [68–70]. The truncated USP is proposed to derive from alternative usage of translation start sites or protease cleavage [68,69,71]. We detected both isoforms of USP, and observed that the levels of both isoforms, particularly the full-length USP, are significantly increased during the pupal stages (Fig 4A). To facilitate biochemical analyses of USP (see below), we generated a polyclonal USP antibody in guinea pig. Similar to the USP monoclonal antibody [72,73], this new polyclonal antibody also specifically recognizes the two isoforms of USP (S5A Fig), and reveals a similar expression pattern of USP during development (Fig 4A). Interestingly, the protein levels of CDK8 and CycC are increased after the L3 wandering stage (Fig 4A).
Because the major changes in EcR, USP, CDK8, and CycC levels occurred during the L3 larval to pupal transition (Fig 4A), we performed our subsequent analyses of cdk8 and cycC mutants at the L3 wandering stage and the white prepupal stage. During the L3 wandering stage, the level of the full-length USP protein was significantly increased in cdk8 and cycC mutants (Fig 4B). The level of EcR-B1 was increased in the mutants, particularly in the cdk8 mutant larvae (Fig 4B). In white prepupae, the level of EcR-B1 was also significantly increased in cdk8 and cycC mutants, but the level of full-length USP was similar to the control (Fig 4C). Thus, the total protein levels of EcR are higher in cdk8 and cycC mutants than in controls during the L3 wandering stage and the white prepupal stage, while the total protein levels of USP are higher in cdk8 and cycC mutants during the L3 wandering stage.
Since the expression of the EcR-USP target genes is reduced in cdk8 and cycC mutants (Fig 2), we did not expect that protein levels of EcR and USP would be increased in mutants at the same stage (Fig 4B and 4C). Thus we examined whether the subcellular distribution of EcR or USP were affected in cdk8 or cycC mutants by performing immunostaining of the salivary glands from the L3 wandering larvae. Both EcR (S6A Fig) and USP (S6A’ Fig) were localized in the nuclei in wild-type salivary glands. In cdk8 and cycC mutant glands, the levels of EcR (S6B and S6C Fig) and USP (S6B’ and S6C’ Fig) in both nucleus and cytoplasm appear to be slightly elevated compared to the control (S6A and S6A’ Fig, respectively), which is supported by quantification of these images using ImageJ (S6D and S6E Fig). These results suggest that the cytoplasmic levels of EcR or USP and nuclear levels of USP were increased in cdk8 and cycC mutants.
Since immunostaining is not a robust quantitative approach, we fractionated nuclear soluble and cytoplasmic fractions of total proteins and analyzed the levels of EcR and USP by Western blot. The full-length USP protein was significantly increased in the nuclear soluble fraction of samples from cdk8 and cycC mutants during the late L3 wandering and WPP stages (Fig 4D, left panel). In addition, the nuclear EcR levels were higher in cdk8, and to a lesser extent, cycC mutants, than the control at the late L3 and WPP stages (Fig 4D). In contrast, when analyzing the cytoplasmic fraction from the early L3 to WPP stage, we observed that USP level was a bit higher in cdk8 and cycC mutants than the control, but there was no obvious difference in cytoplasmic EcR levels (Fig 4D, right panel). Nevertheless, these analyses show that the increased EcR and USP proteins in cdk8 or cycC mutants are predominantly localized in the nuclei during early and late L3 stage, suggesting that the subcellular localization of EcR and USP are not affected in the cdk8 or cycC mutants.
Interestingly, the salivary gland cells in the cdk8 and cycC mutants are smaller than the control of the same stage (compare S6B and S6C with S6A Fig), in addition to weaker DAPI staining (S6B” and S6C” Fig, compared to the control in S6A” Fig). The sizes of salivary gland cells positively correlate to the DNA content [74]. The giant polytene chromosomes are produced from successive rounds of DNA endoreduplication. At the molecular level, DNA endoreduplication is controlled by periodical E2F1-activated expression of cyclin E (cycE) gene followed by transient degradation of E2F1 protein, which is mediated by the CRL4 (CDT2) ubiquitin ligase [75]. Previously, we have reported that CDK8-CycC negatively regulates E2F1 activity in Drosophila [76]. As measured by qRT-PCR, the levels of E2F1 targets genes, such as CG7670, cycE, MCM5, mus209 (encoding PCNA), Orc5, rnrL, and stg, are indeed significantly increased in the cdk8 or cycC mutant salivary glands (S7 Fig). These results suggest that the smaller salivary glands in the cdk8 and cycC mutants are likely caused by dysregulated E2F1 activity and endoreduplication.
An alternative model to explain the apparent discrepancy between the increased protein levels of EcR-USP and the decreased EcR target gene expression in cdk8 or cycC mutants is that the CDK8-CycC complex is required for EcR-USP binding to the promoters of EcR target genes. In this model, the accumulated EcR-USP in nuclei may not effectively stimulate the target gene expression in the absence of CDK8 or CycC. To test this hypothesis, we used the antibodies against EcR or USP to ascertain protein localization on polytene chromosomes, which provide a straightforward method for rapid detection of the genome-wide localization of chromatin-binding proteins [77,78]. In polytene chromosome spreads from wild-type larvae, EcR (Fig 5A) and USP (Fig 5A’) antibodies stain distinct bands that largely overlap with each other. However, we could hardly detect any signal of anti-USP staining on polytene chromosome spreads from cdk8 and cycC mutants that were prepared and imaged under the same conditions (Fig 5B’ and 5C’). Similarly, the signal of the anti-EcR antibody staining was significantly reduced on the polytene chromosome from the cdk8 and cycC mutants (Fig 5B and 5C), compared to the control (Fig 5A). To validate the consequence of this reduction of EcR-USP binding to polytene chromosomes, we examined the expression of EcR-target genes in the cdk8 and cycC mutant salivary glands at the L3 wandering stage using qRT-PCR. Similar to the data from whole-body analysis (Fig 2B), the levels of EcR activated genes were significantly reduced in cdk8 and cycC mutant salivary glands than the control (Fig 5D). These observations suggest that the recruitment of EcR and USP to their target promoters is defective in cdk8 and cycC mutants.
To further validate these observations, we performed chromatin immunoprecipitation (ChIP) assay to examine whether the presence of USP at EcR target gene promoters, such as E74, E75, E78, and Hsp27, was affected in cdk8 and cycC mutant larvae. As shown in Fig 5E, the binding of USP to the promoters of these EcR-USP target genes was diminished in cdk8 and cycC mutants. These data are consistent with the reduced binding of EcR and USP to the polytene spreads in cdk8 and cycC mutants (Fig 5B and 5B’ and 5C and 5C’). Taken together, these observations suggest that the CDK8-CycC complex is required for the recruitment of EcR and USP to their target genes.
To test the possibility that CDK8-CycC interacts with EcR or USP in vivo, we analyzed whether EcR or USP could co-immunoprecipitate with CDK8 in white prepupae. As shown in Fig 6A, EcR-B1 co-immunoprecipitated endogenous CDK8. Similarly, CDK8 was co-immunoprecipitated with USP (Fig 6B). These results suggest that CDK8 can interact with the EcR-USP complex in vivo. To test whether other Mediator subunits can co-immunoprecipitate with EcR and USP, we performed mass spectrometry analysis for proteins that immunoprecipitated with either EcR or USP in wild-type white prepupae. As shown in Fig 6C, multiple Mediator subunits, including the subunits of the CDK8 submodule, Med12 (encoded by kohtalo or kto), Med13 (encoded by skuld or skd [79–81]), CDK8, and CycC, co-immunoprecipitated with EcR and USP. This assay also identified several known cofactors for EcR-USP, such as Hsp70, Taiman, Smr, Rig, dDOR, and Utx (S2 Table). These results validated and significantly expanded our co-immunoprecipitation data (Fig 6A and 6B), suggesting that the Mediator complexes may function as transcriptional cofactors for EcR-USP.
To address whether the interaction between EcR and USP is affected by CDK8-CycC, we tested whether USP could co-immunoprecipitate EcR in cdk8 and cycC mutants as efficient as control during the white prepupal stage. As expected, USP co-immunoprecipitated with EcR-B1 in the control; however, despite the elevated levels of EcR and USP in the mutants (Fig 6D’), much less EcR-B1 could be co-immunoprecipitated with USP in cdk8 or cycC mutants (Fig 6D), consistent with the reduced USP binding to EcR targets in the mutants (Fig 5). This result suggests that CDK8-CycC normally functions to enhance the EcR-USP interaction, which is required for EcR-USP binding to the promoters of EcR target genes.
Many transcriptional cofactors of nuclear receptors are known to possess a conserved signature amino acid motif LXXLL (where L is leucine and X is any amino acid), as the interaction surfaces [65,66]. For example, via this LXXLL motif, the Mediator subunits (MED1 and MED14) and other cofactors interact with mammalian nuclear receptors, such as androgen receptor, estrogen receptor, glucocorticoid receptor, thyroid hormone receptor and RXR [65,66,82,83]. This LXXLL motif is also found in several transcription coactivators for EcR [13], such as Taiman, TRR, Rig, dDOR, dDEK, Hsp90, and Hsc70 [24–29,84]. Interestingly, we have found that CDK8, but not CycC, has a LXXLL motif that is highly conserved from yeasts to human (Fig 6E). The X-ray crystal structure of the human CDK8-CycC complex shows that this leucine-rich motif is localized on the surface of the CDK8 protein [52]. In addition, the two LXXLL motifs in MED1 are vertebrate-specific and they are not present in flies or worms (S8A Fig), while Drosophila Med14 has one LXXLL motif that is conserved from flies to humans but not in worms (Fig 6F).
To test whether EcR or USP may directly interact with CDK8 and Med14, we performed yeast two-hybrid assays. We focused on the EcR-B1 isoform, because cdk8 and cycC mutants resemble EcR-B1 mutants (Fig 1) and EcR-B1 is the major isoform that controls the larval–pupal transition [53]. Similar to other nuclear receptors, EcR and USP contain a ligand-independent activation function (AF1) domain at their N-termini, followed by a DNA-binding domain (DBD) and a ligand-binding domain (LBD) that contains the ligand-dependent activation function (AF2) (Fig 6G) [6,13]. We observed that EcR-AF1, but not EcR-AF2, could directly bind to CDK8 (Fig 6H). In contrast, CDK8 did not bind to either AF1 or AF2 of USP (Fig 6H). Similarly, EcR-AF1, but not EcR-AF2, directly interacted with the fragment of Med14 that contains the LXXLL motif (Fig 6H). One caveat of these analyses is that the ligand 20E is not present in this assay, thus it is possible that the ligand may be required for EcR-AF2 to interact with CDK8, Med14, or other Mediator subunits. Nevertheless, these data suggest that the Mediator complexes are involved in regulating EcR-dependent gene expression through direct interactions between EcR and CDK8 or Med14 (S8B Fig).
Recently, we have reported that CDK8-CycC plays a key role in regulating lipogenesis in Drosophila and mammals by directly inhibiting the transcriptional activity of SREBPs [46]. Since the wandering behavior triggered by a pulse of 20E may mark a fundamental transition in energy metabolism from SREBP-dependent lipogenesis in feeding larvae to lipolysis in nonfeeding pupae, our data showing that CDK8 regulates EcR- and SREBP-dependent transcription prompt us to hypothesize that CDK8-CycC may integrate feeding-stimulated lipogenesis and ecdysone-regulated metamorphosis during the larval–pupal transition (Fig 7A).
To assess the plausibility of this hypothesis, we first analyzed the protein levels of CDK8, CycC, SREBP, EcR and USP from mid-L3 larval stage (92 hr AEL) to WPP stage (120 hr AEL) by Western blot. In our experiments, the larvae started moving out of food approximately104 hr AEL, wandering stage occurred between 108 and 116 hr AEL, and then they reached WPP stage at approximately120 hr AEL. As shown in Fig 7B, the level of CDK8 is significantly increased during the wandering stage, which coincides with the abrupt increase of EcR and USP proteins. In contrast, the protein levels of CycC and SREBP were not significantly altered.
To test whether the levels of EcR-USP and SREBP correlate with the expression of their target genes, we analyzed the expression of their target genes using qRT-PCR. We observed that the mRNA levels of cdk8 and cycC are gradually increased during L3 (Fig 7C and 7D), which is supported by our measurement of their levels from early L3 (84 hr AEL) to pupal stage (72hr APF) (S9A Fig). Although the expression of usp is not significantly increased, the mRNA levels of EcR and EcR-target genes, such as E74, E75, and E78, are significantly increased during the wandering stage (Figs 7E–7H and S9B). In contrast to EcR and EcR target genes, the mRNA levels of SREBP, and particularly SREBP-target genes, such as dFAS, dACC and dACS, are significantly decreased during the wandering and WPP stages (Figs 7I–7K and S9C). Importantly, the patterns of change for SREBP target genes and EcR target genes appear opposite, and the transition occurs during the wandering stage, suggesting that the onset of wandering stage may represent a turning point for the increase of CDK8 and EcR-USP but the opposite trend of SREBP activity during the late L3 stage. The wandering behavior is accompanied by the cessation of feeding, thus the wandering stage may mark the major shift from lipogenesis in feeding larvae to EcR-regulated pupariation. These changes are suggestive and correlative, thus we performed additional experiments to test the relationship between nutrient intake and activities of SREBP and EcR as described below.
Because CDK8-CycC directly regulates the transcriptional activity of both SREBP and EcR-USP, we sought to examine whether CDK8-CycC was actively involved in coordinating lipogenesis and metamorphosis in response to changes in nutrient intake triggered by wandering behavior. Since starvation of feeding larvae prematurely turns off nutrient intake, we asked whether starvation of the feeding larvae could precociously regulate the CDK8-SREBP/EcR network outlined in Fig 7A. Drosophila larvae reach critical weight between 80 hr and 82 hr AEL, and continue feeding for about 20 hr before the onset of wandering stage [85]. Therefore, we starved larvae during the first half of the post-critical weight feeding stage (84–100 hr AEL), and then analyzed levels of CDK8, CycC, SREBP, EcR, and USP by Western blot. As shown in Fig 8A, the levels of CDK8, EcR-B1, and the full-length USP are barely detectable in normal feeding larvae during 84–100 hr AEL, but all of them are significantly increased after 4–8 hr of starvation (88 or 92 hr AEL). In contrast, the level of nuclear SREBP was decreased after 12 hr of starvation (96 hr AEL), while CycC was not significantly affected by starvation (Fig 8A). These data show that starvation indeed leads to precocious reduction of mature form of SREBP and up-regulation of CDK8, EcR, and USP. Interestingly, the mRNA levels of these factors are not significantly affected by starvation (Figs 8B, 8C and 8E, S10A and S10B), suggesting post-transcriptional regulation of CDK8, EcR, USP, and SREBP by starvation.
Next, we examined whether the expression of EcR and SREBP target genes was affected by starvation using qRT-PCR. Consistent to reduced level of mature SREBP protein and increased CDK8, SREBP target genes such as dFAS and dACS are strongly reduced after 8 hr of starvation (Figs 8F and 8G). Although EcR and USP levels are significantly increased after starvation, expression of EcR target genes, such as E74, E75 and E78, is not significantly affected by starvation (Figs 8D, S10C and S10D), suggesting that increase of EcR-USP alone is not sufficient to induce EcR target gene expression. To test whether the ecdysone biosynthesis is affected by starvation, we measured ecdysteroid titer and found no significant effect of starvation on ecdysone biosynthesis during the 84–100 hr AEL (S10E Fig). Although it is unclear whether ecdysone biosynthesis is accelerated by starvation between 100 and 120 hr AEL (see below, Fig 8H), this observation (S10E Fig) may explain why EcR target genes are not induced by elevated EcR-USP in starved larvae during the 16-hr period that we analyzed. Together, these results suggest that starvation precociously up-regulates CDK8-CycC and EcR-USP, but down-regulates SREBP and SREBP activity, all post-transcriptionally.
Furthermore, we analyzed the effect of starvation on the timing of the larval–pupal transition. We observed that starvation of the wild-type larvae after they reached critical weight led to approximately 6 hr earlier onset of pupariation (Fig 8H, red line) and formation of smaller pupae than control (S10F Fig). These observations are consistent to the predicted effects on the CDK8-EcR/SREBP network when nutrient intake is stopped early by starvation (Fig 7A).
Previously, we reported that refeeding of the starved larvae strongly activated the expression of lipogenic genes such as dFAS, while over-expression of CycC in fat body significantly hampered the refeeding-induced dFAS expression [46]. Therefore, to further analyze the effect of nutrition and feeding on the CDK8-EcR-SREBP network, we tested whether refeeding of starved larvae could have opposite effects on the CDK8-EcR/SREBP regulatory network to starvation. Specifically, we starved wild-type larvae at 84 hr AEL for 10 hr, and then collected the refeeding larvae after they were transferred back to normal food for 0, 1, 2, 3, 6, or 9 hr (Fig 8H, blue line; Fig 9A). We observed that refeeding for 1 to 3 hr potently reduced the protein levels of CDK8, EcR and USP (Fig 9B). Except EcR, the mRNA levels of cdk8 and usp are not obviously affected by refeeding (Figs 9D and 9E, and S11B), suggesting a post-transcriptional regulation of these factors by refeeding. Similar observations were made after refeeding for 6 or 9 hr (Fig 9C and S12). Importantly, these changes are opposite to the effect of starvation (Fig 8A), supporting the inhibitory effects of feeding or refeeding on CDK8 (Fig 7A). Although both EcR and USP levels are reduced in refed larvae, expression of EcR-target genes are not obviously affected (S11C–S11E Fig and S12F and S12G Fig). Perhaps, biosynthesis of 20E or other cofactors for EcR-USP dependent transcription are not present in refed larvae in the time window that we analyzed. Indeed, the refed larvae could pupariate, but with approximately 8 hr of delay (Figs 8H and S10F). In addition, we did not observe any obvious changes in the level of mature SREBP proteins, but expression of SREBP and the SREBP target genes were significantly increased in refed larvae (Figs 9F–9H and S12H–S12K), suggesting a potent stimulatory effect of refeeding on SREBP activity. Taken together, these results are largely consistent with the model that CDK8-CycC links the nutrition intake to EcR-USP and the activity of SREBP, suggesting that CDK8-CycC functions as a signaling node for coordinating lipid homeostasis and developmental timing in response to nutrient cues (Fig 7A).
Through EcR-USP, ecdysone plays pivotal roles in controlling developmental timing in Drosophila. In this study, we show that cdk8 or cycC mutants resemble EcR-B1 mutants and CDK8-CycC is required for proper activation of EcR-target genes. Our molecular and biochemical analyses suggest that CDK8-CycC and the Mediator complexes are directly involved in EcR-dependent gene activation. In addition, the protein levels of CDK8 and CycC are up-regulated at the onset of the wandering stage, closely correlated with the activation of EcR-USP and down-regulation of SREBP-dependent lipogenesis during the larval–pupal transition. Remarkably, starvation of the feeding larvae leads to premature up-regulation of CDK8 and EcR-USP, and precocious down-regulation of SREBP, while refeeding of the starved larvae results in opposite effects on the CDK8-SREBP/EcR network. Thus, we propose that CDK8-CycC serves as a key mediator linking food consumption and nutrient intake to EcR-dependent developmental timing and SREBP-dependent lipogenesis during the larval–pupal transition.
The Mediator complex is composed of up to 30 different subunits, and biochemical analyses of the Mediator have identified the small Mediator complex and the large Mediator complex, with the CDK8 submodule being the major difference between the two complexes [38,39,86]. Several reports link EcR and certain subunits of the Mediator complex. For example, Med12 and Med24 were shown to be required for ecdysone-triggered apoptosis in Drosophila salivary glands [87–89]. It was recently reported that ecdysone and multiple Mediator subunits could regulate cell-cycle exit in neuronal stem cells by changing energy metabolism in Drosophila, and specifically, EcR was shown to co-immunoprecipitate with Med27 [90]. However, exactly how Mediator complexes are involved in regulating EcR-dependent transcription remains unknown. Our data suggest that CDK8 and CycC are required for EcR-activated gene expression. Loss of either CDK8 or CycC reduced USP binding to EcR target promoters, diminished EcR target gene expression, and delayed developmental transition, which are reminiscent of EcR-B1 mutants [53]. Importantly, our mass spectrometry analysis for proteins that co-immunoprecipitate with EcR or USP has identified multiple Mediator subunits, including all four subunits of the CDK8 submodule. Taken together, previous works and our present work highlight a critical role of the Mediator complexes including CDK8-CycC in regulating EcR-dependent transcription.
How does CDK8-CycC regulate EcR-activated gene expression? Our biochemical analyses show that CDK8 can interact with EcR and USP in vivo and that CDK8 can directly interact with EcR-AF1. These observations, together with the current understanding of how nuclear receptors and Mediator coordinately regulate transcription, suggest that CDK8-CycC may positively and directly regulate EcR-dependent transcription (S8B Fig). Our yeast two-hybrid analysis indicates that the recruitment of CDK8-CycC to EcR-USP can occur via interactions between CDK8 and the AF1 domain of EcR. Interestingly, this assay also revealed a direct interaction between EcR-AF1 and a fragment of Med14 that contains the LXXLL motif. In future work, it will be interesting to determine whether CDK8 and Med14 compete with each other in binding with the EcR-AF1, whether they interact with EcR-AF1 sequentially in activating EcR-dependent transcription, and how the Mediator complexes coordinate with other known EcR cofactors in regulating EcR-dependent gene expression.
In cdk8 or cycC mutants, the binding of USP to the promoters of the EcR target genes is significantly compromised, even though nuclear protein levels of both EcR and USP are increased. It is unclear how CDK8-CycC positively regulates EcR-USP binding to EcREs near promoters. CDK8 can directly phosphorylate a number of transcription factors, such as Notch intracellular domain, E2F1, SMADs, SREBP, STAT1, and p53 [42,43,46]. Interestingly, the endogenous EcR and USP are phosphorylated at multiple serine residues, and treatment with 20E enhances the phosphorylation of USP [70,91,92]. Protein kinase C has also been proposed to phosphorylate USP [93,94]. It will be interesting to determine whether CDK8 can also directly phosphorylate either EcR or USP, thereby potentiating expression of EcR target genes and integrating signals from multiple signaling pathways.
Although we favor a direct role for CDK8-CycC to regulate EcR-USP activated gene expression, we could not exclude the potential contribution of impaired biosynthesis of 20E to the developmental defects in cdk8 or cycC mutants. For example, the expression of genes involved in synthesis of 20E, such as sad and spok, is significantly reduced in cdk8 or cycC mutant larvae. Indeed, the ecdysteroid titer is significantly lower in cdk8 mutants than control from the early L3 to the WPP stages, and feeding the cdk8 mutant larvae with 20E can partially reduce the retardation in pupariation. Nevertheless, impaired ecdysone biosynthesis alone cannot explain developmental defects that we characterized in this report for the following reasons. First, feeding cdk8 or cycC mutants with 20E cannot rescue the defects in pupal morphology, developmental delay, and the onset of pupariation. Second, the expression of EcRE-lacZ reporter in cdk8 or cycC mutant salivary glands cannot be as effectively stimulated by 20E treatment as in control. Third, knocking down of either cdk8 or cycC in PG did not lead to obvious defects in developmental timing. Therefore, the most likely scenario is that the cdk8 or cycC mutants are impaired not only in 20E biosynthesis in the PG, but also in EcR-activated gene expression in peripheral tissues. Defects in either ecdysone biosynthesis or EcR transcriptional activity will generate the same outcome: diminished expression of the EcR target genes, thereby delayed onset of pupariation.
How CDK8-CycC regulates biosynthesis of ecdysone in PG remains unknown. Several signaling pathways have been proposed to regulate ecdysone biosynthesis in Drosophila PG, including PTTH and Drosophila insulin-like peptides (dILPs)-activated receptor tyrosine kinase pathway and Activins/TGFβ signaling pathway [95,96]. Interestingly, CDK8 has been reported to regulate the transcriptional activity of SMADs, transcription factors downstream of the TGFβ signaling pathway, in both Drosophila and mammalian cells [97,98]. Thus, it is conceivable that the effect of cdk8 or cycC mutation on ecdysone biosynthesis may due to dysregulated TGFβ signaling in the PG.
Our effort to explore the potential role of food consumption and nutrient intake on CDK8-CycC has resulted an unexpected observation that the protein level of CDK8 is strongly influenced by starvation and refeeding: starvation potently increased CDK8 level, while refeeding has opposite effect, and both occur post-transcriptionally (Figs 8 and 9). The importance of this observation is highlighted in two aspects. First, considering the generally repressive role of CDK8 on Pol II-dependent gene expression, up-regulation of CDK8 may provide an efficient way to quickly tune down most of the Pol II-dependent transcription in response to starvation; while down-regulation of CDK8 in response to refeeding may allow many genes to express when nutrients are abundant. Second, it will be necessary to test whether the effects of nutrient intake on CDK8-CycC is conserved in mammals. If so, considering that both CDK8 and CycC are dysregulated in a variety of human cancers [43], the effects of nutrient intake on CDK8 may have important implications in not only our understanding of the effects of nutrients on tumorigenesis, but also providing nutritional guidance for patients with cancer.
Major dietary components including carbohydrates, lipids, and proteins, can strongly influence the developmental timing in Drosophila [2]. Excessive dietary carbohydrates repress growth and potently retard the onset of pupariation [99–101]. One elegant model proposed to explain how high sugar diet delays developmental timing is that high sugar diet reduces the activity of the Target of Rapamycin (TOR) in the PG, thereby reducing the secretion of ecdysone and delaying the developmental transition [102]. Previously, we reported that insulin signaling could down-regulate CDK8-CycC, and that ectopic expression of CycC could antagonize the effect of insulin stimulation in mammalian cells, as well as the effect of refeeding on the expression of dFAS in Drosophila [46]. Although the mRNA levels of TOR and insulin receptor (InR) are not significantly affected in cdk8 or cycC mutants (Fig 3B), it is necessary to further study whether and how different dietary components may regulate CDK8-CycC in the future.
Our developmental genetic analyses of the cdk8 and cycC mutants have revealed major defects in fat metabolism and developmental timing ([46]; this work). De novo lipogenesis, which is stimulated by insulin signaling, contributes significantly to the rapid increase of body mass during the constant feeding larval stage. This process is terminated by pulses of ecdysone that trigger the wandering behavior at the end of the L3 stage, followed by the onset of the pupariation. Insulin and ecdysone signaling are known to antagonize each other, and together determine body size of Drosophila. The genetic interaction is established, but the detailed molecular mechanisms are not [1,25,103,104]. The SREBP family of transcription factors controls the expression of lipogenic enzymes in metazoans and the expression of cholesterogenic enzymes in vertebrates [105,106]. Our previous work shows that CDK8 directly phosphorylates the nuclear SREBP proteins on a conserved threonine residue and promotes the degradation of nuclear SREBP proteins [46]. Consistent with the lipogenic role of SREBP and the inhibitory role of insulin to CDK8-CycC [46], the transcriptional activity of SREBP is high while the levels of CDK8-CycC and EcR-USP are low prior to the onset of wandering stage. Subsequently during the wandering and non-mobile, non-feeding pupal stage, the transcriptional activity of SREBP is dramatically reduced, accompanied by the significant accumulation of CDK8-CycC and EcR-USP (Fig 7).
The causal relationship of these phenomena was further tested by our starvation and refeeding experiments. On the one hand, we observed that the levels of CDK8, EcR and USP are potently induced by starvation, while the mature SREBP level and the transcriptional activity of SREBP are reduced by starvation (Fig 8). Starvation of larvae prior to the two nutritional checkpoints in early L3, known as minimum viable weight and critical weight, which are reached almost simultaneously in Drosophila, will lead to larval lethality; while starvation after larvae reach the critical weight will lead to early onset of pupariation and formation of small pupae [9,85,107,108]. Thus, this nutritional checkpoint ensures the larvae have accumulated sufficient growth before metamorphosis initiation [2,85]. If we regard the status with high CDK8, EcR, and USP as an older or later stage, these results indicate that starvation shifts the regulatory network precociously, which is consistent with the regulatory network outlined in Fig 7A and the observed premature pupariation (Fig 8H). On the other hand, our analyses of refed larvae show that refeeding potently reduced the levels of CDK8, EcR and USP (Fig 9). If we consider the status with low CDK8, EcR, and USP as a younger or earlier stage, these results indicate that refeeding delays the activation of this network, which is consistent with our model (Fig 7A) and delayed pupariation as observed (Fig 8H). Taken together, our results based on starved and refed larvae suggest that CDK8-CycC is a key regulatory node linking nutritional cues with de novo lipogenesis and developmental timing (Fig 7A).
The larval–pupal transition is complex and dynamic. Although the expression of SREBP target genes fit well with the predicted effects of starvation and refeeding, the expression of EcR targets during the stage that we analyzed does not reflect the changes in the protein levels of EcR and USP (Figs 8 and 9). It is reasonable to consider that CDK8-CycC and EcR-USP are necessary, but not sufficient, for the activation of EcR target genes. One possibility is that there is a delay on synthesis of 20E or other cofactors that are required for EcR-activated gene expression in response to starvation. Indeed, we measured the 20E levels during the first 16 hr of starvation and observed no significant difference between fed and starved larvae (S10E Fig). It will be necessary to further analyze the effect of starvation on 20E synthesis at later time points in the future.
Taken together, we propose a model whereby CDK8-CycC functions as a regulatory node that coordinates de novo lipogenesis during larval stage and EcR-dependent pupariation in response to nutritional cues (Fig 7A). It is likely that pulses of 20E synthesized in the PG, and subsequent behavioral change from feeding to wandering, ultimately trigger the transition from SREBP-dependent lipogenesis to EcR-dependent pupariation. The opposite effects of CDK8-CycC on SREBP- and EcR-dependent gene expression suggest that the role of CDK8 on transcription is context-dependent.
In conclusion, our study illustrates how CDK8-CycC regulates EcR-USP-dependent gene expression, and our results suggest that CDK8-CycC may function as a regulatory node linking fat metabolism and developmental timing with nutritional cues during Drosophila development.
The null alleles of cdk8 (cdk8K185) and cycC (cycCY5) strains were provided by Drs. Muriel Boube and Henri-Marc Bourbon [50]. The EcRE-lacZ reporter and ubi-Gal4 lines were obtained from Dr. Keith Maggert. The P[hs-usp] transgenic line [72,73] was obtained from the Bloomington Drosophila stock center. Embryos from cycCY5 germline clones were generated using the Flipase recombinase-mediated dominant female sterile technique [109]. All flies were maintained on standard cornmeal-molasses-yeast medium at 25°C.
The anti-USP monoclonal antibody was provided by Dr. Rosa Barrio Olano. The anti-CycC polyclonal antiserum (peptide antibody in rabbits) was provided by Dr. Terry Orr-Weaver. Anti-EcR common (DDA2.7) and anti-EcR-B1 (AD4.4) monoclonal antibodies were obtained from Developmental Studies Hybridoma Bank, and the anti-actin (MA5-11869) monoclonal antibody was purchased from Thermo Scientific (Rockford, IL). The anti-CDK8 polyclonal antibody was generated by immunizing rabbits using peptide AA355~372 (KREFLTDDDQEDKSDNKR) as the antigen, anti-SREBP polyclonal antibody was generated using peptide AA360~378 (KDLLQLGTRPGRASKKRRE) as the antigen, and both were performed by Thermo Scientific. The antisera were purified by GST-CDK8 (AA1~372) or GST-SREBP (AA1~451) fusion proteins, respectively, using the protocol as described previously [110]. The anti-USP polyclonal antibody was generated by immunizing guinea pigs with GST-USP (full length) as the antigen, performed by Covance Research Products (Denver, PA). These fusion proteins were generated using the protocol described previously [111].
We generated the tagged genomic cdk8 or cycC (approximately 7.5-kb) rescue constructs using backbone of the pVALIUM20 vector, which can be used for site-specific insertion with the PhiC31 integrase system [112]. For subcloning, we first linearized the pVALIUM20-gypsy-MSC10 vector by EcoRI (NEB). The gDNA segments for cdk8 and cycC were PCR amplified from bacterial artificial chromosome (BAC) clones (CH322-104A8 for cdk8 locus and CH321-46N21 for cycC locus) from the BACPAC Resources Center (http://bacpac.chori.org/home.htm). To ensure the fidelity of these PCR reactions, we used a high-fidelity DNA polymerase PrimeSTAR Max (Takara, Cat# R045A) and then purified all segments by gel extraction (QIAEX II). To join four DNA segments (pVALIUM20 backbone, two gDNA segments and one EGFP segment) seamlessly in a single reaction, we used In-Fusion HD system developed by Clontech (639649). This system requires that the sense and antisense PCR primers contain a 15bp overlap with the neighboring segment and 20–30bp segment specific sequence.
The primers with the 15bp overlapping sequence underlined are listed below: Cdk8 IN-1L: 5′-GTGGCTAGCAGAATTCAGGCACCCATTGGCGATG; Cdk8 IN-2: 5′-GTTGAAGCGCTGGAAGTTCTGCT; Cdk8 IN-3(EGFP): 5′-TTCCAGCGCTTCAACATGGTGAGCAAGGGCGAGGAG; Cdk8 IN-4(EGFP): 5′-TGTATCAGTCTCTCACTTGTACAGCTCGTCCATGCCG; Cdk8 IN-5: 5′- TGAGAGACTGATACATGCAGCATTTTTTC; Cdk8 IN-6LL: 5′- GGCTCTAGATGAATTATGCTCGCTGATTCCACGATCAG; CycC IN-1L: 5′- GTGGCTAGCAGAATTTCCTTCGAGGATCGCACCTG; CycC IN-2: 5′-ACGCTGAGGCGGTGGTTTC; CycC IN-3(EGFP-ATG): 5′-ATGCCACCGCCTCAGCGTGTGAGCAAGGGCGAGGAGCTG; CycC IN-4(EGFP): 5′-TATGAAGCTCTTCTACTTGTACAGCTCGTCCATGCCG; CycC IN-5: 5′-TAGAAGAGCTTCATAATCATTCATCATTAGC; and CycC IN-6L: 5′-GGCTCTAGATGAATTTGCTGGACCTATACAGACGCACG.
For the In-Fusion reaction, 100 ng of enzyme-digested, gel-purified vector were mixed with the PCR amplified segments at a molar ratio of 1 vector to 2 of each DNA segment in a total of 10 μl system buffered by In-Fusion HD Enzyme premix and the subsequent steps were carried out following the manufacturer’s instructions. The positive clones were selected and characterized by restriction enzyme digestion and sequencing. The rescue constructs were inserted into the second chromosome (attP40 site at 25C6) with the service provided by Genetic Services, Inc. This design facilitates genetic recombination since the endogenous cdk8 and cycC genes are on the third chromosome.
The microarray analyses were described previously [46], and the data sets can be found in the ArrayExpress database (http://www.ebi.ac.uk/arrayexpress/; accession number E-MTAB-1066).
The RNA isolation, reverse transcription, the qRT-PCR analyses, and primers for the lipogenic enzymes were performed as described previously [46]. The primers used in the qRT-PCR assay are listed in S3 Table, and Rp49 gene was used as the control. Primers for InR, kni, mld, nvd, tor, and vvl are adapted from [61].
Quantification of ecdysteroids in whole larvae was performed as described by [61,113] with the following modifications. Briefly, animals were homogenized in 0.25 ml 75% methanol, and then the supernatants were collected following centrifugation at 14,000 g for 15 min. The pellets were re-extracted in 0.1 ml methanol. The supernatants were combined, evaporated using a SpeedVac, and then re-dissolved in 0.5 ml ELISA buffer (Cayman Chemical). Ecdysteroids were measured using a commercial ELISA kit (Cayman Chemical) that detects 20E equivalents. Standard curves were generated using 20E (Cayman Chemical), and absorbance was measured at 405 nm on a microplate photometer (Thermo Scientific).
The rescue experiments were performed as described previously [63]. Briefly, cdk8K185 and cycCY5 homozygous mutants at late L2 and early L3 larvae were collected and placed in groups of 10 individuals in new vials containing food with 20E (Alexis or Cayman Chemical) ranging from 2.0 μM to 2.0 mM (Figs 3 and S4), and w1118 larvae were treated in parallel as the control. Pupae were collected and photographed under a microscope.
The EcRE-lacZ reporter line was recombined with cdk8K185 or cycCY5 mutant to generate the following genotypes: “w1118; EcRE-lacZ; cdk8K185/TM6B”, “w1118; EcRE-lacZ; cycCY5/TM6B”. The “w1118; EcRE-lacZ; +” line was used as the control. To ensure that we compare the salivary glands that are at the same developmental stage, we dissected the salivary gland from mid-L3 homozygous larvae (non-TM6B) of all these genotypes, and separated into two halves in Grace’s insect medium (HiMedia). One half was treated in Grace’s medium with 1μM 20E, while the other half from the same larva was cultured in Grace’s insect medium as the control. After 2.5 hr incubation at 25 ˚C, salivary glands were stained with X-gal solution (3.0 mM K4[Fe(CN)6], 3.0 mM K3[Fe(CN)6] in PBS with X-gal stock solution (8% in DMSO) added to a final concentration of 0.2%) at 37°C for 1 hr in the dark. The stained salivary glands were then transferred into 80% glycerol in PBS, mounted and photographed with a Leica DM2500 microscope.
The salivary glands from the third-instar wandering larvae were dissected in PBS (phosphate buffered saline, pH7.4) and fixed in 5% formaldehyde in PBS for 10 min. After washing in PBT (PBS with 0.2% Triton X-100, pH7.4) for 4 times for 1 hour, the salivary glands were blocked in PBTB (0.2% BSA, 5% normal goat serum in PBT) for 1 hour at room temperature. The glands were then incubated with the anti-EcR-common DDA2.7 antibody (1:100) and anti-USP antibody (1:2,000, polyclonal antibody from guinea pigs) at 4 ˚C overnight on a nutator. After rinsing with PBT for 4 times, the glands were incubated with secondary antibodies (BODIPY-conjugated goat anti-mouse antibody 1:500 in PBTB; Alex594-conjugated goat anti-guinea pig antibody, 1:2,500 in PBTB) at room temperature for 2 hr. After standard nuclear counterstaining with DAPI (4′,6-Diamidino-2-phenylindole dihydrochloride, Sigma), the salivary glands were mounted on slides with Vectashield mounting media (Vector lab). For immunostaining of polytene chromosome, we followed the protocol described previously [114], and the following antibodies were used: anti-EcR-common DDA2.7 (1:200 in PBTB), anti-USP (Guinea pig, 1:200), BODIPY-conjugated goat anti-mouse antibody (1:500), and Alex594-conjugated goat anti-guinea pig antibody (1:1,000). Confocal images were taken with a Nikon Ti Eclipse microscope, and images were processed by Adobe Photoshop CS6 software.
We separated the cytoplasmic, nuclear soluble, and nuclear insoluble fractions of protein extractions by following the protocol as described [115]. Western blot analysis was performed as previously described with minor modifications [46]. For whole cell extracts, homogenized larvae or pupae were lysed in a buffer containing 50 mM Tris HCl (pH 8.0), 0.1 mM EDTA, 420 mM NaCl, 0.5% NP-40, 10% glycerol, 1 mM dithiothreitol (DTT), 2.5 mM phenylmethanesulfonylfluoride (PMSF), protease and phosphatase inhibitors (the cOmplete Protease Inhibitor Cocktails and PhosSTOP, Roche Applied Science). Supernatants were collected after centrifugation at 2,000 g for 15 min at 4°C. Protein concentrations were measured with a Bradford protein assay kit (Bio-Rad). A given amount of whole cell extract was mixed with 4x Laemmli sample buffer (Bio-Rad). After boiling for 5 min, the proteins were resolved by 8% SDS-PAGE gel and transferred to PVDF membrane. The following antibodies were used: anti-EcR-common DDA2.7 (1:250), anti-USP (guinea pig, 1:2,000), anti-USP (monoclonal antibody, 1:1,000), anti-CDK8 (polyclonal antibody from rabbit, 1:50), anti-CycC (rabbit polyclonal antibody, 1:2,000), anti-dSREBP (rabbit polyclonal antibody, 1:100), and anti-actin monoclonal antibody (1:4,000, Thermo Scientific). The membranes were incubated with the corresponding HRP-conjugated secondary antibodies (1:2,500–1:10,000, Jackson ImmunoResearch) for 1 hr at room temperature. After washing, the HRP signals were visualized by the Western Lightening Plus ECL (PerkinElmer) according to the manufacturer’s instructions.
The co-IP assay was performed as described previously with minor modifications [116]. Briefly, the IP complex was prepared with 35 μL Magnetic Protein G beads (28-9670-66, GE Healthcare Life Sciences) and 5 μg primary antibody or IgG in 500 μL PBS and put on the rotator for 12–16 hr at 4°C. After incubation, the IP complex was washed with PBS twice and eventually removed all PBS. Lysates of 30 white prepupae per sample were prepared in the lysis buffer (150 mM NaCl, 50 mM Tris pH 8.0, 5 mM EDTA, 5 mM DTT, 0.1 mM PMSF, 0.5% NP-40, 2 mM Na3VO4, and protease inhibitor from Roche Applied Science). 400 μL of lysates were pre-cleared with 20 μL Magnetic Protein G beads on a rotator for 1 hr at 4°C, then the beads were discarded, and the lysates were mixed with IP complex and put on the rotator for 12–16 hr at 4°C. The IP complex was washed with lysis buffer (without protease inhibitor) five times, added 60 μL 2X sample buffer, denatured for 3 min at 95°C, and further analyzed by Western blot.
Each IP sample was run as a gel plug and proteins in the gel plug were reduced, carboxymethylated, digested with trypsin using standard protocols. Peptides were solubilized in 0.1% trifluoroacetic acid, and analyzed by Nano LC-MS/MS (Dionex Ultimate 3000 RLSCnano System interfaced with a Velos-LTQ-Orbitrap (ThermoFisher, San Jose, CA). Sample was loaded onto a self-packed 100 μm x 2 cm trap (Magic C18AQ, 5 μm 200 Å, Michrom Bioresources, Inc.) and washed with Buffer A (0.2% formic acid) for 5 min with a flow rate of 10 μl/min. The trap was brought in-line with the analytical column (Magic C18AQ, 3 μm 200 Å, 75 μm x 50 cm) and peptides fractionated at 300 nL/min using a segmented linear gradient: 4%–15% B (0.2% formic acid in acetonitrile) in 35 min, 15%–25% B in 65 min, 25%–50% B in 55 min. Mass spectrometry data was acquired using a data-dependent acquisition procedure with a cyclic series of a full scan acquired in Orbitrap with resolution of 60,000 followed by MS/MS (acquired in the linear ion trap) of the 20 most intense ions with a repeat count of two and a dynamic exclusion duration of 30 sec.
Peak lists in the format of MASCOT Generic Format (MGF) was generated using the Proteome Discover 1.4 (ThermoFisher). Data were searched against latest flybase Drosophila melanogaster protein database (madmel-all-translation-r6.03.fasta) using a local version of the Global Proteome Machine (GPM) XE Manager version 2.2.1 (Beavis Informatics Ltd., Winnipeg, Canada) with X!Tandem SLEDGEHAMMER (2013.09.01) to assign spectral data [117,118]. Precursor ion mass error tolerance was set to ±10 ppm and fragment mass error tolerance set to ±0.4 Da. Cysteine carbamidomethylation was set as a complete modification, methionine oxidation and deamidation at asparagine and glutamine residues were set as variable modifications. All LC-MS data were analyzed together in a MudPit analysis and individual data extracted to ensure that peptides that could be assigned to more than one protein were assigned consistently for all samples. The resulting identifications were filtered by peptide log GPM expectancy score (log(e)<-1.5).
Whole-animal extracts prepared from L3 wandering larvae or white prepupae of w1118 (control), cdk8K185, or cycCY5 homozygous mutants were used for ChIP assays according to the protocols described previously [119]. Briefly, the fixed materials were sonicated using an Ultrasonic Processor Cell Disruptor (Branson S-450D) at 50% power output for 60 sec (2-sec-long pulse with 1 minute interval on ice). We prepared triplicate biological samples for each genotype, and for each sample, we used 2.0 μg Guinea Pig anti-USP for IP or 2.0μg normal serum isolated from the same Guinea Pig before immunization as a control. SYBR Green PCR Master Mix (Invitrogen) was used in qPCR reactions. For the qRT-PCR, the following primers were designed using Primer Express (Applied Biosystems) based on the EcR ChIP-Seq data (Kevin White’s lab): Hsp27 F 5′-GCAACAAACAAAAGAACGGC-3′, Hsp27 R 5′-TTTCAGAGTGCAACAGAGCTTG-3′; E74 F 5′-TCGGTCAAAAGCAGAGTTCACA-3′, R 5′-ATTTCTCTGCAACTGCTCCC-3′; E75 F 5′-AGGCCTGGCTGGCTGTTACTTA-3′, R 5′-CGGAGAGTTGAAGGCGAGTTT-3′; and E78 F 5′-ATGACGTTGCCCACAAGTCATT-3′, R 5′-ACAGTTGCCTTGGCTTCTTCG-3′. Fold enrichment was calculated and normalized using the Guinea Pig normal serum as the negative control.
To investigate physically interactions between EcR/USP and CDK8/Med14 in yeast cells, we have cloned EcR/USP into pGBKT7 (bait vector) and Drosophila Med14/CDK8 into pGADT7 (prey vector) (Clontech, Mountain View, CA) using the procedures described previously [120]. The following primers were used: for EcRB1-AF1 BamHIS 5′-CAGGATCCTGAAGCGGCGCTGGTCGAAC-3′ and EcRB1-AF1PstIAS 5′-CACTGCAGACCTGAAGATATAGAATTCACCGAATCGC-3′; for EcRB1-AF2 BamHIS 5′-CAGGATCCCTGATGAAATATTGGCCAAGTGTCAAGC-3′ and EcRB1-AF2 PstIAS 5′-CACTGCAGGATGGCATGAACGTCCCAGATCTC-3′; for USP-AF1 (1-100AA): dUSP-1EcoRIS 5′-CAGAATTCATGGACAACTGCGACCAGGACGC-3′ and dUSP-100BamHIAS 5′-CAGGATCCGCTGCCGCTCAGCGGATGGTT-3′; for USP-AF2 (206-509AA): dUSP-206EcoRIS 5′-CAGAATTCAGCTCTCAAGGCGGAGGAGGAGGA-3′ and dUSP-509BamHIAS 5′-CAGGATCCCTACTCCAGTTTCATCGCCAGGCC-3′; for Med14 (159-320AA): dMed14-159EcoRIS 5′-CAGAATTCCTAATTGTACATACTGTCTACATACGATCGG-3′ and dMed14-320BamHIAS 5′-CAGGATCCGGTAGTCTTTTCCTTGAGTTGAACCAC-3′.
At least three independent biological repeats were included for each genotype, all error bars indicate standard deviation, and t-tests were performed using Microsoft Excel (S1 Data). Statistical significance was shown in figures, and fold changes of 1.5 or greater were considered as biologically significant. All biochemical analyses were repeated at least three times and the representative results were shown.
|
10.1371/journal.pcbi.1000145 | Intramolecular Cohesion of Coils Mediated by Phenylalanine–Glycine Motifs in the Natively Unfolded Domain of a Nucleoporin | The nuclear pore complex (NPC) provides the sole aqueous conduit for macromolecular exchange between the nucleus and the cytoplasm of cells. Its diffusion conduit contains a size-selective gate formed by a family of NPC proteins that feature large, natively unfolded domains with phenylalanine–glycine repeats (FG domains). These domains of nucleoporins play key roles in establishing the NPC permeability barrier, but little is known about their dynamic structure. Here we used molecular modeling and biophysical techniques to characterize the dynamic ensemble of structures of a representative FG domain from the yeast nucleoporin Nup116. The results showed that its FG motifs function as intramolecular cohesion elements that impart order to the FG domain and compact its ensemble of structures into native premolten globular configurations. At the NPC, the FG motifs of nucleoporins may exert this cohesive effect intermolecularly as well as intramolecularly to form a malleable yet cohesive quaternary structure composed of highly flexible polypeptide chains. Dynamic shifts in the equilibrium or competition between intra- and intermolecular FG motif interactions could facilitate the rapid and reversible structural transitions at the NPC conduit needed to accommodate passing karyopherin–cargo complexes of various shapes and sizes while simultaneously maintaining a size-selective gate against protein diffusion.
| The nuclear pore complex is a molecular filter that gates macromolecular exchange between the cytoplasm and the nucleoplasm of cells. It contains a size-selective diffusion barrier at its center composed of proteins named FG nucleoporins. These nucleoporins feature large, structurally disordered domains that are highly decorated with phenylalanine–glycine (FG) sequence motifs. The dynamic structure of these disordered FG domains excludes them from classical structural biology analyses such as X-ray crystallography; thus, new approaches are needed to characterize their shape. Here computational and biophysical approaches were used to elucidate the ensemble of structures adopted by the FG domain of a nucleoporin. The analyses showed that the FG motifs function as intramolecular cohesion elements that compact the shape of the FG domain, forcing it to adopt loosely knit globular configurations that are constantly reconfiguring. Within the nuclear pore complex, dozens of these nucleoporin FG domains may stack as loosely knit globules forming a porous sieve that gates molecular diffusion by size exclusion.
| The nuclear pore complex is a supramolecular protein structure in the nuclear envelope that controls nucleo-cytoplasmic traffic and communication (Figure 1A) [1]. A key NPC architectural feature is a poorly understood semi-permeable diffusion barrier at its center, which allows passive diffusion of particles less than 3–4 nm in diameter (or 30–40 kDa in mass for a folded protein) and opens to allow facilitated transport of larger particles up to 39 nm in diameter [2]. The NPC is composed of ∼30 proteins or nucleoporins (nups) that are present in multiple copies [3],[4]. Among these, a group that contains numerous phenylalanine-glycine repeats (FG nups) (a subset is shown in Figure 1B) line the transport conduit of the NPC (Figure 1A). These FG nups function as stepping-stones for karyopherin movement across the NPC [5],[6] and as structural elements of the NPC protein diffusion barrier [7],[8].
The three dimensional structure of S. cerevisiae FG nups is unusual because their 150–700 amino acid (AA) FG domains are natively unfolded [9] in their functional state [6]. Since there are ∼150 FG nups in each NPC [4], it is currently hypothesized that its transport conduit is lined and/or flanked by 150 natively unfolded FG domains. Together these FG domains constitute ∼12% of the total NPC mass or >6.5 MDa of its ∼55 MDa structure in yeast [10]. The FG domains of nups were initially hypothesized to function as repulsive entropic bristles that create a virtual gate at the NPC periphery [11],[12], and later as cohesive polypeptide chains that form a hydrogel at the NPC center [8],[13],[14]. More recently, an analysis of all nup FG domains in S. cerevisiae indicated that some FG domains (the GLFG-rich domains) bind to each other weakly via hydrophobic attractions between their FG motifs, whereas other FG domains (the FxFG-rich domains) do not form such cohesions [7]. Despite the fact that different subtypes of FG domains are defined by their content of FxFG, GLFG or SAFGxPSFG motifs, their ability to interact with each other (i.e., their cohesiveness) seems to correlate best with the AA composition of the sequences between FG motifs, rather than with the specific FG motif [7]. Hence, the human FG nups may also interact with each other, despite having only one GLFG-rich nup among its eleven members [3].
It is generally assumed that natively unfolded proteins have some preferred 3-D structures dictated by intra-molecular cohesion [15],[16]. Current evidence that the FG domains of nups have some structure is based on CD and FTIR spectroscopic analysis, which indicates that FG domains have anywhere from 5% to 20% α-helical and β-sheet content at any given moment [9], yet the locations of such structures in the protein are probably ever-changing. The conformational flexibility inherent to natively unfolded proteins and protein domains such as those in the FG nups, places them beyond the reach of classical structural biology tools such as X-ray crystallography and homology-based computational methods [17]–[20]. However, it is clear that these and other unfolded proteins participate in a wide range of key cell biological processes [21]–[23] and that their native plasticity bestows specific functional properties, such as rapid molecular interaction times and the ability to bind multiple proteins simultaneously [24]. In the case of nucleoporin FG domains, a key function is to bind multiple karyopherins [6] with very rapid interaction times [25]. Thus, in contrast to folded proteins, the structure of natively unfolded proteins must be described as a dynamic ensemble of interconverting conformers.
Since traditional experimental methods for elucidating protein structure cannot be used with natively-unfolded proteins, new approaches are needed to study and describe their dynamic ensemble of structures. In this emerging area of research, Jha et al [26],[27] have recently introduced a general statistical coil model, and Bernado et al [28],[29] have estimated the nuclear magnetic resonance (NMR) measured residual dipolar couplings (RDCs) [30] from dynamic simulations to characterize the ensemble-averaged conformations of α-synuclein. Also, Ollerenshaw et al [31] have applied a native-centric topological model to understand the essential folding/unfolding dynamics SH3 domains, and Pappu and co-workers have characterized poly-glutamines as a function of chain-length conformational sampling by molecular dynamics (MD) and Monte Carlo simulations [32]. Most of these computational investigations suggest the existence of a preferred ensemble of conformers for each protein, rather than suggesting pure random coils.
Here we conducted molecular dynamics simulations and biophysical measurements on a small FG domain from the yeast nucleoporin Nup116 (Q02630) to test the hypothesis that phenylalanines in its FG motifs function as intramolecular cohesion elements that impart structure. Apart from its cell biological significance, we chose this protein as a model system to investigate how a combination of molecular dynamics simulations and biophysical measurements can be used to characterize the ensemble of structures adopted by a natively unfolded protein, such as the FG domain of a nucleoporin.
In the analysis that follows we first used MD simulations to generate a statistical ensemble of coil conformations for a 111 AA region of the Nup116 FG domain containing ten FG motifs (wild-type), and of a mutant version thereof lacking the phenylalanines in the ten FG motifs (F>A mutant) (Figure 1B and 1C). The MD trajectories were then analyzed to evaluate the degree of secondary structure, the overall dimensions of the protein conformations, and the contribution of the FG motifs to the intramolecular cohesion of coils (i.e., compaction) in the dynamic ensemble of nup structures. The simulated FG domains were then expressed in bacteria, purified to homogeneity, and analyzed by NMR spectroscopy and sizing columns to quantify their average shape through measurements of diffusion coefficient and Stokes radii. Finally, mathematical and biophysical analyses were combined to estimate the tertiary structure that best describes the natively unfolded domain of the representative FG nucleoporin.
Twenty independent MD simulations were performed at 300 K (25°C) on the wild-type (6 ns) and F>A mutant (5 ns) versions of a Nup116 FG domain (AA 348–458) starting from a fully-extended conformation. The goal of these simulations was to sample the conformational distribution of the proteins as close as possible to their native distribution in solution. As soon as the simulations started, within the first 100 ps, the extended FG domains collapsed into a more cohesive or compact ensemble of structures with small patches of unstable (see below) secondary structure. Since the wild-type and mutant FG domains are highly flexible and disordered, the resulting end-structures from each of the twenty simulations did not resemble one another as expected for natively unfolded proteins (see Figure 1C for representative examples). Despite the fact that the nup structures were ever-changing (see below), the ensemble of structures for each did “converge” to a similar size early in the simulation according to various metrics of size, which changed little in the last 3 ns. This was evidenced by a constant radius of gyration (Figure S1) and by statistical analyses that showed no significant change in the range of Rg values during the last 3 ns (data not shown).
To describe quantitatively the structural dynamics of the FG domains, we calculated the auto-correlation function of a vector of the 118 Φ and 118 Ψ angles along the peptide backbone of FG domain structures sampled every 1 ps from the MD trajectory. Figure S2 shows the autocorrelation functions with a 200 ps window from the final 3 ns of simulation of all twenty wild-type and F>A mutant FG domain simulations, along with the comparable auto-correlation function from the MD simulation of a control protein that is folded (fibroblast growth factor 1). For each of the replicate nup simulations, the correlation in the Φ–Ψ angles dropped from 1 to 0.738 (±0.031) or 0.741 (±0.023) in 1 ps for the wild-type or mutant FG domains, respectively, and then slowly decayed over 200 ps to 0.672 (±0.037) or 0.665 (±0.029), respectively. In contrast, for the control protein fibroblast growth factor 1, the autocorrelation function dropped to only 0.875 in 1 ps, and then to 0.864 over the 200 ps auto-correlation window. These results indicate that the AA chain backbone of wild-type and mutant FG domains is constantly changing structure and is highly dynamic in comparison to a folded protein.
The ensemble of structures for each of the twenty MD trajectories generated for the wild-type and mutant FG domains were sampled at 1 ps intervals during the final 3 ns of the simulations, yielding a total of 60,000 structures for each protein. The secondary structure content was then analyzed in detail to determine the fraction of time during the simulations that each AA residue spent as part of a “helical” structure (either an α-helix or a 310-helix). In general, no significant difference in overall helical content between wild-type and mutant FG domains was observed. The alpha- and 310- helical structures that did form ranged in size from 2–6 AA residues and did not persist for more than 35 ps on average (data not shown). The maximum duration of an α-helix and a 310-helix was 97 and 699 ps, respectively (data not shown).
Using the same set of 60,000 structures, two measures of protein compactness were calculated: the radius of gyration (Rg) and the end-to-end distance between terminal residues. The average (±1 standard deviation) end-to-end distance for the wild-type FG domain simulated at 300 K was 20.42 Å (±9.51), and for the mutant was 20.69 Å (±7.78) (data not shown). The predicted radius of gyration was 14.52 Å (±1.18) for the wild-type and 14.41 Å (±1.24) for the mutant FG domain (Figure 2A). The simulations sampled different regions of conformation space because significant run-to-run variations were observed in the probability distributions for each structural parameter. The similar Rg and end-to-end distance values obtained for the wild-type and mutant FG domains implied that both proteins occupy equivalent hydrodynamic volumes. However, this conclusion was at odds with two different quantitative measurements of the physical dimensions of purified FG domains (see below).
Interestingly, it has been reported that increasing MD simulation temperature can yield protein dimensions that more closely resemble those obtained by NMR protein conformation measurements [33],[34]. Indeed, when we extended the nup MD simulations for an additional 1 ns at 325 K (52°C) or at 350 K (77°C), a very different picture emerged (Figure 2A). At 325 K there was a slightly greater difference in the average radius of gyration between the wild-type (15.11±1.43 Å) and the mutant (15.76±2.58 Å) FG domains. Five of the twenty mutant simulations now had an average Rg greater than 18 Å, but all the wild-type simulations had an average Rg below 18 Å, indicating that the mutant FG domain is larger (Figure 2A). In addition, the average end-to-end distance for the wild-type FG domain was 20.84 Å (±10.75) compared to 24.26 Å (±13.16) for the mutant (data not shown). At 350 K, there was a much greater difference between their radii of gyration (Rg). The average Rg was 17.40 Å (±3.11) for the wild-type FG domain and 23.68 Å (±6.05) for the mutant domain (Figure 2A and 2B). At 350 K, fifteen of the twenty mutant simulations had an average Rg greater than 20 Å, compared to only three for the wild-type simulations (data not shown). Consistently, the average end-to-end distance for the wild-type FG domain was 29.95 Å (±16.04) compared to 52.56 Å (±25.31) for the mutant (Figure 2C). The larger dimensions obtained for the wild-type and mutant FG domains at 325 K and 350 K compared to 300 K were likely due to thermal “melting” during the additional 1 ns of simulation. These data combined provide a first indication that the F>A mutant Nup116 FG domain is not as intramolecularly cohesive or compact as the wild-type version.
As a way of assessing the dynamic structure of the FG domain, particularly from the point of view of the FG motifs, we plotted the distances between the backbone β-carbons (Cβ) for the ten sites that correspond to the phenylalanine (Phe, F) or to the substitute alanine (Ala, A) residues in the various FG motifs. The distances used were from the MD simulations at 350 K, which yielded structures (Figure 1C) that better reproduced the dimensional difference between the wild-type and mutant FG domains measured by NMR analysis and in sieving columns (see below). The distance analysis yielded 45 F–F (or A–A) distances for each structure (Table S1). Probability distributions for each Cβ-to-Cβ distance were calculated and analyzed looking for significant differences between the wild-type and mutant FG domain configurations. To estimate the sharpness of the Cβ-to-Cβ distance distributions, the number of 1 Å wide bins that had greater than 10% of the probability distribution was counted; no bin had more that 20% of the probability distribution. This metric was calculated for all 45 Phe–Phe Cβ-to-Cβ distances in all twenty replicates of the FG domain simulations. In stable tertiary structures, these distances occur as one sharp-peak distribution around the equilibrium inter-residue distance; in a fully random ensemble of structures, they occur as a very broad distribution; and in semi-structured proteins, they occur as one or more intermediate-width distributions. Figure 3A shows two representative examples of the probability distributions obtained from the MD trajectories at 350 K; the values shown correspond to the distribution of distances between phenylalanine F84 and F93 in the wild-type FG domain (Figure 1C) or alanine A84 and A93 in the F>A mutant domain. Overall, for the wild-type FG domain simulations, the majority of the simulations analyzed had more than three peaks; by comparison, only a minority of the F>A mutant simulations analyzed exhibited a similar level of sharpness in the distance distributions (data not shown). These results provide tentative evidence that the wild-type FG domain is more structured than the mutant. Repeating this analysis to include only peaks with the distance distributions of <15 Å yielded a very similar result (data not shown).
To permit comparisons of the average inter-residue distances, the probability distributions obtained for the Cβ-to-Cβ distances were fit to a single Gaussian distribution even though in some cases there were multiple distinct peaks (Figure 3A). This was only a rough approximation to the observed probability distribution, but the assumption was justified in the context that these ensemble of structures were to be used (see below). After all, these structures are rapidly inter-converting and the width of the Gaussian is broad enough to accommodate all of the major peaks in the distribution. For example, in the case of the F84–F93 distance distribution, a Gaussian centered at 16.2 Å with a width of 11 Å covered both peaks at 10 and 20 Å (Figure 3A).
Probability distributions of all inter-residue distances obtained from the MD simulations were subjected to clustering using the Pearson squared correlation. This was done to determine how any two of the distributions sampled in regular intervals of the MD simulations are correlated with each other. The correlation coefficient does not depend on the specific measurement units used because other correlation coefficients, such as Euclidian distance metric, yielded similar clustering effects (data not shown). Figure 3B shows the matrix intensity plots of the correlations for the wild-type and mutant FG domains. The indices of the matrix correspond to the various Phe–Phe pairs (or Ala–Ala pairs) listed in Table S1. Indices 1 through 9 correspond to the distance from the first Phe (at position 13; see Figure 1C) to the other 9 Phe (positions 23 through 113), while indices 10–18 correspond to similar distances from the second Phe (at position 23) to the other eight Phe's (positions 32 through 113) and so on. Altogether, the clustering analysis showed that there is a stronger correlation between the various Phe–Phe distributions in the ensemble of wild-type FG domain structures than between the various Ala–Ala distributions in the ensemble of mutant FG domain structures. This indicated that the wild-type FG domain is generally more ordered than the mutant.
To obtain a broader view of the dynamical correlation between inter-residue distances in the FG domains, the Pearson correlation coefficient between all 990 distinct interresidue distances in all 20 simulation replicates of each FG domain were calculated, yielding 19,800 correlation coefficients. In this case, a value of 1.0 would indicate a perfect linear correlation between two inter-residue distances (as one inter-residue distance grew larger, the other would grow by a proportionate amount); a value of 0.0 would indicate no correlation between a distance pair; and a value of −1.0 would indicate perfect anticorrelation. Figure S3 shows back-to-back histograms of the resulting correlation coefficients. For the wild-type FG domain, 9.0% of the observed correlation coefficients had values above 0.7 versus 7.1% in the F>A mutant. This demonstrated that the ensemble of wild-type FG domain structures shows a bias towards higher correlation coefficients between inter-residue distances than the F>A mutant domain, indicating more structural coherence in the wild-type FG domain than in the mutant.
To better describe the relationship between FG motifs in the FG domain, the distances between F–F pairs (or substitute A–A pairs) were also categorized into groups representing distances of 10–15, 15–20, or >20 Å. These are shown in Figure 3C as thick red, medium blue, or thin green lines, respectively. A list of all distances for both proteins is given in Table S2. Among the F–F distances in the wild-type FG domain, seven F–F pairs are less than 15 Å apart (red text in Table S2 and red-thick lines in Figure 3C). In contrast, the F>A mutant FG domain had only two A–A pairs with such short distances. In the wild-type FG domain, four F–F pairs were in the range of 15–20 Å apart (blue), while seven F–F pairs were farther than 20 Å (green). In the F>A mutant, there were eight A–A pairs with distances in the mid-range (15–20 Å), and three pairs showing distances greater than 20 Å. These results demonstrate that the intramolecular distances between FG motifs in the wild-type and mutant FG domains are quantifiably different from each other. There was a tendency for FG motifs in the wild-type FG domain to be proximal to each other (i.e., to cluster), which was absent in the mutant. This conclusion is consistent with the hypothesis that the FG motifs in the wild-type Nup116 FG domain interact intra-molecularly in a manner similar to what has been observed for intermolecular interactions between this Nup116 FG domain and other FG domains of nups [7].
The structural predictions made by the in silico modeling prompted us to seek physical evidence that the phenylalanine residues in FG motifs function as structural cohesion elements that form putative intra-molecular interactions within the Nup116 FG domain. In principle, a change in the dynamic ensemble of FG domain structures resulting from the substitution of all Phe's to Ala's could be detected by NMR. A less-ordered mutant FG domain would exhibit a slower diffusion coefficient. The wild-type and F>A mutant versions of the Nup116 FG domain were purified to `homogeneity and subjected to NMR analysis. Plots of the one-dimensional 1H NMR spectra are shown in Figure 4 (left panels). It was anticipated that the hydration of the FG domains would be significantly different from that of ordered, globular proteins [35],[36] due to the lack of stable folded structures in the FG domains [9]. When presaturation of the water was used there was a significant reduction in the intensity of the amide region of the nup spectrum due to fast exchange with the solvent protons [37],[38]. This observation, combined with the narrow chemical shift dispersion of the amide resonances (7.9–8.5 ppm) in both nup spectra, was a clear indication that both FG domains are natively unfolded and highly dynamic. NMR experiments conducted at lower temperatures (5 and 10°C, compared to 25°C) gave similar results, but offered no significant improvement in the spectral dispersion (data not shown).
Experimental self-diffusion measurements (intensity vs. product of the area of gradient pulse strength and the diffusion length) of the FG domains and the corresponding exponential fits are also shown in Figure 4 (right panels). The data yielded self-diffusion coefficients (Dsexpt) values of 13.17 (±0.26) and 12.18 (±0.12)×10−11 m2 s−1 for the wild-type and mutant FG domains, respectively. This indicated slower diffusion for the less-ordered mutant FG domain. Despite the mass of the F>A mutant domain being smaller (11.9 kDa) than wild-type (12.6 kDa) (due to the replacement of 10 Phe for Ala) the diffusion constant of the mutant was smaller on average, suggesting that its effective hydrodynamic volume is larger. As expected for unfolded proteins [39], the diffusion of the wild-type and mutant FG domains was significantly slower than a folded protein of higher molecular weight [39],[40], indicating that the wild-type and mutant FG domains have unfolded structures that sample a relatively large conformational space.
To further characterize the hydrodynamic properties of the wild-type and mutant Nup116 FG domains, each was analyzed by FPLC in a sieving column to determine its Stokes radius. The expectation was that the less ordered mutant FG domain would occupy more hydrodynamic space and would elute faster from the sizing column. Purified wild-type and mutant versions of the Nup116 FG domain were subjected to size-fractionation through an FPLC Superdex 75 column and their elution profiles were compared to that of commonly-used size standards, such as carbonic anhydrase (29 kDa, Rs = 23.5Å), ovalbumin (45 kDa, Rs = 29.8Å), and BSA (68 kDa, Rs = 35.6Å). The Stokes radius for the wild-type FG domain was measured at 25.2 (±0.6) Å (Table 1), which is larger than carbonic anhydrase despite the FG domain having less than one-half the mass. This highlighted the fact that the FG domain is natively unfolded. The F>A mutant domain eluted faster from the sieving column and migrated as a particle with a Stokes radius equivalent to 27.1 (±0.6) Å, which is larger than the wild-type FG domain despite the mutant having less mass (Table 1). This apparent loss of compaction for the mutant FG domain compared to the wild-type (a ∼20% change in hydrodynamic volume) was consistent with its slower NMR diffusion coefficient (Figure 4) and with the computationally-predicted difference in hydrodynamic dimensions between them at 350 K (Figure 2). The observed loss of intra-molecular cohesion in the mutant FG domain supported our hypothesis that FG motifs within the natively-unfolded FG domain of Nup116 interact intramolecularly via phenylalanines that cluster through hydrophobic attractions. In essence, the hydrophobic interactions between FG motifs likely bias the arrangement of coils within an FG domain to form an ensemble of dynamic non-random tertiary structures with a quantifiable level of intramolecular cohesion.
The hydrodynamic volume or Stokes radius of a protein in different structural configurations (e.g., a folded globule, a molten globule, a premolten globule, a coil, an extended coil) can be estimated from its mass using mathematical equations [41]. These equations were derived from the analysis of large data sets containing experimentally-determined hydrodynamic values for proteins in those structural configurations. Here, using the mass of the wild-type and mutant Nup116 FG domains, we calculated their hypothetical Stokes radius in each structural configuration and compared these predicted values to our experimentally-measured Stokes radii values (Table 1). The goal was to identify the structural configuration of each FG domain that best matched the biophysical measurement obtained for its hydrodynamic volume. For the wild-type FG domain, a predicted native pre-molten globule structure matched best its measured Stokes radius, and for the F>A mutant, a predicted native coil structure was the best match (gray boxes, Table 1). These results support the hypothesis that GLFG motifs in nucleoporins function as intra-molecular cohesion elements, because their absence caused a loss of compaction in the Nup116 FG domain, shifting its dynamic ensemble of structures from native premolten globular configurations to native coil configurations.
We have used a combined computational and biophysical approach to characterize the dynamic ensemble of structures adopted by a natively unfolded or intrinsically unstructured protein. Specifically, we characterized the ensemble of conformations adopted by a fragment of the FG domain of the S. cerevisiae nucleoporin Nup116 (AA 348–458) and of a mutant version thereof (F>A) lacking the phenylalanines in its predominantly GLFG motifs (Figure 1). Both FG domains were found to be highly dynamic and disordered, yet contained quantifiable structural differences between them. The MD simulations predicted a more cohesive and/or compact ensemble of structures for the wild-type FG domain compared to the F>A mutant based on the average radius of gyration and end-to-end distances (Figure 2). This structural prediction was supported by the inter-phenylalanine or inter-alanine distance analysis (i.e., the distance between wild-type or mutant FG motifs, respectively) (Figure 3C), which indicated shorter distances between the FG motifs in the wild-type domain; and by the Pearson correlations of F–F (or A–A) pair distances in the FG domains (Figure 3B), which indicated that the Nup116 FG domain has increased probability of sampling geometries that are more ordered when the phenylalanines in the FG motifs are present. The structural predictions made by the MD simulations were confirmed by direct physical examination of purified FG domains, through NMR-based measurement of their hydrodynamic properties (Figure 4) and by measurement of their hydrodynamic radii in sieving columns (Table 1). In all of the analyses, the wild-type FG domain was found to be more compact than the mutant domain. Unlike the simulations at 350 K, the lower temperature simulations (e.g., at 300 K) did not reproduce this difference in hydrodynamic volumes. Hence, the MD simulations at the higher temperature (350 K) for this class of natively-unfolded proteins may reproduce more accurately their physical properties in solution. As a caveat, the magnitude of the size difference between the nups simulated at 350 K is larger than the magnitude of the difference in their physical dimensions as measured in the sizing columns (Table 1). Notwithstanding, the simulation values matched, within the experimental error of the simulations (s.d. ±18%), the measured Stokes radii for the purified FG domains. Since only 20 single-molecule simulations were used to predict the dimensions of the FG domains, whereas ∼35 trillion molecules were used to accurately measure their average dimension in the sieving columns (s.d. ±2%), it seems likely that a greater number of simulations for greater time-periods could increase the congruency between simulated and measured values.
The mass and physical dimensions of the Nup116 FG domain fragment analyzed here (AA 348–458; MW = 12.6 kDa; Rs = 25.2±0.6), together with the scaling relations developed by Uversky's group [41], led us to conclude that this Nup116 FG domain fragment is best described as a dynamic ensemble of native pre-molten globular structures (Table 1). This structural information can in turn be used to predict the physical dimensions of the full-length Nup116 FG domain (AA 1–960; see [42]) based on its mass and assuming that it also adopts native premolten globular structures. Using the scaling relations, which convert protein mass to physical dimensions in any of a number of structural configurations [41], we estimated that the entire Nup116 FG domain would occupy a hydrodynamic volume equivalent to a 12-nm-diameter sphere (Table S3). For comparison, its volume would be equivalent to a 16-nm-diameter sphere if it were to adopt less compact native-coil configurations; or to a 19-nm-diameter sphere if it were to adopt extended-coil configurations; or to a 7-nm-diameter sphere if it adopted a tightly folded configuration (data not shown). Likewise, size estimations can be done for other full-length nucleoporin FG domains that have similar AA composition and FG motif type as Nup116 (e.g., the GLFG nup subfamily shown in Figure 1B) [42]. Such analysis predicts that their FG domain dimensions would be equivalent to spheres with diameters of 7, 7, 11, and 7 nm for Nup49 (AA 1–251) (Q02199), Nup57 (AA 1–255) (P48837), Nup100 (AA 1–800) (Q02629), and Nup145n (1–216) (P49687), respectively, assuming native premolten globular configurations for each case (Table S3). These predicted dimensions for the FG domains are generally consistent with the 16–46% larger dimensions reported for the full-length FG nups containing the FG domain, the folded NPC anchoring domain and a Protein A tag (Table S3) [43]. Interestingly, all of these FG domains including the Nup116 FG domain appear to be large enough to butt against each other locally within the NPC (at least within a single spoke and probably between adjacent spokes) given their close anchoring at the NPC (see paragraph below) [43],[44], yet appear to be too small to span across the NPC transport conduit from their tether sites within the NPC scaffold (∼19 to 32 nm away from the conduit center; Table S3) to the space occupied by the FG domains of nups anchored at the opposite side (see Figure 1A and Figure S5A). This is because the NPC transport conduit has an estimated radius of 19 nm [2],[43],[44], which is significantly larger than the estimated diameter for these FG domains (∼7–12 nm). Notwithstanding, large fluctuations in the dimensions of FG domains, which are intrinsic to natively unfolded structures, or steric hindrance effects caused by the spatial confinement between closely-anchored FG domains [12],[45] could cause the FG domains to extend further out into the transport conduit (Figure S5B). The cohesive properties between FG domains within the conduit [7],[14] (also see below), or the cross-linking action of karyopherins within the conduit (i.e., karyopherins appear to bind multiple FG motifs in different FG domains simultaneously) [45]–[47] could transiently stabilize some of the extended FG domain conformations (Figure S5B and S5C, respectively) [14].
The evidence presented here suggests that the FG motifs in Nup116 function as structural, intramolecular cohesion elements that bias the arrangement of coils within the FG domain and condense its dynamic ensemble of structures into more cohesive, less disordered states. In the case of the Nup116 FG domain examined, its FG motifs are responsible for shifting its ensemble of structures from native-coil configurations (as seen for the mutant) to native pre-molten globular configurations (Table 1 and Figure 1C). In principle, all types of FG motifs (GLFG, FxFG, SAFG, PSFG, etc.) [42] could exert cohesion through hydrophobic pairing, stacking, zippering, or otherwise clustering of the aromatic ring of phenylalanine side chains through energetically favorable aromatic edge-to-face interactions, as opposed to less favorable face-to-face (π–π) interactions [48]. Interestingly, a report by Dhe-Paganon et al., defined a “phenylalanine zipper” motif within the hydrophobic core of APS, which is critical for APS dimerization [49]. There, the aromatic side chains of ten phenylalanine residues are uniquely stacked to form a zipper that is stabilized by helical secondary structures in the protein backbone. Although FG domains do not appear to have stable secondary structures, residues surrounding the FG motif, such as the leucine residue of GLFG motifs or the second phenylalanine residue in FxFG motifs, could enhance the hydrophobic clustering effect by increasing the local hydrophobicity of the Phe residue in the FG motif and/ or by influencing the orientation of its Phe ring. A two dimensional representation of the Nup116 FG domain AA sequences in a hydrophobic cluster analysis (HCA) [50],[51] illustrates its hydrophobic “LF” patches very well (Figure S4). Although HCA is most commonly used in determining hydrophobic clusters in helical patterns [31], it is also informative in the absence of a structural fold because it allows the identification of hydrophobic features between nearby AAs. The HCA analysis highlighted LF patches and MFMF-connections in the wild-type Nup116 FG domain, which were missing in the F>A mutant domain. This implied that the F>A mutant FG domain is less compact because it does not have hydrophobic patches and connections to make intra-molecular interactions.
Our finding that the FG motifs can function as intramolecular cohesion elements has important implications for the general architecture and function of the NPC, especially if the ability of these FG motifs to mediate intramolecular cohesion of coils functionally mimics their demonstrated ability to mediate intermolecular cohesion between FG domains [7]. Indeed, the representative FG domain of Nup116 analyzed here and the FG domains of other (but not all) FG nups (Nup49, Nup57, Nup100, nNup145, and Nup42) engage in homotypic and heterotypic interactions with each other in vitro and in vivo via FG motifs [7]. By analogy to the Nup116 FG domain, this group of cohesive FG domains may also exhibit intramolecular cohesion of their own FG motifs. At the NPC, such intra- and intermolecular FG motif interactions could be in competition with each other, possibly causing the FG domains to fluctuate between monomeric and polymeric states (Figure S5B). Alternatively, such interactions could be in a dynamic equilibrium with each other to form a metastable quaternary structure (Figure S5A). According to the two-gate and the hydrogel models of NPC architecture, the FG motifs of nups within the NPC conduit engage in intermolecular cohesions with each other to form a highly flexible network of cohesive polypeptide chains, which forms a size-selective sieve or gate [7]–[9],[13],[14]. The cohesiveness of such network(s) is presumably maintained by the weak but numerous interactions between FG motifs [52]. However, if the intra- and intermolecular interactions between FG motifs were in competition with each other at the NPC, then intramolecular cohesions could effectively prevent the FG domains from forming a network altogether by causing them to “fold back” on themselves (i.e., an autoinhibitory mechanism).
What type of FG motif interaction dominates at the NPC, either intramolecular or intermolecular, is indeed an important question whose answer may rely largely on four parameters: the distance between FG domain anchor sites at the NPC, the volume of space occupied by each FG domain, the space available at the NPC for each FG domain, and the steric hindrance effect between neighboring FG domains [12],[45],[53]. As discussed above, the estimated dimensions for the “cohesive” GLFG-rich domains of yeast nups (7–12 nm diameter spheres) combined with the close proximity between their anchor points within each spoke (most are ≤5 nm apart from others anchored adjacently and ≤10 nm from others anchored above or below in the z-axis; see Table S3) implies that at least within a spoke and probably between adjacent spokes the FG domains of these nups butt against each other to occupy overlapping space [43],[44]. This close positioning could allow or even promote the formation of a supra-molecular quaternary structure of cohesive FG domains at the NPC through a multitude of inter-molecular FG motif interactions. This structural assembly could take the form of a meshwork of intertwined polypeptide chains [14],[52], or alternatively, based on the data presented here, the assembly could take the form of a doughnut-shaped array of laterally-cohesive, native pre-molten globules (as depicted in Figure 1A and Figure S5A). Most importantly, local reversible shifts in the equilibrium between intra- and intermolecular FG motif interactions could facilitate the fast structural changes in the NPC permeability barrier, which are presumably coupled to the passage of karyopherin-cargo complexes of different shapes and sizes during transit across the NPC (Figure S5C). As karyopherin–cargo complexes disrupt (either by mass action or by direct interaction with the nup FG domain) intermolecular FG motif interactions during transit (as predicted for all FG nups in the hydrogel model, or for a discrete subset of nups in the two-gate model) [7],[8],[14], the FG motifs liberated as a result would become available to form intramolecular interactions. This could cause the FG domains to fold back on themselves (i.e., to compact), effectively opening the permeability barrier by suddenly occupying less space.
It remains to be determined whether other types of nup FG domains, which do not display intermolecular cohesions with each other via FG motifs [7], can nevertheless form intramolecular cohesions of their own FG motifs to adopt compact configurations. According to the “virtual gate” [11], the “oily spaghetti” [54] and the “two gate” [7],[55] models of FG domain architecture, these FG domains would exist as highly-extended polypeptide chains, as observed for the Xenopus Nup153 FG domain [12]. Interestingly, in the case of Nup153, its extended FG domain appears to compact upon binding a karyopherin [45]. Clearly, a more detailed structural characterization of the various FG domains as they interact with karyopherins and each other is needed to fully understand the dynamic and highly flexible structure of the NPC transport conduit.
MD simulations of individual FG domains were started from a fully extended backbone structure (i.e., with the Φ and Ψ angles set to 180° for all residues except for the three proline residues, which put a 60° bend in the sequence). A different random number seed was chosen for each of the different simulations to randomize the initial atom velocities. Twenty separate simulations were run for either 6 ns (wild-type) or 5 ns (mutant) each using different initial atomic velocities and analyzed at 1 ps intervals. The wild-type fragment required an additional nanosecond of dynamics to have its radius of gyration converge. All MD simulations were performed with AMBER [56]–[58] using implicit solvent models. Each molecule was simulated in the presence of a Generalized Born/Surface Area (GB/SA) implicit solvent model [59] that calculates an effective solvation energy as an empirical parameter multiplied by the exposed surface area of different atom types. Each molecule was simulated using the GB/SA implicit solvent implementation in Amber versions 7 and 8. Each system is energy-minimized using 100 cycles of conjugate gradients. Constant-temperature molecular dynamics at 300 K with a coupling constant of 2.0 ps was performed on the minimized systems using the standard partial charges for the Amber force field and Bondi radii for the atoms. Bonds containing hydrogens were constrained using SHAKE and a time step of 2 fs was used in all simulations. A cutoff of 250 Å was used for the electrostatic interactions, which for this system is equivalent to infinity. The salt concentration (Debye-Huckel screening) was set at 0.15 M. Secondary structure analysis: For the final 3 ns of each simulation, the structure was analyzed every 1 ps using a standard program for identifying secondary structure from atomic coordinates (Define Secondary Structure of Proteins; DSSP) [60]. Radius of gyration and end-to-end distance analyses: For the final 3 ns of each simulation, radii of gyration and the end-to-end distances between terminal residues were calculated using the program CARNAL and ptraj, distributed with AMBER 7 [57],[58]. High temperature molecular dynamics simulations: Molecular dynamics were performed at elevated temperatures for each of the 20 wild-type and mutant FG domain simulations. The GB/SA simulations were all restarted after 5 ns coupled to a heat bath at 325 or 350 K with all other parameters of the simulation kept the same. The simulations were run for 1 ns, and the final 500 ps were used for analysis.
To determine the degree of dynamical change in the ensemble of FG domain structures, the autocorrelation function was derived for a vector (total vector length = 236) composed of the 118 Φ and 118 Ψ angles (in the 111 AA nucleoporin sequence with a 9 AA N-terminal tag; see Figure 1) along the peptide backbone of structures sampled every 1 ps from the MD trajectory. The autocorrelation function was computed as the dot-product of successive Φ–Ψ vectors using every 1 ps step as a new time origin and calculating the correlation function out to 200 ps. Several different time windows were used in the autocorrelation calculation, but all gave the same result. For comparison purposes, the same autocorrelation function was calculated for the first 118 Φ–Ψ angles (out of 130) for an MD trajectory of a well-folded protein (fibroblast growth factor 1; PDB ID = 1AXM).
The MD trajectories of 45 F–F distances between the Cβ atoms were analyzed to calculate the probability distributions. Systematically, each of the F–F distance was interrogated and all of the distances were binned (1 Å bins from 0 to 60 Å) to form a histogram of distance distributions. Probability distributions were calculated for each of the twenty simulations independently, and the values obtained were averaged at the end. A similar procedure was adopted for the mutant FG domain where the distance between the Cβ atoms of the Ala residue was used. Final probability distributions were used without any normalization.
As a first approximation, the probability distributions were fit to a Gaussian distribution (probability versus distance). This is a conservative approach and is expected to be valid considering the number of structures generated (60,000) during the molecular dynamics simulations and in the absence of any constraints. The center of the Gaussian is considered as the mean distance between the F–F (or A–A), while the width at half-maximum is used as the allowed variation in the constraint. Clustering analysis: The correlation between different F–F probability distribution reflects the degree to which these variables (F–F distances) are related. The most common measure of correlation is the Pearson Product Moment Correlation (http://www.r-project.org/) and reflects the degree of linear relationship between the two variables. In order to determine whether probability profiles of the F–F interaction correlate, a similarity matrix with a Pearson square metric was calculated. The correlation was used to indicate the presence (or absence) of relationship between various F–F interactions.
The coding sequence for the representative 111 AA Nup116 FG domain was amplified from genomic S. cerevisiae DNA using PCR and was cloned into the vector pGEX-2TK in frame with the coding sequence for glutathione S-transferase (GST) at the 5′ end, and in frame with the coding sequence for six contiguous histidines at the 3′ end. Site directed mutagenesis was then used to alter the coding sequence for the mutant F>A FG domain. The correct coding sequences were confirmed by DNA sequence analysis. The FG domains were expressed in a E. coli BL21+ strain as fusion proteins with GST (glutathione S transferase) at the N-terminus and a HIS tag (six contiguous histidine residues) at the C-terminus. Glutathione coated Sepharose beads were then used to isolate each GST-FG domain fusion from crude bacterial cell extracts. The isolated FG domains were eluted from the beads by specific thrombin proteolysis of the GST tag. Nickel-coated beads were then used to capture and isolate the FG domain through its C-terminal His-tag, and the captured proteins were eluted from the beads using imidazole. Finally, the eluates were concentrated in a Centricon 3 unit and were size fractionated in an FPLC Superdex 200 sizing column that was equilibrated in 50 mM NaH2PO4, pH of 6.5 for the NMR analysis, or in an FPLC Superdex 75 column equilibrated in 20 mM Hepes, pH 6.8, 150 mM KOAc, 2 mM Mg(OA)2 for determination of Stokes radii.
Tandem-affinity purified wild-type and F>A mutant Nup116 FG domains were subjected to size-fractionation through an analytical-scale FPLC Superdex 75 column. FG domains (100 µl of 7.5 mg/ml) were injected at a flow rate of 0.5 ml/min at 4°C into a column that was preequilibrated in 20 mM Hepes pH 6.8, 150 mM KOAc, 2 mM Mg(OAc)2, and 0.5 ml fractions were collected. The FG domain elution profiles were monitored by UV absorbance at 280 nm and by SDS-PAGE analysis of the eluates. The nup elution profiles were compared to those of carbonic anhydrase (29 kDa, Rs = 23.5 Å), ovalbumin (45 kDa, Rs = 29.8 Å), and BSA (68 kDa, Rs = 35.6 Å), which served as molecular size standards. The elution volume of the standards was plotted in relation to their Stokes radii, allowing for estimation of the FG domain Stokes radii from the resulting linear regression formula.
NMR experiments were performed on tandem-affinity purified FG domains dissolved in 50 mM NaH2PO4, pH 6.5. Final protein concentrations were ∼0.5 mM for both wild-type and mutant FG domains. NMR experiments were performed in a Varian INOVA 600 MHz spectrometer equipped with a 5 mm probe with a single-axis (along Z) shielded magnetic field gradients. One dimensional 1H NMR experiments were obtained using the water suppression scheme 1-3-3-1 Water-gate [61]. Self-diffusion coefficient measurements were obtained using a BPP-SED (bipolar-gradient pulse pair selective echo dephasing) sequence [62].
Translational diffusion tensor values were calculated based on the beads-model approximation of García de la Torre and Bloomfield [63]. This method has been used successfully to calculate translational as well as rotational diffusion tensors of proteins [40],[64]. All atoms were considered as beads of equal size (σ = 5.1 Å). The overall isotropic translational self-diffusion coefficient was calculated by taking the average of the principal values of the diffusion tensor.
The hydrodynamic radius (Rh) for the wild-type and mutant Nup116 FG domains was calculated from the radius of gyration (Rg) values obtained from the simulations using the scaling relationship given in [39]. For native proteins, the scaling relationship is Rh = Rg/0.77, and for proteins in strong denaturing conditions, the scaling relationship is Rh = Rg/1.06. For the wild-type and mutant Nup116 FG domains simulated at 300 and 325 K, the hydrodynamics radius was obtained by Rh = Rg/0.77. In the 350 K simulations, some of the protein conformations were highly extended (as in denaturing conditions) and a single scaling value was not appropriate. In this case, if the Rg for a structure was less than 30.7 Å for wild-type and 29.6 Å for the mutant, it was scaled by 1/0.77; if the Rg was greater, the value was scaled by 1/1.06. The Rg cutoff values of 30.7 Å (wild-type) and 29.6 Å (mutant) were obtained by using Uversky's relationship: Rh (8 M urea) = (0.22)*M0.52, where M is the molecular mass [41]. The molecular mass for the simulated wild-type FG domain was 11,791 Daltons and for the mutant domain was 11,030 Daltons. The calculated Rh values were 22.5±4.0 for the wild-type FG domain and 28.6±5.2 for the mutant. To compare these Rh values to the Stokes radii values (Rs, same as Rh) for the purified FG domains in sieving columns, the contribution of a C-terminal His-tag (6 histidine residues/841 Da), which was added (post simulations) to the FG domains to aid in the purification of only full-length FG domains, had to be factored in. This was done using Uversky's scaling relationship by calculating Rs for the FG domains with the additional tag assuming a native pre-molten globular configuration for the wild-type and a native coil configuration for the mutant (see Table 1). The Rs estimated from the molecular dynamics simulations for the wild-type and mutant FG domains were multiplied by the ratio (Rs (His-tag)/ Rs (no-tag)) to yield the final values of 23.1 (±4.1) Å for the wild-type FG domain and 29.6 (±5.4) Å for the mutant FG domain reported in Table 1.
|
10.1371/journal.pntd.0002738 | Contrasting Patterns in Mammal–Bacteria Coevolution: Bartonella and Leptospira in Bats and Rodents | Emerging bacterial zoonoses in bats and rodents remain relatively understudied. We conduct the first comparative host–pathogen coevolutionary analyses of bacterial pathogens in these hosts, using Bartonella spp. and Leptospira spp. as a model.
We used published genetic data for 51 Bartonella genotypes from 24 bat species, 129 Bartonella from 38 rodents, and 26 Leptospira from 20 bats. We generated maximum likelihood and Bayesian phylogenies for hosts and bacteria, and tested for coevoutionary congruence using programs ParaFit, PACO, and Jane. Bartonella spp. and their bat hosts had a significant coevolutionary fit (ParaFitGlobal = 1.9703, P≤0.001; m2 global value = 7.3320, P≤0.0001). Bartonella spp. and rodent hosts also indicated strong overall patterns of cospeciation (ParaFitGlobal = 102.4409, P≤0.001; m2 global value = 86.532, P≤0.0001). In contrast, we were unable to reject independence of speciation events in Leptospira and bats (ParaFitGlobal = 0.0042, P = 0.84; m2 global value = 4.6310, P = 0.5629). Separate analyses of New World and Old World data subsets yielded results congruent with analysis from entire datasets. We also conducted event-based cophylogeny analyses to reconstruct likely evolutionary histories for each group of pathogens and hosts. Leptospira and bats had the greatest number of host switches per parasite (0.731), while Bartonella and rodents had the fewest (0.264).
In both bat and rodent hosts, Bartonella exhibits significant coevolution with minimal host switching, while Leptospira in bats lacks evolutionary congruence with its host and has high number of host switches. Reasons underlying these variable coevolutionary patterns in host range are likely due to differences in disease-specific transmission and host ecology. Understanding the coevolutionary patterns and frequency of host-switching events between bacterial pathogens and their hosts will allow better prediction of spillover between mammal reservoirs, and ultimately to humans.
| Bats and rodents are important hosts for emerging human diseases. While a large body of research has focused on viral pathogens in these hosts, the diversity, evolution, and transmission of their bacterial pathogens remains relatively unstudied. We conducted co-evolutionary analyses of two bacterial genera know to be pathogenic in humans, Bartonella and Leptospira, along with their bat and rodent hosts. We found that Bartonella had a significant pattern of coevolution with both bat and rodent hosts, while Leptospira in bats showed a lack of congruence with its bat hosts and a high number of host switching events. Our statistically driven approach to understand the frequency of host switching events in these mammal–bacterial systems can be easily applied to other host–pathogen systems, including viruses, to assess the likelihood of zoonotic spillover.
| Bats and rodents are the two most diverse and geographically widespread orders of mammals [1], [2], and are important reservoirs for a growing number of emerging infectious diseases (EIDs) with significant impacts on public health. Bats are reservoir hosts of several viral pathogens of high consequence, including Henipaviruses, Ebola and Marburg viruses, lyssaviruses, Severe Acute Respiratory Syndrome coronavirus, and likely Middle Eastern Respiratory Syndrome coronavirus [3]–[5]. Rodents are known reservoirs of hantaviruses, arenaviruses, Lassa fever virus, plague and other bacterial zoonoses [6]. Over the last two decades, the majority of research on bat and rodent zoonotic diseases has focused on viral infections (Figure S1). While the number of virus-related publications for bats has had a marked rise over the past decade, research on bacteria in bats has remained consistently low (Figure S1). The evolutionary relationships between these important mammalian hosts and their known bacterial pathogens has been little studied to date [7], [8].
Bats and rodents are evolutionarily ancient orders of mammals, with periods of diversification dating back 75 and 85 million years ago, respectively, thus allowing ample time for pathogens and hosts to coevolve [9]. Bats and rodents make up 60% of all extant mammal species while exhibiting a wide-range of life-history and ecological traits. Ecological, evolutionary, and life-history traits can influence pathogen richness and cross species transmission, or spillover, in these bat and rodent hosts [5], [8]–[11]. The peridomestic habits of these mammals also likely increase the frequency of human contact and facilitate disease spillover [12], [13]. Anthropogenic alterations that increase exposure to bats and rodents, including expanding agricultural operations, bushmeat hunting, and climate change, may increase the opportunity for diseases to emerge in human populations in the future [14]. How these ecological and life history factors may affect the coevolutionary patterns between reservoir host species and their associated pathogens is an open question, but will depend on characteristics related to pathogen transmission and host ecology.
The evolutionary patterns of hosts and their known pathogens can be used to quantify the frequency of spillover events within and between reservoir hosts, and is a crucial first step for developing predictive models for zoonotic disease emergence. Previous research has demonstrated how these coevolutionary studies can shed light on specific instances of host switching, cospeciation, and other events in coronaviruses and their bat hosts [15], as well as malaria parasites and their avian hosts [16]. However, to our knowledge, no comparative cophylogenetic analysis of bacterial pathogens has been applied yet to bat and rodent hosts. Here we examine host-pathogen evolution in bats and rodents using two bacterial genera, Bartonella spp. and Leptospira spp., known to cause neglected tropical diseases in humans.
The genus Bartonella consists of globally distributed and highly diverse alpha-proteobacteria that infects a wide-range of mammals. After infection, the bacteria eventually enters erythrocytes and endothelial cells and can persist asymptomatically in a wide range of mammalian reservoir hosts [17]. The disease is mainly transmitted through arthropod vectors including fleas, flies, lice, mites, and ticks [18]–[21]. Thus, the transmission and evolution of Bartonella species in mammals is the result of a complex relationship between multiple hosts, vectors, and pathogens. Bartonella has been reported with high prevalence and genetic diversity from numerous recent studies in bats [22]–[25] and rodents [26]–[30]. Bartonella is recognized as a neglected tropical disease, and there are indications of human infections derived from neighboring wildlife populations. In Thailand, genetic studies have indicated highly similar Bartonella strains between infected humans and nearby rodent populations [31]–[34]. Neighboring rodents have also been implicated as a possible source for bartonellosis in the United States and Nigeria [35], [36].
Leptospira is a genus of spirochete bacteria which also has a wide geographical distribution [37], and has been recognized as an important emerging pathogen due to its increasing incidence in both developing and developed countries [38]. Leptospires are maintained in nature by a large variety of wild and domestic animal hosts, and the bacteria colonize their kidneys and are excreted in their urine [39]. Rodents were the first recognized carriers, though the bacteria has been isolated from almost all screened mammals. Recently, bats have been found to carry Leptospira in Madagascar, Australia, Peru, and Brazil, and seroprevalence has been recorded to be as high as 35% [40]–[44]. Unlike Bartonella spp., Leptospira spp. are not vector-borne, and transmission to humans and other hosts is primarily through contact with water and environments contaminated with infected animal urine [45]. While most research has focused on rodents reservoirs of leptospirosis, recent genetic studies have also indicated bats as carriers of the bacteria [41], [44].
In order to better understand the evolutionary dynamics of Bartonella and Leptospira in bat and rodent hosts, we compiled available genetic information from hosts and bacterial pathogens to determine cophylogenetic patterns on a global scale. Evidence of cophylogeny can be used to test hypotheses of coevolution, and a lack of congruence between host and pathogen phylogenies can identify pathogen spillover, or interspecific transmission, events [46]. Long associations through evolutionary time can lead to reciprocal adaptations in both the hosts and their parasites, as well as concurrent divergence events in the two lineages. Evolutionary events including strict codivergence, parasite duplication, parasite extinction, and parasite host switching, will either strengthen or diminish the congruence between host and parasite [47]. Patterns of host-parasite or host-pathogen congruence may also vary geographically. For example, host specificity of Bartonella was observed in Old World bats in Kenya [24], while bats in Peru and Guatemala in the New World appeared to have no specific Bartonella-bat relationships [22], [48]. In contrast, consistency is observed for Bartonella in rodents, with host-specificity apparent in both Old and New World [49]–[51].
The primary goal of this paper is to examine the global co-evolutionary patterns of bats, rodents and their associated bacterial pathogens – using Bartonella and Leptospira as case studies. We specifically test for evolutionary congruence between bat host species and Bartonella and Leptospira, as well as rodent host species and Bartonella. Analysis of rodent Leptospira was unfortunately excluded due to a lack of comparable sequence datasets and host taxonomic diversity. Although there is a long history of research on leptospirosis in rodents, the publicly available sequence data that has been obtained thus far covers only a handful of rodent host species distributed across 3 genes: secY, flab,and lipL3 [52]–[56]. We also test whether evolutionary patterns and bacterial host specificity differ between the New World and Old World bat and rodent hosts, as was previously observed for Bartonella [22], [24], [48]. Finally, we conduct event-based cophylogeny analyses to reconstruct likely evolutionary histories for each group of pathogens and hosts.
Sequence data used for analyses were obtained by searching for relevant papers from 1900–2013 through online sources PubMed, Web of Science, and Google Scholar using keywords “Bartonella*” and “Leptospir*” combined with “bat OR Chiroptera*” or “rodent*”. All Bartonella and Leptospira sequences from bat or rodent hosts identified to the species level were compiled into our datasets (Tables S1, S2, S3). Bat hosts include individuals in the Artibeus, Brachyphylla, Carollia, Coleura, Desmodus, Eidolon, Glossophaga. Hipposideros, Lonchophylla, Micronycteris, Mimon, Miniopterus, Monophyllus, Myotis, Nyctalus, Otomops, Phyllostomus, Promops, Pteronotus, Rousettus, Rhinophylla, Sturnira, Triaenops, Uroderma, and Vampyressa genera (Table S1, S3). Rodent hosts include individuals in the Acomys, Aethomys, Apodemus, Callosciurus, Clethrionomys, Dryomys, Gerbillus, Glaucomys, Jaculus, Mastomys, Microtus, Mus, Myodes, Niviventer, Pachyuromys, Peromyscus, Psammomys, Rattus, Rhabdomys, Sekeetamys, Spermophilus, Tamias, Tamiasciurus, Tatera, and Urocitellus genera (Table S2). Only unique genotypes were included in the dataset. The largest comparable genetic datasets consisted of the partial citrate synthase gene (gltA) for Bartonella and 16S rRNA gene for Leptospira, and these were selected for analysis. Cytochrome b gene sequences from all bat and rodent host species were obtained from GenBank (Tables S4, S5), as this mitochondrial gene has proven to be useful for within Order, species-level resolution of mammalian phylogenies [57]–[59]. For host species that did not have an available cytochrome b sequence, the most closely related species with available sequence was used as a substitute for host-parasite associations. For bats, we made four substitutions: Hipposideros armiger for Hipposideros commersoni, Phyllostomus hastatus for Phyllostomus discolor, Promops centralis for Promops nasutus, and Triaenops persicus for Triaenops menamena. Our results suggest that these genus-level host substitutions do not disrupt overall co-phylogenetic patterns. For all species, host taxonomy was synonymized according to Mammal Species of the World 3rd Edition [60].
In total, we compiled sequences from 51 Bartonella genotypes (38 New World, 13 Old World) from 24 bat species (15 New World, 9 Old World), and 129 (20 New World, 109 Old World) Bartonella genotypes from 38 rodent species (4 New World, 35 Old World). We also compiled sequences from 26 Leptospira genotypes (19 New World, 7 Old World) from 20 bat species (14 New World, 6 Old World). Insufficient genetic data of one gene for Leptospira in rodents precluded their use in the analysis; therefore only Leptospira in bat hosts was examined.
Bacterial and host species sequences were imported from GenBank into Geneious Pro 5.0.4. Sequences for each bacterial genus and their corresponding bat and rodent hosts were each aligned using default parameters in MUSCLE [61] as implemented in Geneious [62]. Outgroup taxa, obtained from GenBank, were included in each alignment, and were chosen based on previous species-level phylogenies. The outgroup for Bartonella was Brucella melitensis [25], for Leptospira was Leptonema illini [44], and for the bat and rodent hosts was the duck-billed platypus, Ornithorhynchus anatinus HQ379861 [63]. In order analyze the difference in host-specificity between Old and New World geographic regions, each alignment was further divided into Old and New World. Alignments were inspected visually and ends were trimmed and gaps found in only one non-outgroup sequence were deleted due to high likelihood of sequencing error. After these edits, this resulted in 1,133 base pairs (bp) for cytb bat sequences, 338 bp for gltA Bartonella sequences, and 1,246 bp for 16S Leptospira sequences.
Maximum likelihood (ML) phylogenetic trees were generated using RAxML 7.0.4 [64] implemented with the Cyberinfrastructure for Phylogenetic Research (CIPRES) Portal (www.phylo.org) using the substitution model GTRMIX, which determines an optimal tree by comparing likelihood scores under a GTR+G model. The number of bootstrap replicates were determined using the previously described stopping criteria. In order to corroborate the phylogenies as determined through ML, Bayesian inference (BI) host phylogenies were also generated using MrBayes 3.1.2 [65]. We utilized a GTR+I+G substitution model, with 10,000,000 generations, sampling every 5000th generation with 4 heated chains and a burn in length of 1,000,000.
To visualize host-bacteria associations, tanglegrams were generated from the best ML trees in TreeMap 3.0 [66]. For cophylogenetic analyses, we utilized both global fit as well as event-based methods. We selected programs that are capable of accounting for evolutionary patterns given association of parasite species to multiple hosts, as well as the presence of multiple parasites in a single host.
Global-fit methods were used to quantify the degree of congruence between two given host and parasite topologies, and identify the individual associations contributing to the cophylogenetic structure [67]. First, global-fit analysis was tested using distance-based ParaFit [68], using matrices of patristic distances calculated from maximum likelihood host and parasite phylogenies in R 3.0.1 [69]. With an additional matrix of host-parasite links, ParaFit analyses [68] were also performed in R using package ape [70] with 999 permutations to implement a global test as well as individual links. Each individual host-bacteria interaction is determined to be significant if either its ParaFit 1 or Parafit 2 p-value≤0.05, and these significant interactions are shown in solid lines in the tanglegrams.
As ParaFit tends to be liberal with its values, we also implemented newly developed program Procrustean Approach to Cophylogeny (PACo) [71] in R using packages ape and vegan [72] in order to obtain, and potentially corroborate, comparable global goodness-of-fit statistics with Parafit global values. PACo differs from ParaFit by utilizing Procrustean superimposition, in which the parasite matrix is rotated and scaled to fit the host matrix. Thus, PACo explicitly tests the dependence of the parasite phylogeny upon the host phylogeny.
We then used event-based program Jane 4 [73] to determine the most probable coevolutionary history of the associated host and parasites, again using the ML host and bacteria trees as input. We assigned different relative costs to 5 possible evolutionary events, in a method similar to previous research efforts [74]. We performed analyses with 100 generations, population sizes of 100, and a default cost setting matrix of 0 for cospeciation, 1 for duplication of parasites, 2 for duplication and host switch, 1 for loss of parasite, and 1 for failure to diverge. In further runs, we changed one of the possible events to a cost of 10 each time, rendering that event prohibitively expensive. By further exploring the parameter space this way, we determined how these changes affected the overall costs of the optimal evolutionary history.
The topology of the BI tree was identical to that of the ML tree, except for a few branches with low support values. Thus, only the ML trees are presented here and used for further cophylogenetic analyses. Phylogenies tend to be well supported for more recent divergence events, but not deeper nodes. Nodes with bootstrap values ≥50 are labeled on all tanglegrams (Figures 1–6).
Based on a default cost setting, we calculated the optimal number of each type of coevolutionary event, to minimize total cost, for each host-pathogen association (Table 1). In order to account for the different sample sizes of each of the phylogenies, we divided the number of cospeciation and host switch events by the number of parasites in each association (Table 1). The resulting ratios can then be compared across the different associations in order to see the overall impact of each event given the number of parasites. Based on these calculations, Leptospira and bat host associations have the greatest number of host switches per parasite (0.731), while Bartonella and rodent host associations have the fewest (0.264). Leptospira and bat host associations also have the greatest number of cospeciations per parasite (0.231), while Bartonella and rodent host associations have the fewest (0.132).
We also compared cophylogenetic fit between bacteria and Old World vs. New World host species. Bartonella had nearly twice as many host switches per parasite in New World (0.474) as compared to Old World bats (0.278). Bartonella in rodents had the opposite trend, with more than twice as many host switches per parasite in Old World (0.339) as compared to New World rodents (0.150). There were also approximately three times as many cospeciation events per parasite for Bartonella in the Old World for both groups of hosts (bats: 0.333, rodents: 0.156) compared to the New World (bats: 0.132, rodents: 0.050). For Leptospira, the differences between Old and New World cophylogenetic patterns for both host switches and cospeciations were minimal.
We explored a wide-range of cost parameters in order to determine the effect of removing different evolutionary event options from each analysis (Table S6). Increasing the cost of host switching events had the greatest overall impact on total cost. In fact, the role of the host switching events in the overall coevolutionary pattern was so strong that even with a potentially prohibitive cost of 10, the solution still proposed between 2–4 host switch events for Bartonella in New World bats and Old World rodents.
We found significant coevolutionary congruence between Bartonella and both their rodent and bat hosts at a global level, while the relationship between Leptospira and their bat hosts was non-significant. Event-cost results support the global-fit findings, with the rodent-Bartonella and bat-Bartonella associations having the least number of host switches per parasite, which indicates greater evolutionary congruence over time. Co-evolution of bartonellae and their mammalian hosts also remains significant when New and Old World datasets are analyzed separately. The evolutionary pattern in bat hosts is driven mostly by a few strong host-parasite interactions, with 51% of individual associations significant. In comparison, a greater proportion, 67%, of the individual rodent-bacteria associations are significant, indicating stronger coevolutionary interactions throughout these lineages. In fact, in the New World association of rodents and Bartonella, a full 100% of the host-parasite links were significant. In contrast, for Leptospira and their bat hosts, there is only 1 significant individual in analysis of the entire data set, and no significant host-parasite links when the data are analyzed separately as Old vs. New World. We note that the sample sizes for Bartonella in New World rodents and Leptospira in Old World bats are both small, and that the observed patterns could change with the addition of more data. Similarly, the relatively short sequence available of only one gene for both Bartonella and Leptospira may limit the resolution and nodal support for the pathogen phylogenies we obtained. These issues can only be addressed with additional sampling and genetic sequencing to complement these sparse datasets. For example, while our analysis of Bartonella-host relationships was limited by the availability of gltA fragments, the use of multi-gene phylogenies would be a more robust approach given the confounding effect of recombination [75]. Despite these potentially confounding factors, our preliminary analysis suggests a strong signal was present for some host-pathogen relationships and at a host order and pathogen genus level these trends were generalizable.
Event-cost methods corroborate the non-significant coevolutionary history of Leptospira and bats. Interestingly, the number of cospeciations per parasite is also the highest for Leptospira and bats, although they also have the highest number of host switches per parasite. Since their overall coevolutionary relationship is nonsignificant, this suggests that for the bat-Leptospira system, coevolutionary relationships are driven mostly strongly by the host switching events rather than cospeciation. Exploring the parameter space of cost structures further supports our findings. For all associations, maximizing the cost of host switching results in the largest overall change in the total cost (Table S6). This indicates that host switching is an “expensive” evolutionary event, and our finding of frequent and well-supported host switching in the bat-Leptospira system suggest that there are intrinsic ecological and transmission factors driving this.
One explanation for the different coevolutionary patterns between Bartonella and Leptopira may be differences in the modes of transmission and infection dynamics for each pathogen. As a vector transmitted parasite, Bartonella has an additional evolutionary step in adapting to an arthropod organism as well as a mammalian host. Combined, this can exert greater evolutionary selection and act as a selective force driving speciation. Further, Bartonella forms persistent, often asymptomatic, infections in its hosts [17], and some evidence even suggests that Bartonella may be acting as a symbiont more than a pathogen [18], [76], [77]. Many Bartonella species are also likely transmitted by only one arthropod species [78], and this specificity can then be translated to a greater coevolutionary pattern between the disease and eventual mammalian host. In bats, the arthropod vectors include blood-feeding bat flies, from which Bartonella has been sequenced and cultured [77], [79]. Host specificity of these arthropods may help to maintain the high diversity of Bartonella and long-term coevolutionary patterns between bat flies and their Bartonella parasites [77]. However no in-depth cophylogenetic analyses have been conducted for these bacteria and their known arthropod vectors, and this is an area for future exploration. Additional studies on arthropod ecology, e.g. bat fly, population structure, dispersal, ecology, and host specificity will also help to clarify the role of bat hosts vs. arthropod vectors in the evolution of Bartonella [77], [80]. Additionally, Bartonella is an intracellular bacteria which can survive only within erythrocytes and endothelial cells [17]. This requires a finer adaptation to the host's cells in order for bacterial penetration. In summary, Bartonella infection dynamics favor vector transmission, and the specific host-vector relationships, potential vertical transmission in vectors, and intracellular nature of the bacteria allow for co-evolutionary relationships to develop over time.
In contrast, Leptospira spp. are not vector-transmitted and instead are transmitted via environmental contamination. Leptospires are able to survive outside of their hosts, and can persist in water bodies when shed in animal urine [81]. Although the vast majority of Leptospira infections are mild, a small proportion involve multiple organ systems and develop various complications resulting in a case fatality in human patients of about 40% [45]. As contact with urine and contaminated water is the main form of disease transmission, physical proximity to environmental sources can play a large role in influencing host-pathogen interactions [82]. Thus it is possible that geographic overlap of the host species will better predict similarity in the bacteria they carry rather than the phylogenetic relatedness of the hosts. Overlapping geographic distribution of host species has been found to be an important determinant of pathogen sharing in primates [83]. The role of environmental transmission is most likely why we observed frequent host switching events and a lack of coevolutionary patterns in the Leptospira lineages we studied. Further investigations of Leptospirosis disease dynamics, including shedding, transmission, and immunity, in bat populations is warranted, as well as their zoonotic potential given the propensity towards cross-species transmission.
We originally hypothesized a difference in the strength of coevolutionary relationships between Old and New World host species, since previous research in bats had indicated host specificity in the Old World but not New World for Bartonella [24]. While the mechanism for this observation was not clear, it may be hypothesized that a greater degree of congruence between host-bacteria phylogenies in the Old World may be due to longer evolutionary time for the establishment of mutualistic relationships with mammalian hosts [24]. Yet, in our larger datasets, we did not see this pattern emerge, and our results indicated that coevolutionary patterns are generalizable globally. For Bartonella, significant coevolutionary congruence with hosts was evident globally and across host ranges, while for Leptospira, the lack of a coevolutionary relationship in bat hosts was evident in both the Old and New World. However, it is interesting that we observed a stronger relationship between rodents and Bartonella than between bats and Bartonella. There are two possible explanations for this. First, in mammalian evolutionary history, rodents existed for a longer period with 4.1 million years earlier time of origin and a 10 million year difference in time of basal diversification between the two [9]. Thus it is possible that there has been a longer time for parasite-host relations to coevolve in rodents and create stronger patterns. However it is not clear that these hosts have been infected with the two pathogens in question over their entire evolutionary history, and further detecting such deep evolutionary divergences is confounded by genetic saturation and nucleotide homoplasy. A second explanation is that the ecological differences between bats and rodents may explain the observed differences in host-specificity. Unlike rodents, a number of bat species form large multi-species gatherings and are more likely to have direct ecological overlap between host species (e.g. many thousands of individuals from >8 bat species roosting together in a single cave site in Mexico [84]). Similarly, at sites across the tropics, an extraordinary numbers of bat species can exist in sympatry, e.g. >70 species sharing tropical forest habitat in Krau Wildlife Reserve, Malaysia [85]. The gregarious aggregations of highly mobile individuals, often between multiple species, may help to explain differences in the global coevolutionary patterns observed between bats and rodents. While there has been growing scientific interest in these ecological and life-history host traits to explain viral sharing in bats and rodents [5], [8], [10], the role that these traits may play in bacterial pathogen diversification and spillover has been little investigated to date.
Overall, it is likely that the interplay of multiple factors, including geographic overlap, pathogen transmission pathways, infection dynamics, and host ecological and evolutionary history, that contribute to the contrasting coevolutionary patterns evident in mammal-bacteria interactions we observed. Further research is warranted to better understand and tease apart these contributing factors, and we recognize some of the limitations of this preliminary study. First, our analysis was limited by the availability of comparable data sets for a given gene and host taxonomic group. This precluded us from examining Leptospira in rodents; and resulted in low support values from some nodes in our phylogenies. In the future, using multiple genes or full genome data, for a greater number of bat and rodent taxonomic groups and bacterial microbes once they are available, will allow for more robust taxonomic analyses. Also, in addition to host phylogeny that we examine here, future data collection and analyses should focus on arthropod vector host specificity and phylogenetic relationships to better predict specificity within Bartonella. Future investigations should also consider the role of host geographic range and niche overlap to explain pathogen sharing between hosts. The application of spatial analyses of wildlife hosts for both Bartonella and Leptospira will provide valuable information on transmission potential based on the role of contact vs. cophylogeny. We predict that species with overlapping ranges will share more similar communities of Leptospira than non-overlapping bat species, regardless of their phylogenetic relatedness. For Bartonella there is also likely to be a geographic effect, as interaction among bats of different species within multi-species roosts, or shared habitats, could be an important factor for bacterial pathogen sharing.
Lastly, this work is particularly important because it involves two emerging, neglected tropical diseases with known, sylvatic wildlife reservoirs. Bartonella has been of concern as an emerging zoonoses due to its ability to induce life-threatening illnesses such as endocarditis, myocarditis, meningoencephalitis, and contributing to chronic debilitating disease, all while being difficult to diagnose in humans as well as animals [86]. Leptospirosis is a constant concern to public health authorities, and annual global incidence of severe leptospirosis has been estimated as 500,000 [87]. Elucidating the diversity and coevolutionary patterns of these bacteria in their natural hosts and understanding the frequency and causes of host-switching events, will help us better predict spillover from the mammal reservoirs into humans. Disruption of strict coevolutionary patterns, as we observed for both bacterial genera, to varying degrees, provides a framework to forecast pathogen spillover potential to any mammalian host, including humans [88]. The methods that we employed here to study bacterial disease in bats and rodent hosts are broadly applicable to a wide range of other disease types, including viruses in their mammalian hosts. By expanding these tools to better understand the evolutionary past of pathogens within and among wildlife hosts, we gain information to better predict the outbreaks of the future.
|
10.1371/journal.pntd.0002067 | Cross-reactive Neutralizing Antibody Responses to Enterovirus 71 Infections in Young Children: Implications for Vaccine Development | Recently, enterovirus 71 (EV71) has caused life-threatening outbreaks involving neurological and cardiopulmonary complications in Asian children with unknown mechanism. EV71 has one single serotype but can be phylogenetically classified into 3 main genogroups (A, B and C) and 11 genotypes (A, B1∼B5 and C1∼C5). In Taiwan, nationwide EV71 epidemics with different predominant genotypes occurred in 1998 (C2), 2000–2001 (B4), 2004–2005 (C4), and 2008 (B5). In this study, sera were collected to measure cross-reactive neutralizing antibody titers against different genotypes.
We collected historical sera from children who developed an EV71 infection in 1998, 2000, 2005, 2008, or 2010 and measured cross-reactive neutralizing antibody titers against all 11 EV71 genotypes. In addition, we aligned and compared the amino acid sequences of P1 proteins of the tested viruses.
Serology data showed that children infected with genogroups B and C consistently have lower neutralizing antibody titers against genogroup A (>4-fold difference). The sequence comparisons revealed that five amino acid signatures (N143D in VP2; K18R, H116Y, D167E, and S275A in VP1) are specific for genogroup A and may be related to the observed antigenic variations.
This study documented antigenic variations among different EV71 genogroups and identified potential immunodominant amino acid positions. Enterovirus surveillance and vaccine development should monitor these positions.
| Recently, enterovirus 71 (EV71) has caused life-threatening outbreaks in tropical Asia. EV71 has one single serotype but can be phylogenetically classified into 3 main genogroups and 11 genotypes (A, B1∼B5 and C1∼C5). In Taiwan, nationwide EV71 epidemics with different predominant genotypes occurred in 1998(C2), 2000–2001(B4), 2004–2005(C4), and 2008(B5). In this study, historical sera from children infected with these 4 genotypes were collected to measure cross-reactive neutralizing antibody titers against 11 genotypes. In addition, amino acid sequences of P1 proteins of the tested viruses were compared. Serology data showed that children infected with genogroup B and C consistently have lower neutralizing antibody titers against genogroup A (>4-fold difference). Antigenic variations between genogroup B and C could be detected but did not have a clear pattern. Five amino acid signatures are specific for genogroup A and may be related to the observed antigenic variations. Vaccine development should monitor the antigenic and genetic variations to select vaccine strains.
| Human enteroviruses include over 100 serotypes and usually cause self-limited infections, except polioviruses and enterovirus 71 (EV71) which frequently involve neurological complications [1], [2]. Although EV71 was first described in 1969, a retrospective analysis shows that this virus circulated in the Netherlands as early as 1963 [3]. Recent molecular evolution studies predicted that EV71 could have emerged in the human population around 1941 [4]. Globally, two patterns of EV71 outbreaks have been reported: small-scale outbreaks with low mortality and large-scale outbreaks with high mortality. The latter pattern occurred in Bulgaria with 44 deaths in 1975, in Hungary with 45 deaths in 1978, in Malaysia with 29 deaths in 1997, in Taiwan with 78 deaths in 1998, in Singapore with 5 deaths in 2000, and recently in China with more than 100 deaths every year after 2007. Due to its tremendous impact on healthcare systems, development of EV71 vaccines is a national priority in some Asian countries [2].
EV71 has one single serotype as measured by hyperimmune animal antiserum but can be phylogenetically classified into 3 genogroups (A, B and C) and 11 main genotypes (A, B1∼B5 and C1∼C5) by analyzing the most variable capsid protein sequences (VP1) [1]. Recently, one new genogroup was only detected in India [5]. Genogroup A viruses were isolated in 1970 in the United States and were not detected globally again until 2008. In an investigation of a HFMD outbreak in central China in 2008, Yu et al identified five EV71 isolates which were closely related to genotype A based on analysis of the VP1 gene [6]. In contrast, genogroups B and C are widely circulating in the world with different evolution patterns [7], [8]. Interestingly, genogroup replacements of EV71 have been well documented in Taiwan and Malaysia [1], [2]. In Taiwan, nationwide EV71 epidemics with different predominant genotypes occurred in 1998 (C2), 2000–2001 (B4), 2004–2005 (C4), and 2008 (B5) [9]–[11]. In this study, sera from EV71-infected children were collected to measure cross-reactive neutralizing antibody titers against different genotypes, which are critical to understand the drivers of genogroup replacement and viral diversity, and for selection of vaccine strains.
Institutional review board approvals were obtained from Chang Gung Memorial Hospital, and National Taiwan University following the Helsinki Declaration. Written informed consents were obtained from parents/guardians on behalf of all child participants.
To avoid confounding effects of EV71 re-infections, historical sera were collected from young children who were under 5 years of age and infected with different EV71 genotypes in 1998 (genotype C2, 10 sera), 2000 (genotype B4, 5 sera), 2005 (genotype C4, 2 sera), 2008 (genotype B5, 5 sera), or 2010 (genotype C4, 3 sera) [10], [12]–[14]. These sera were used to measure cross-reactive neutralizing antibody titers against all 11 EV71 genotypes.
Twelve strains of the 11 EV71 genotypes were used in the study, including two genotype C4 viruses which were isolated in 2005 and 2008, respectively. Eight of these twelve viruses were isolated in Taiwan and the other four viruses (genotype A, B2, B3 and C3) had not circulated in Taiwan (Table 1). All viruses were amplified in rhabdomyosarcoma (RD) cells using Dulbecco's Minimum Essential Medium (DMEM) containing fetal bovine serum 2% v/v and penicillin/streptomycin. Virus titers (50% tissue culture infectious doses, TCID50) were determined in RD cells using the Reed-Muench method.
The P1 region of the EV71 genome encodes four capsid proteins including VP1, VP2, VP3 and VP4 proteins, which are involved in the induction of immune response and the infection of cells [15]–[17]. Therefore, the P1 regions of 11 EV71 genotypes were sequenced to identify correlations between genetic and antigenic variations. Viral genomic RNA was extracted from 140 µL of virus culture isolates using a QIAmp Viral RNA kit (Qiagen, USA) according to the manufacturer's instructions. cDNA of EV71 was synthesis by SuperScript II Reverse Transcriptase (Invitrogen, USA). PCR reactions were performed by specific primers and KAPA HiFi DNA Polymerase (Kapa Biosystems, USA). Primers used in this study are listed in Supporting Table S1. Nucleotide sequences of P1 regions (2586 bp) were aligned and analyzed by the Mega 4 software (Molecular Evolutionary Genetics Analysis software version 4.0) [18]. Phylogenetic trees were constructed by the neighbor-joining method using the Maximum Composite Likelihood method and the prototype CA16 strain (CA16/G-10) as the outgroup virus. The reliability of the tree was estimated using 1,000 bootstrap replications. Nucleotide sequences analyzed in this study have been submitted to GenBank.
Laboratory methods for measuring EV71 serum neutralizing antibody titers followed standard protocols [19], [20]. Briefly, 50 µL of two-fold serially diluted sera and virus working solution containing 100 TCID50 of EV71 were mixed on 96-well microplates and incubated with RD cells. A cytopathic effect was observed in an inverted microscope after an incubation period of 4–5 days. Each serum dilution includes three replicates and the neutralization titers were read as the highest dilution that could result in a reduction of the cytopathic effect in at least two of three replicate wells. Each test sample was run simultaneously with cell control, positive serum control, and virus back titration. If the ratios of neutralizing antibody titers between different genotypes were greater than 4, we measured neutralizing antibodies titers at least three times to confirm the accuracy of tests.
Large tabular serological data are hard to summarize and are recently analyzed using antigenic cartography (i.e., antigenic map) [11], [21], [22]. Briefly, antigenic cartography is a way to visualize and increase the resolution of serological data, such as neutralization data. In an antigenic map, the distance between a serum point S and antigen point A corresponds to the difference between the log2 of the maximum titer observed for serum S against any antigen and the log2 of the titer for serum S and antigen A. Thus, each titer in a neutralization assay table can be thought of as specifying a target distance for the points in an antigenic map. Modified multidimensional scaling methods are used to arrange the antigen and serum points in an antigenic map to best satisfy the target distances specified by the neutralization data. The result is a map in which the distance between points represents antigenic distance as measured by the binding assay [11]. In this study, an antigenic map was generated using a web-based analytic tool [22].
Neutralizing antibody titers were log transformed to calculate the geometric mean titers (GMTs), and their 95% confidence intervals (95% CI). The GMTs of cross-reactive neutralizing antibody titers were further used to generate an antigenic map using a web-based analytical tool [22]. The relative positions of strains and antisera were adjusted such that the distances between strains and antisera in the map represent the corresponding ratios between homologous and heterologous neutralizing antibody titers. Differences between homologous and heterologous neutralizing antibody titers were tested for statistical significance by the nonparametric tests (NPAR1WAY Procedure) using SAS software (SAS Institutes, Cary, NC).
Nucleotide sequences analyzed in this study have been submitted to GenBank (accession numbers JN874547–JN874558).
Twenty-five sera were collected from 25 young children who were infected with EV71 genotype C2, B4, C4, B5, and C4 in 1998, 2000, 2005, 2008 and 2010, respectively. Cross-reactive neutralizing antibody titers against 11 EV71 genotypes are shown in Table 2. Overall, all EV71-infected children had detectable neutralizing antibody titers against 11 EV71 genotypes. Interestingly, homologous neutralizing antibodies titers were not always higher than heterologous neutralizing antibody titers.
As shown in Table 2, serum neutralizing antibody titers against the homologous genotype (C2) in children infected in 1998 varied over 100-fold and they were grouped into two groups (low and high titers) for further analysis. In addition, children infected in 2005 and 2010 were merged for further analysis because they were all infected with genotype C4. GMTs of neutralizing antibody titers against 11 genotypes are shown in Figure 1. Overall, children infected with genotype C2, C4, B4 and B5 had lower GMTs (>4-fold difference) against genotype A than other genotypes. In contrast, antigenic variations between genogroup B and C did not have a clear pattern. We further merged neutralizing antibody titers against different genotypes within the same genogroup to calculate GMT for further comparisons. As shown in Figure 2, children infected with genotype C2 and C4 had similar GMT against genogroup B and C but children infected with B4 and B5 had higher GMTs against genogroup B than against genogroup C. We further constructed the antigenic map using GMT of cross-reactive neutralizing titers presented in Figure 1. Overall, genotypes in genogroup B and C clustered together and genotype A was found to be outside of the cluster (Figure 3).
To investigate the correlation between genetic and antigenic variations of EV71 genotypes, nucleotide and deduced amino acid sequences of P1 regions of 11 EV71 genotypes (12 viruses) were analyzed. Pairwise comparisons of P1 regions have shown that the nucleotide (amino acid) differences within EV71 genogroup were 0.049∼0.151 (0.005∼0.015) for Genogroup B and 0.042∼0.135 (0.005∼0.013) for Genogroup C and the nucleotide (amino acid) differences were 0.209∼0.224 (0.02∼0.026) between Genogroup A and B, 0.21∼0.235 (0.018∼0.024) between Genogroup A and C, and 0.188∼0.228 (0.021∼0.032) between Genogroup B and C (Table 3). Overall, the nucleotide differences in the P1 region within genogroup were much lower than that between genogroups (0.042∼0.151 vs. 0.188∼0.235) but the differences in amino acid sequences were not as abundant as found in nucleotide sequences (0.005∼0.015 vs. 0.018∼0.032) (Supporting Table S2). Genetic variations in VP1, VP2, VP3 and VP4 genes were also analyzed (Supporting Table S2). Interestingly, nucleotide differences in VP1∼VP4 were similar but no amino acid differences were observed in VP4 gene, which may exclude influence of VP4 on antigenic evolution of EV71.
Phylogenetic analyses based on nucleotide sequences of the P1, VP1 and VP1+VP3 regions are shown in Figure 4. Overall, the phylogenetic trees generated using the P1 and VP1+VP3 regions indicated that genogroups B and C were distinct from the genotype A; however, the phylogenetic tree based on the VP1 region suggested that genogroup A is clustered with genogroup C. Overall, the phylogenetic relationship among the EV71 genotypes did not match with the antigenic relationship observed in this study. To further determine the amino acid differences related to the observed antigenic variations shown in Fig 1 and 3, the deduced amino acid sequences of P1 regions were aligned to reveal that five amino acid signatures (N143D in VP2; K18R, H116Y, D167E, and S275A in VP1) are specific for genogroup A and may be related to the observed antigenic variations (Figure 5).
EV71 has one single serotype as measured by hyperimmune animal antiserum but antigenic variations have been reported recently in human studies [9]–[11]. Using sera collected from young children with primary infection of genotype B5, two studies detected partial antigenic differences between genogroup B and C but not between viruses in the same genogroup (B5 and B4 viruses) [9], [10]. Kung et al. did not detect significant antigenic differences between genotypes B4 and C4 viruses using acute-phase sera from EV71 inpatients [23]. A serological survey in healthy Japanese children and adults detected partial antigenic differences between genotype B5 and A viruses but not among different genotypes in genogroup B and C that had previously circulated in Japan [24]. By constructing an antigenic map using 14 children sera, however, Huang et al. detected antigenic differences between genogroup B and C, and also between B5 and B4 viruses [11]. In a monkey study, Arita et al. [25] found that monkeys immunized with live-attenuated EV71 vaccine (genotype A) induced similar (<4-fold difference) antibody responses against genotype B1 but lower (≧4-fold difference) antibody responses against genotype B4, C2 and C4. In our study, we found that children infected with genotype C2, C4, B4 and B5 had lower GMTs (≧4-fold difference) against genotype A than other genotypes but antigenic variations between genogroup B and C did not have a clear pattern, which is different from the Huang study [11]. It is hard to compare different studies which had small sample size and employed different human sera and laboratory procedures, in particular the cell lines (RD cells vs. Vero cells) and virus strains used in the neutralization assay. A network to harmonize laboratory procedures including standard sera and viruses is required to make the comparison possible. Moreover, the clinical and epidemiological significance of the antigenic variation requires longitudinal serological studies to clarify.
Most clinical studies, including our study faced the limitation of small sample size due to the difficulty of collecting large amounts of serum samples from young children. Ideally, suitable animal models should be developed to generate a panel of antisera for monitoring EV71 antigenic variations, as ferrets served for influenza surveillance [26]. Representative EV71 clinical isolates could be selected for monitoring antigenic variations using the animal antisera. The clinical isolates with significant antigenic variations detected using animal antisera would be further evaluated using children post-infection sera.
Currently, five EV71 vaccine candidates are under evaluation in clinical trials, including three genogroup C viruses and two genogroup B viruses [27]. Based on the cross-reactive neutralizing antibody presented in the current study, genogroup B and C viruses are expected to induce protective neutralizing antibodies against genogroup B and C viruses but not genogroup A viruses. Interestingly, genogroup A viruses have disappeared for over 35 years but re-emerged in China in 2008. In an investigation of a HFMD outbreak in central China in 2008, Yu et al identified five EV71 isolates which were closely related to genotype A based on analysis of VP1 genes but these genogroup A viruses did not spread widely [6]. Reasons for the reemergence of genotype A in central China are not clear, and the full genomic sequences of the isolates should be performed to clarify the issue. Recently, novel genotype C2-like viruses were detected in Taiwan in 2008 and children infected with genotype C4, C5, B4 and B5 viruses had much lower (>100-fold) serum cross-reactive neutralizing antibody titers against the novel C2-like virus than against the homologous viruses. Interestingly, these novel C2-like viruses were recombinants of genotype C2 and B3 viruses but they did not spread widely [9]. Based on historical poliovirus studies, immunodominant neutralizing epitopes mainly locate on VP1 and VP2 proteins. Recently, binding sites of two EV71 mice neutralizing monoclonal antibodies were identified using synthetic peptide technology to locate at amino acid position 211–225 of VP1 protein and amino acid position 136–150 of VP2 protein, respectively [16]. The importance of these linear epitopes in the human immune response is not clear. In the current study, we combined human serological data and viral genetic sequence data to identify five amino acid positions (4 on VP1 protein and 1 on VP2 protein) related to antigenic variations. Only one of these five positions (VP2-143) was also identified in the mice monoclonal antibody studies. The clinical significance of these five positions needs to be verified using reverse genetics to generate mutant viruses. Recently, the 3-dimensional structures of EV71 capsid proteins have been published [28], [29]. Structural studies elucidating interaction between EV71 capsid proteins and neutralizing antibodies will help understand the mechanism of vaccine-induced immunity and design better vaccines.
Traditionally, the phylogenetic relationship of EV71 genotypes has been widely analyzed using VP1 nucleotide sequences [1] . Interestingly, a recent study found that the VP1-based phylogenetic tree is not similar to the complete genome-based phylogenetic tree [7]. Our study also found that the phylogenetic trees based on VP1 and P1 nucleotide sequences differ slightly. Specifically, genogroup A is close to genogroup C in the VP1-based phylogenetic tree but this relationship was not found in the P1-based phylogenetic trees. It is well known that enteroviruses including EV71 frequently recombine at the junction of structural (P1) and non-structural (P2 or P3) genes [8], [30]. Therefore, the P1 gene is suitable for phylogenetic analysis but the complete genome is required for detection of gene recombination. However, the P1 gene (about 3000 nucleotides) is much larger than the VP1 gene (about 890 nucleotides) and the P1 gene may not be readily available. The combined VP1+VP3 gene (about 1600 nucleotides) is much shorter than the P1 gene but could generate a similar phylogenetic tree to that based on the P1 gene. Overall, the VP1 gene is good enough for defining genotypes of genogroup B and C viruses, but it would be better to analyze the phylogenetic relationship between genogroup A viruses and other genogroup viruses based on the VP1+VP3 or P1 genes.
From an evolutionary perspective, a recent analysis of 628 EV71 VP1 sequences estimated that EV71 emerged in the human population around 1941 and evolved more quickly in the past 20 years [4]. It is unclear why EV71 has evolved more quickly in the past 20 years. Recombination, being a common occurrence among enteroviruses, might be the likely explanation for the emergence of EV71, but it would require full genome analysis to better understand the mechanism of EV71 evolution, which is critical to long-term success of EV71 vaccination programs.
|
10.1371/journal.pmed.1002819 | Government policy interventions to reduce human antimicrobial use: A systematic review and evidence map | Growing political attention to antimicrobial resistance (AMR) offers a rare opportunity for achieving meaningful action. Many governments have developed national AMR action plans, but most have not yet implemented policy interventions to reduce antimicrobial overuse. A systematic evidence map can support governments in making evidence-informed decisions about implementing programs to reduce AMR, by identifying, describing, and assessing the full range of evaluated government policy options to reduce antimicrobial use in humans.
Seven databases were searched from inception to January 28, 2019, (MEDLINE, CINAHL, EMBASE, PAIS Index, Cochrane Central Register of Controlled Trials, Web of Science, and PubMed). We identified studies that (1) clearly described a government policy intervention aimed at reducing human antimicrobial use, and (2) applied a quantitative design to measure the impact. We found 69 unique evaluations of government policy interventions carried out across 4 of the 6 WHO regions. These evaluations included randomized controlled trials (n = 4), non-randomized controlled trials (n = 3), controlled before-and-after designs (n = 7), interrupted time series designs (n = 25), uncontrolled before-and-after designs (n = 18), descriptive designs (n = 10), and cohort designs (n = 2). From these we identified 17 unique policy options for governments to reduce the human use of antimicrobials. Many studies evaluated public awareness campaigns (n = 17) and antimicrobial guidelines (n = 13); however, others offered different policy options such as professional regulation, restricted reimbursement, pay for performance, and prescription requirements. Identifying these policies can inform the development of future policies and evaluations in different contexts and health systems. Limitations of our study include the possible omission of unpublished initiatives, and that policies not evaluated with respect to antimicrobial use have not been captured in this review.
To our knowledge this is the first study to provide policy makers with synthesized evidence on specific government policy interventions addressing AMR. In the future, governments should ensure that AMR policy interventions are evaluated using rigorous study designs and that study results are published.
PROSPERO CRD42017067514.
| Despite global commitments to reduce antimicrobial resistance and protect the effectiveness of antimicrobials, most countries have not yet started implementing government policies to reduce their overuse and misuse of antimicrobials.
To the best of our knowledge, no evidence syntheses have attempted to identify the policy options available to government policy makers to tackle antimicrobial resistance by reducing antimicrobial use in humans.
We searched 7 academic databases to identify impact evaluations of government policy interventions aiming to reduce human antimicrobial use that were published in any language before January 28, 2019.
We found 69 studies that evaluated government policy interventions to reduce antimicrobial use around the world. From these, we were able to describe 17 different types of policies that governments have used to tackle this major driver of antimicrobial resistance in humans.
Commonly used policy strategies included public awareness campaigns and antimicrobial guidelines; however, other policy strategies focused on vaccination, stewardship, and changing regulations around prescribing and reimbursement.
We found 4 randomized controlled trials and 35 studies using rigorous quasi-experimental designs. The remaining 30 studies used uncontrolled and descriptive study designs.
Our systematic evidence map suggests that governments have a variety of policy options at their disposal to respond to the growing threat of antimicrobial resistance.
Unfortunately, most existing policy options have not been rigorously evaluated, which limits their usefulness in planning future policy interventions.
To avoid wasting public resources, governments should ensure that future antimicrobial resistance policy interventions are evaluated using rigorous study designs, and that study results are published.
| Antimicrobial resistance (AMR) is currently high on the global political agenda. This attention has opened a rare policy window for achieving meaningful action on AMR [1–7]. Although the potential for AMR has been recognized since the earliest days of antibiotics [8], the misuse and overuse of antimicrobials has persisted over decades, contributing to the development of resistance [9]. AMR is now expected to have severe consequences for human health, social well-being, and economic development. AMR has already rendered some infections untreatable using existing antimicrobials [10,11], and global projections suggest that AMR could derail the Sustainable Development Goals, driving an estimated 24 million people into extreme poverty and exacerbating global economic inequality [12], and potentially resulting in tens of millions of deaths [13].
Successfully overcoming the threat posed by AMR will require multi-sectoral and multi-jurisdictional cooperation to protect the effectiveness of existing and future antimicrobials [2,3,5]. Recent political initiatives addressing AMR, including the 2016 United Nations resolution [14] and the 2017 Berlin Declaration of the G20 Health Ministers [15], are promising signs that governments and international agencies are mobilizing to act on AMR. The 194 member states of the World Health Organization (WHO) agreed to develop national AMR action plans by 2017 [16], and countries have largely responded [17].
Despite these positive steps, most countries have not yet started implementing policies to reduce their overuse and misuse of antimicrobials [17]. Evidence from high-income countries suggests that reducing antimicrobial use is associated with lower rates of resistance [9], yet there is limited evidence on what types of government policy interventions effectively reduce antimicrobial use. Typically, government policy changes are useful tools for improving public health when the health threat requires widespread change and uniform compliance with a set of minimum standards [18]. However, research on AMR has principally focused on changing the prescribing behaviours of individual physicians [19], rather than creating large-scale reductions in antimicrobial use through population-wide interventions.
Given that governments are currently grappling with the challenge of implementing AMR policies under their recently developed national action plans, a focus on the potential impact of government policy interventions on antimicrobial use is timely. Governments are currently attempting to weigh the merits of numerous types of policy interventions that could safely reduce antimicrobial use, utilizing policy levers such as legislation, taxation, economic incentives, funding support, public awareness campaigns, and regulation of professionals and businesses whose work might affect AMR [18]. Policy makers would benefit from a tool that catalogues and assesses the government policy responses that have been used in various contexts and health system settings. Thus we undertook a systematic evidence mapping project to support evidence-informed action on AMR at the government level, by identifying, describing, and assessing the full range of government policy interventions aiming to reduce human antimicrobial use that have been implemented and evaluated.
A protocol describing the full methods of this project was published in advance [20] and registered in PROSPERO (CRD42017067514). Deviations from the protocol are noted below, and the paper has been reported in accordance with the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guidelines [21]. In brief, we produced an evidence map that identifies government policy interventions aiming to reduce antimicrobial use in humans. To be included in the evidence map, studies had to (1) clearly describe a government policy intervention aiming to reduce human antimicrobial use and (2) apply a quantitative design to measure the impact.
We searched 7 electronic databases from medicine and the social sciences (MEDLINE, CINAHL, EMBASE, PAIS Index, Cochrane Central Register of Controlled Trials, Web of Science, and PubMed [articles not indexed in MEDLINE]) from inception to January 28, 2019, without language or date limits. Targeted web searching was used to identify grey literature, and the ProQuest Dissertations & Theses database was used to identify dissertations. We contacted subject-matter experts in each of WHO’s 6 regions to identify missing studies.
We screened titles and abstracts against 3 inclusion criteria: (1) the evaluated intervention was a policy intervention defined as an intervention enacted by a government or government agency at the federal, state, provincial, or municipal level that aimed to change antimicrobial use through education, restriction, incentivization, coercion, training, persuasion, changing the physical or social context, modelling appropriate behaviour, or reducing barriers to action [22]; (2) the study quantitatively evaluated the effect of the intervention; and (3) the study assessed an outcome measure related to human antimicrobial use such as consumption, dosing, prescribing, or sales of an antibiotic, antiviral, antiparasitic, or antifungal drug. Examples of interventions include regulating the sales of antimicrobials, restricting the use of last-resort antibiotics, and launching public awareness campaigns. Titles and abstracts were each independently screened by 2 reviewers (SRVK and S Jones, A Srivastava, or RN), and disagreements were resolved by consensus. The full text of potentially relevant studies was screened by 2 reviewers (SRVK and MN or RN). Non-English articles were translated using Google Translate, or a translation was requested from the corresponding author.
Data on study characteristics, study participants, interventions, analyses, and measured effects were extracted in duplicate by 2 reviewers (SRVK and MN or RN) using a customized data extraction tool (see S2 Text). In consultation with SJH and JMG, SRVK grouped studies according to the Behaviour Change Wheel framework’s intervention functions and our definition of policy intervention [22]. Where appropriate, studies were coded with multiple Behaviour Change Wheel intervention functions; however, studies were coded with a single Behaviour Change Wheel policy approach. We inductively identified and described policy options based on groupings of similar interventions, and coded studies according to their region, study design, and the intervention functions of the Behaviour Change Wheel. WHO regions were used to group countries; the region of the Americas was subdivided into Canada/US and Latin America.
From 13,635 abstracts, we identified 69 evaluations of government policy interventions to reduce human antimicrobial use. Fig 1 shows the full summary of screening and inclusion. Of the 69 included studies, 67 focused on antibiotics and 2 on antimalarials; no studies aimed to reduce the use of other antimicrobial agents. The majority of policies targeted healthcare workers (n = 44) or healthcare workers and the community (n = 13), while the remaining 12 policies exclusively targeted a community audience. We found evaluations in 4 of the 6 WHO regions—the Americas (n = 24), Western Pacific (n = 22), Europe (n = 21), and Africa (n = 2)—but did not identify any evaluations from the South East Asian region or the Eastern Mediterranean region. Of the 69 included studies, 67 were published in English and 2 were published in Spanish.
Using our definition of policy intervention, we organized studies according to the policy categories of the Behaviour Change Wheel framework. Table 1 describes the interventions of the included studies. The largest grouping of policies was regulatory interventions (n = 27), followed by guidelines (n = 18), communication policies (n = 17), legislation (n = 3), and fiscal measures (n = 3). One evaluation was identified for service provision policies; however, we did not identify any social planning policies. Regulatory policies (n = 20/27) and legislation (n = 3/3) were largely organized at the national level. Communication policies were organized at different levels of government; 8 were at the national level, 3 were at the state/provincial level, and 6 were at the regional level (municipality, county, or other geographic unit). Similarly, 12 of the guidelines were at the national level, 2 were at the state/provincial level, and 4 were at the regional level. One of the fiscal measures policies was at the national level, and 2 were at the regional level. The sole service provision policy was organized at the national level.
The majority of the 69 included studies were retrospective evaluations using routinely collected data from health insurance databases or electronic health records (n = 46), or sales data from IMS Health (n = 14). Four of the included studies were randomized controlled trials, 3 were non-randomized controlled trials, 7 used non-randomized controlled before-and-after designs, 25 used time series designs, 18 used uncontrolled before-and-after designs, 10 used descriptive methods, and 2 used cohort study designs (Box 1). The included studies predominantly used antibiotic consumption measured as defined daily doses (n = 25) or physician prescribing rates (n = 26) as an outcome measure.
Among the 69 included evaluations, we identified 17 distinct policy options that have been evaluated for their ability to reduce antimicrobial use. Table 2 summarizes these policy options and lists the studies that evaluated specific manifestations of them. By far the most common of these policy options were informational strategies, including public awareness campaigns (n = 17), which informed healthcare workers and/or the public about AMR and antimicrobial overuse, and antimicrobial guidelines (n = 13), which provided information to healthcare workers on the preferred use of antimicrobial drugs or preferred treatments for resistant infections. These strategies were widely used across most regions, and in particular represented a large proportion of the interventions evaluated in Canada/US [52,53,59,61,62,69,74,76–78,80,81] and Europe [63,65,67,68,70,71,73,75,79,82,90,91].
Other policy options were less commonly reported and tended to group regionally, as can be seen in Fig 2. For example, 7 studies were categorized as “prescription requirement” policies, which were regulatory and legislative policies essentially banning the sale of over-the-counter antibiotics by requiring a prescription from a healthcare professional [24,25,39,40,44,85,86]. These policies were implemented starting in the late 1990s in countries or regions in Latin America where over-the-counter antibiotic sales were not previously prohibited, or where existing regulations were not enforced. Countries in WHO’s Western Pacific region tried a diverse range of strategies, most of which were evaluated only once or twice. These policies were largely implemented in China [27,29,33,42,43,45–49,60,66,89], South Korea [32,49,57,87], and Taiwan [26,30], and included disclosure requirements for hospitals to post their antibiotic use rates online, professional regulation strategies that changed the codes of practice around the prescribing and dispensing of antibiotics by different health professions, and reimbursement penalties for physicians, who were not paid for their services unless their prescriptions met the guidelines for antibiotic prescribing. Along similar lines, 1 European country (Denmark) explored using reimbursement penalties targeting patients, where the national health insurance plan reimbursed patients a smaller proportion of the antibiotic cost than was previously reimbursed [41]. Three studies in Canada/US [34,36,38] and 1 each in Europe [31] and the Western Pacific [30] also tried reimbursement restrictions, where the patient was not reimbursed the cost of an antibiotic by their national health insurance plan unless the physician met particular guidelines such as proving the existence of an infection or consulting with an infectious disease specialist. In the African region, we found only 2 studies, both targeting the use of antimalarial drugs, and both employing published guidelines to change prescribing [50,51]. Some national-level interventions were evaluated in multiple studies using different methods, populations, and evaluation time frames; these included the ban on over-the-counter sales of antimicrobials in Chile [24,25,44], the Antibiotics Are Not Automatic campaign in France [70,71,73,82], and the national essential medicines policy in China [29,45,46,49].
Around the world, governments are currently working to develop policy responses to the growing threat of AMR. In our evidence map, we identified 69 evaluation studies looking at the impact of policy interventions on antimicrobial use across 4 of WHO’s 6 regions. From this search, we were able to identify 17 different policy options and examples of each that governments can use to inform their future AMR policies.
Many of the policy options identified in this map were evaluated in only a few studies. These evaluations were highly regionalized, which likely results from similarities between contexts and health systems within regions of a country, or in neighbouring countries. Policy makers in other parts of the world who operate with similar contextual problems or health systems may find these policies useful models for policy development in their countries. For example, the prescription requirement policies [24,25,39,40,44,85,86] were implemented in 5 Latin American countries where over-the-counter antibiotic sales were formally or informally permitted. While this type of policy would not be useful in Canada and the US, where prescriptions are already required, the regulations and legislation used in Latin America may be a useful model for many other countries, such as those in Africa, the Eastern Mediterranean, South East Asia, and the Western Pacific that currently allow over-the-counter sales of antibiotics, and where overuse of antimicrobials is likely to decrease in response to this restriction.
Similarly, we identified several policies that used electronic medical records and national health insurance systems to change physician and patient behaviours around antimicrobial use. These policies, including restricted reimbursements and reimbursement penalties for patients and prescribers, were used in high-income jurisdictions such as Canada, Sweden, and Taiwan. These policies take a different approach, targeting overuse through restrictive and coercive financial mechanisms.
Given the complexity of AMR, and the need to balance conservation of antimicrobial effectiveness with ensuring access to appropriate antimicrobials for those who need them [2], there is unlikely to be a “silver bullet” intervention that solves the global AMR problem. These 17 government policy interventions offer a starting point for countries to adapt to their local context. Since most of these 17 policies have been evaluated only once or twice and in particular contexts, it would be unwise to draw strong conclusions about their effectiveness. Indeed, many of these interventions were evaluated using low-quality, non-randomized designs; while many systematic reviews would exclude studies on this basis, we retained them in our evidence map to ensure that we captured the widest range of policy options possible. To avoid future waste of public resources, and in line with WHO recommendations for national action on AMR [93], governments should ensure that AMR policy interventions are evaluated using rigorous study designs and that study results are published.
Not surprisingly, our evidence map found that public awareness campaigns and guidelines were commonly used strategies for reducing antimicrobial use across all regions. These educational approaches are traditional public health strategies and have been promoted by both WHO and the UK’s Review on Antimicrobial Resistance [13]. While launched at the government level, many of these programs and policies still focus on changing the practice of individual prescribers, usually physicians, rather than targeting other healthcare professionals or altering healthcare structures to reduce overuse and misuse of antibiotics. Different governments have different policy levers at their disposal, including the ability to implement complex regulatory, legislative, fiscal, and service provision policies, which could potentially bring about more dramatic change than policies focused on individual prescriber behaviour change. Many approaches to reducing antimicrobial consumption can only be implemented by governments, including many of the policies we identified (e.g., professional regulation, restricted reimbursement, and prescription requirements) as well as policies identified by others in academic literature for which we did not find any evaluations (e.g., creating human-only classes of antimicrobials [1], banning direct-to-consumer advertising [3,6], and using tax or fiscal measures [94]). Given that governments can employ a broad range of policy options beyond public awareness campaigns and guidelines, the full range of possible AMR policies should be further explored.
Our evidence map represents the first systematic effort, to our knowledge, to identify government policy interventions and specific policy mechanisms for reducing human antimicrobial use. We worked with 3 research librarians from 3 disciplines and contacted experts around the world to identify published and grey literature on government and AMR policies. However, we recognize that there are implemented policies (e.g., [95]) that have not been captured in this evidence map. We suspect that these studies have not been evaluated, or have not been evaluated with respect to antimicrobial use, or the results of these studies have not been made public.
As with many studies about AMR, we were unable to directly investigate the human health impact of government action on AMR due to the complex relationships among AMR, the use of antimicrobials in humans, animals, and agriculture, and health outcomes. This complexity will continue to be a challenge for AMR research until such a time as “one health” monitoring systems for both antimicrobial use and AMR improve. However, reductions in antimicrobial use are a more immediate measure of policy impact, and large-scale reductions in antimicrobial use are likely to lead to lower levels of resistance [96]. Reducing antimicrobial use is therefore a valuable target for policy makers tackling AMR at the population level.
Our identification of 17 different policy strategies for reducing human antimicrobial use suggests that governments have a variety of policy options at their disposal for mitigating AMR. However, we also note that most existing policy options have not been rigorously evaluated, and some commonly discussed policy options have not been evaluated for their impact on antimicrobial use. To avoid wasting public resources, governments should ensure that future AMR policy interventions are evaluated using rigorous study designs and that study results are published.
|
10.1371/journal.ppat.1000629 | Bacterial Porin Disrupts Mitochondrial Membrane Potential and Sensitizes Host Cells to Apoptosis | The bacterial PorB porin, an ATP-binding β-barrel protein of pathogenic Neisseria gonorrhoeae, triggers host cell apoptosis by an unknown mechanism. PorB is targeted to and imported by host cell mitochondria, causing the breakdown of the mitochondrial membrane potential (ΔΨm). Here, we show that PorB induces the condensation of the mitochondrial matrix and the loss of cristae structures, sensitizing cells to the induction of apoptosis via signaling pathways activated by BH3-only proteins. PorB is imported into mitochondria through the general translocase TOM but, unexpectedly, is not recognized by the SAM sorting machinery, usually required for the assembly of β-barrel proteins in the mitochondrial outer membrane. PorB integrates into the mitochondrial inner membrane, leading to the breakdown of ΔΨm. The PorB channel is regulated by nucleotides and an isogenic PorB mutant defective in ATP-binding failed to induce ΔΨm loss and apoptosis, demonstrating that dissipation of ΔΨm is a requirement for cell death caused by neisserial infection.
| PorB is a bacterial porin that plays an important role in the pathogenicity of Neisseria gonorrhoeae. Upon infection with these bacteria, PorB is transported into mitochondria of infected cells, causing the loss of mitochondrial membrane potential and eventually leading to apoptotic cell death. Here, we show that PorB enters mitochondria through the TOM complex, similar to other mitochondria-targeted proteins, but then bypasses the SAM complex machinery that assembles all other porin-like proteins into the outer mitochondrial membrane. This leads to the accumulation of PorB in the intermembrane space and the integration of a fraction of PorB into the inner mitochondrial membrane (IMM). In the IMM, ATP-regulated pores are formed, leading to dissipation of membrane potential and the loss of cristae structure in affected mitochondria, the necessary first steps in induction of apoptosis. Our work offers, for the first time, a detailed analysis of the mechanism by which PorB targets and damages host cell mitochondria.
| The genus Neisseria is comprised of the human pathogenic species N. gonorrhoeae (Ngo) and N. meningitidis, which cause gonorrhea and meningitis, respectively. The attachment of bacteria to epithelial cells results in transfer of the outer membrane porin PorB to the host cell cytoplasmic membrane [1],[2] and mitochondria [3],[4]. Infection by Ngo causes loss of membrane potential (ΔΨm) across the inner mitochondrial membrane (IMM) and release of cytochrome c, which is required for activation of caspases and induction of apoptosis [5]. When expressed in host cells, PorB translocates to mitochondria and efficiently causes the breakdown of ΔΨm, but fails to induce the release of cytochrome c and subsequent apoptosis under these conditions [6]. This suggested that PorB is required, but is not sufficient, to induce apoptosis, and that a second signal is needed to induce cytochrome c release during infection.
Pro-apoptotic signals like growth factor withdrawal, DNA damage or cytoskeletal rearrangement lead to the activation of so-called BH3-only proteins, pro-apoptotic members of the Bcl-2 family [7]. Active BH3-only proteins cause the oligomerization and pore formation of Bax and Bak in the outer mitochondrial membrane (OMM) [8],[9]. We recently demonstrated that signaling cascades originating from the initial interaction of gonococci with host cells specifically induce the release of the cytoskeletal associated proteins Bim and Bmf, which are both required for the full induction of apoptosis by gonococcal infection [10]. Bim and Bmf activate proapoptotic Bak and Bax proteins, inducing OMM perforation followed by the release of caspase-activating factors into the cytosol and activation of apoptosis [11]. Thus, Bim- and Bmf-initiated events may act in cooperation with mitochondrial PorB in apoptosis induction.
Whereas targeting of PorB to mitochondria and its crucial role in Ngo-induced apoptosis are well established [3],[6], the molecular mechanism by which PorB causes loss of ΔΨm remains unknown. Studies with yeast mitochondria have indicated that import of PorB, and also other bacterial porins, into mitochondria might follow the same pathway as the endogenous mitochondrial porin, voltage-dependent anion-selective channel (VDAC) [6],[12]. In addition, as bacterial PorB and mitochondrial VDAC are both classical β-barrel proteins, structural similarities may facilitate recognition of bacterial PorB by the mitochondrial protein import machinery.
In general, uptake of newly synthesized proteins from the cytosol is mediated by the TOM complex, the translocase of the mitochondrial outer membrane [13]. Within this complex, the general import pore is formed by Tom40 [14]. At the OMM of yeast mitochondria, both endogenous VDAC [15] and a PorB derivative [6] target Tom40 to enter mitochondria. VDAC and all other mitochondrial β-barrel proteins tested so far are subsequently transferred to the SAM/TOB complex (sorting and assembly machinery) in the mitochondrial outer membrane [16],[17]. Interestingly, the core component of this complex Sam50/Tob55, shows remarkable similarities to the bacterial outer membrane protein Omp85 [18]; Omp85 has been shown to mediate membrane insertion of PorB and other β-barrel proteins in Neisseria meningitidis [19],[20].
Considering the obvious homologies between the Omp85 family members [18], bacterial PorB should be recognized and inserted into the OMM by the SAM/TOB complex. This would also be in agreement with the general rule that β-barrel proteins are found neither in bacterial nor in mitochondrial inner membranes. However, if PorB accumulates in the OMM, it is difficult to explain how it dissipates ΔΨm, since this would require massive ion flux across the IMM.
Here, we investigated the role of mitochondrial targeting of PorB during the course of infection-induced apoptosis. Our data demonstrate that the cooperation of PorB and signaling pathways activated by BH3-only proteins induces the release of cytochrome c and the activation of caspases. We show that PorB avoids Sam50/Tob55 and Sam37/Mas37, two core components of the SAM/TOB complex, to integrate into the IMM. As a result, mitochondria lose their ΔΨm and the structural integrity of the cristae is dramatically altered. We propose that these modifications at the IMM are essential early events in Ngo-induced apoptosis.
Our previous observations that PorB of pathogenic Neisseria efficiently targeted mitochondria and induced ΔΨm loss without triggering the release of cytochrome c ([3] and Fig. 1A), confirmed that these were independent processes, at least in our model. Therefore, we reasoned that PorB-expressing HeLa cells lack signals upstream of mitochondria to efficiently release cytochrome c. Such potential upstream signals are mediated by BH3-only proteins, which we recently identified as necessary factors for gonococci-induced apoptosis [10]. To demonstrate the interplay of PorB-triggered ΔΨm loss and BH3-only protein-induced pathways for the release of proapoptotic factors, we activated Bak in HeLa cells using the cell permeable BH3-mimetic compound BH3I-2. BH3I-2 activates Bak (Fig. 1B) by interfering with its binding to Bcl-XL [21]. We then transfected HeLa cells with the porB gene of strain VPI (PorBNgo) and monitored PorB expression, ΔΨm and cytochrome c distribution, using immunofluoresence microscopy. As a control, PorBNmu of the commensal strain N. mucosa, which does not target mitochondria [6], was expressed. The mitochondria of cells expressing PorBNmu were unchanged in comparison to control cells; in contrast, those expressing PorBNgo lost ΔΨm but stained positive for cytochrome c (Fig. 1A), as previously described [6]. Addition of BH3I-2 had no effect on either the membrane potential or the cytochrome c content of mitochondria of PorBNmu-expressing cells (Fig. 1A). Only PorBNgo-expressing cells stained negative for cytochrome c in the presence of BH3I-2 (Fig. 1A), suggesting that mitochondrial targeting of PorBNgo sensitizes cells for the complete release of cytochrome c upon Bak activation.
To test whether BH3I-2 treatment induces caspase cleavage and activation in PorBNgo-expressing cells, western blot analysis was performed to detect active caspase 3. Caspase 3 remained inactivated in cells treated only with BH3I-2 (Fig. 1C, lane 3). PorBNgo expression alone resulted in minimal caspase activity; however, upon addition of BH3I-2 a sharp increase in activity was elicited (Fig. 1C, lanes 7 and 8). Interestingly, dissipation of ΔΨm by treatment of cells with the uncoupling reagent CCCP (carbonyl cyanide m-chlorophenyl hydrazone) or Antimycin A, an inhibitor of the electron transport chain, also caused activation of caspases when combined with BH3-I2 treatment (Fig. 1C, lanes 4 and 6), although to a lesser extent. These results suggested that PorBNgo-induced ΔΨm loss facilitates the release of cytochrome c, which in turn activates caspase-3.
The release of cytochrome c during apoptosis is a multi-step process that requires a complete remodeling of the IMM [22]. Loss of cristae structure and subsequent condensation of the mitochondrial matrix is often detected in mitochondria devoid of ΔΨm [23],[24]. Accordingly, we tested whether infection and/or PorB expression induce remodeling of mitochondria. Electron microscopy (EM) revealed that most of the mitochondria in infected apoptotic cells were highly condensed and appeared dark (Fig. 2A), a phenomenon explained by an increased electron diffraction of the condensed matrix. When HeLa cells were treated with the caspase inhibitor zVAD prior to infection, at least 60% of the mitochondria from these cells displayed a dark matrix and loss of cristae (Fig. 2A,B). Immunogold-labeling of PorB-transfected cells revealed that most PorB containing mitochondria underwent extensive condensation of the matrix and loss of cristae structure (Fig. 2C,D). We therefore conclude that the presence of PorB triggers a major reorganization of the IMM.
Since induction of ΔΨm loss is most likely the crucial step in the sensitization of mitochondria by PorB and gonococcal infection, we investigated the underlying mechanisms of this process. Since PorB is structurally similar to other β-barrel proteins, it should be recognized by the SAM/TOB complex [25],[26] and insert into the OMM. To test this assumption, we isolated mitochondria from HeLa cells, incubated them with radiolabelled PorB and subsequently subjected these mitochondria to carbonate extraction at different stringencies (pH 10.8 and 11.5) to remove loosely attached PorB. The membrane protein VDAC was present in the pellet at both pH, but the soluble protein Hsp60 was found completely in the supernatant only at more stringent conditions, at pH 11.5. Although the amount of soluble PorB increased with the increase of stringency of carbonate extraction, a fraction of radiolabelled PorB remained associated with membranes even after carbonate extraction at pH 11.5 (Fig. S1A). Upon import into mitochondria, the amount of carbonate-resistant PorB increased with time, but only traces of PorB were detected in the pellet after carbonate-extraction of mock samples containing no mitochondria (Fig. S1B). Thus, carbonate-resistant PorB is not formed independently of mitochondria, but instead is a fully imported and membrane-integrated fraction.
To determine if OMM proteins are involved in the uptake of PorB, we pretreated the isolated mitochondria with trypsin to remove cytosolic domains of OMM proteins and compared the rate of import in pretreated and untreated mitochondria. The amount of carbonate-resistant PorB after import was significantly reduced in trypsin-pretrated mitochondria, reaffirming the dependence of PorB import on the TOM complex (Fig. 3A, Fig. S1E and [6]). Likewise, PorB import was reduced (by 20–30%) in mitochondria of tom40kd-2 cells after 5 days of knockdown induction with Dox (Fig. 3B). Of note, even after knockdown, traces of Tom40 that can function in import are still present in the mitochondria (Fig. 3B and not shown). Importantly, PorB import into mitochondria isolated from pLVTHM cells containing an empty vector was unchanged upon treatment with Dox, excluding a Dox-dependent side effect (not shown). In addition, we have previously shown that import of VDAC into mitochondria isolated from tom40kd-2 cells was reduced, along with other TOM components, including Tom20 and Tom22 [27]. Hence, our findings support the notion that the TOM complex is involved in the import of both VDAC and PorB.
We then addressed the possible role of the SAM complex in the import of PorB using HeLa-derived cell lines with inducible knockdown of Sam50 (sam50kd-2) and Metaxin 2 (mtx2kd-2) [27], a putative mammalian homologue of yeast Sam35. VDAC import into the mitochondria isolated from these cell lines was clearly restricted after the induction of knockdown, confirming our previous data (Fig. S2A,B), [27]) and demonstrating the functional downregulation of Sam50 and Metaxin 2 in these mitochondria. Contrary to our expectations, in Sam50- or Metaxin 2-depleted mitochondria, PorB import was unimpeded (Fig. 3C,D). As observed before [27], in mitochondria with the knockdown of Sam50, levels of Tom40 were likewise reduced in the range of 40–50% (Fig. 3C). However, we did not see any decrease of PorB import into these mitochondria in spite of the reduction of Tom40 amounts. This can be explained by the fact that even a strong Tom40 knockdown of more than 90% affected PorB import only moderately (Fig. 3B); a Tom40 reduction of 40–50% might simply not be sufficient to cause any effects.
Previously published data had suggested that PorB follows the same import route as VDAC [6]. However, our data indicated that the SAM/TOB complex marks a crucial branching point of the pathways. To investigate further, we used a similar approach to most previous studies [25],[26] and used yeast to test whether PorB is a substrate of the SAM complex. PorB transport was monitored in isolated yeast mitochondria using radiolabelled PorB. The protease-protected fraction of PorB was only obtained in the presence of mitochondria (Fig. S1C), ruling out a non-specific aggregation of PorB and confirming import into the mitochondria. As reported previously, the homologous PorB of the non-pathogenic strain N. mucosa was not imported into mitochondria (Fig. S1D; [6]). PorB and other β-barrel protein import efficiencies into mitochondria were then monitored. Similar to our findings with human mitochondria (Fig. 3B,C,D), PorB import efficiencies were comparable to WT levels in the bacterial mutant sam50-1 strain (Fig. 3E), whereas in the yeast mutant tom40-4 import was reduced to levels reported for the mitochondrial porin VDAC [15] (Fig. S1E). Import of VDAC into sam50-1 mitochondria was impaired (Fig. S2C); in contrast, import of the IMM dicarboxylate carrier (DIC) remained unchanged (Fig. S2D), supporting a specific role for the SAM complex in the transport of OMM β-barrel proteins [17],[28]. Similar results were obtained with mitochondria isolated from a Δsam37 strain lacking the Metaxin 1 homologue Sam37/Mas37 [16],[29] (Fig. 3G, S2E & S2F), an essential factor for the mitochondrial import of β-barrel proteins [16]. Collectively, these findings suggest that PorB, although a β-barrel protein, avoids the SAM complex during import into mitochondria.
To determine the precise intramitochondrial location of PorB, we prepared highly purified outer membrane vesicles from yeast mitochondria [30]. Unexpectedly, PorB was not present in the vesicles but only in the pellet fraction (Fig. S3, lane OMV vs. lane MPF). In fact, using these conditions PorB mainly formed aggregates in the cytosol, compromising further analysis. Therefore, we first imported radiolabelled PorB into mitochondria isolated from the WT strain, and then separated mitochondrial membrane vesicles by sucrose density centrifugation. Imported PorB accumulated partially in a carbonate-resistant form, as was the case in human mitochondria, indicating its integration into the membrane (Fig. S1A and data not shown). Whereas PorB accumulated in higher density fractions of the gradient, the other β-barrel protein, OMM porin VDAC, was found in the upper part of the gradient. The majority of PorB accumulated in the same higher density fraction as the γ-subunit of the IMM ATP synthase (F1γ) (Fig. 4A). This fractionation pattern clearly demonstrated that PorB does not accumulate in the OMM and that PorB is at least partially associated with the IMM; however, a significant fraction of the newly imported PorB seems to form aggregates, as suggested by the presence of PorB in the high density fractions 8, 9, and 10 (Fig. 4A).
Cytosolic PorB aggregates have never been observed in mammalian cells, suggesting that they express PorB in a fully import-competent state in vivo. The immediate dissipation of ΔΨm in HeLa cells shows that PorB enters the mitochondria in an active form [6]. We expressed PorB in HeLa cells and determined its sub-mitochondrial localization by immunogold electron microscopy. The vast majority of the PorB gold particles localized to the IMM and matrix, similar to the endogenous IMM protein Tim23 (Fig. 4B,C). For comparison with an endogenous OMM marker, we included labeling of Tom22, which showed a clearly different distribution (Fig. 4B,C). Taken together, these findings show that PorB avoids interactions with the SAM machinery and is instead directed to the intermembrane space (IMS). The distribution of PorB in HeLa cells indicates that, at least in human mitochondria, PorB preferentially associates with the IMM.
It is known from previous multichannel studies that PorB is able to form pores of high conductance [2]. Whereas single channel data have been reported on N. meningitidis PorB [31], the investigations on gonococcal PorBs were restricted to multichannel recordings. However, only single channel analysis can clarify whether gonococcal PorBs can form a high-conductance channel in the IMM. Upon addition of purified PorB to either side of a planar lipid bilayer with lipid composition corresponding to the IMM [32], single-channel currents were readily detected (Fig. 5A). Although PorB exhibited a dynamic gating behavior with a multitude of conductance states (Fig. 5A & Fig. S4, insert), a main conductance state of Λ = 420±15 pS was also identified from a linear current-voltage relationship under symmetrical buffer conditions (Fig. S4). Previous observations with multi-channel recordings [2] indicated a voltage-dependent gating behavior of PorB, raising the question whether PorB channels remain open or closed under the physiological conditions present in the IMM (i.e. at ΔΨm of around 150 mV [33]). Gating transitions were rarely observed at lower voltages (not shown); PorB channels exhibited a typical three step channel closing at voltages above 60 mV (Fig. 5A). Considering the ΔΨm of the IMM and the voltage dependence of PorB, it is obvious that PorB channels incorporated into the IMM would in fact be arrested in a closed state and the capability of PorB to uncouple the mitochondria would be limited.
Since it is known that PorB shows an affinity for nucleotides [2] we analyzed the effect of ATP on PorB channels in detail. The concentration of ATP in mitochondria is known to range from 0.6 to 6.0 mM [34]. After addition of physiologically relevant amounts of ATP to PorB-containing bilayers, two intriguing effects were observed. First, the amplitude of the single-channel conductance was significantly reduced (Fig. 5A,B,D) accompanied by a drastic change in PorB gating behavior (Fig. 5A,D): The channel did not display the typical three step gating transitions but remained mainly in an open state exhibiting increased gating frequency, manifested as flickering (Fig. 5D). The second and physiologically even more important effect was the predominant loss of the voltage-dependent closure of PorB upon ATP addition (Fig. 5C). In intact mitochondria, where ATP is abundant, the channels would be forced to stay open even at a ΔΨm of 150 mV, allowing the flux of large currents across the IMM and subsequent rapid dissipation of ΔΨm. The effects were verified by adding ATP to both sides of the membrane (Fig. 5B,C, left-most, centre-left) or separately to either the trans or the cis compartment, to test for side specificity (Fig. 5B,C, centre-right and rightmost). Since the usual channel closure is inhibited at the side of ATP addition and we do not know the exact orientation in which PorB incorporates into the membrane, it was important to establish that the effect of ATP on the PorB channel was not side-specific (Fig. 5B,C). These observations on single PorB channels allowed us to draw several important conclusions: (i) PorB inserts easily into a lipid bilayer, regardless of membrane potential. (ii) High concentrations of mitochondrial ATP stabilize the PorB channels in an open state. (iii) Considering a typical size and a simplified volume to surface ratio, a single open PorB channel should dissipate ΔΨm in about 0.8 ms (for details see Protocol S1).
Next, we tested if PorB can perforate the IMM in living cells using a calcein quenching assay for infected and porB transfected cells (for details see Protocol S1). Transfection of PorB (Fig. 5E) or infection with Ngo (Fig. 5F & S5) led to the loss of calcein staining after cobalt chloride quenching, confirming IMM permeabilization upon PorB translocation. IMM permeabilization was not a consequence of apoptosis induction since preincubation of the cells with the caspase inhibitor zVAD-fmk failed to prevent IMM permeabilization or the loss of ΔΨm upon Ngo infection (Fig. 5F & S5). In conclusion, these data confirm that PorB is able to form pores in the IMM, leading to its permeabilization and loss of ΔΨm.
We have previously shown that PorB binds ATP via lysine residues potentially located in the PorB channel or the loop 3 region [2]. To identify the specific residues involved, we performed site-directed mutagenesis of several lysine residues to generate PorB derivatives deficient in ATP-binding. Mutant neisserial strains, differing only in the PorB derivative expressed [35], were tested for ATP binding as described previously [2]. Even though the exchange of lysine 98 for glutamine (PorBK98Q) reduced ATP-binding by more than 70% in comparison to WT PorB (PorBNgo) (Fig. 6A), PorBK98Q still clearly localized to mitochondria in transfected cells (Fig. S6). Interestingly, ATP did not affect the channel properties of PorBK98Q. Neither the voltage-dependent gating of the PorBK98Q channel (Fig. 6B,C) nor the voltage-dependent open probability (Fig. 6D,E) was influenced by ATP, confirming that the effects of ATP on the PorB channel depend on lysine residue 98.
To determine the effect of the K98Q mutation in vivo, we infected HeLa cells with isogenic neisserial strains harboring WT PorBNgo (N920) and mutant PorBK98Q (N886) and then measured the loss of ΔΨm by using tetramethylrhodamine ethyl ester perchlorate (TMRE) staining and FACS analysis. Despite both strains having similar infection rates, significantly more cells infected with strain N886 carrying mutant PorBK98Q retained their ΔΨm as compared to cells infected with WT neisserial strain (Fig. 7A,B). Consistently, cells transfected with the PorBK98Q expression construct retained their ΔΨm in a number of cases (Fig. S6). Maintenance of ΔΨm was never observed with the WT PorB construct. Moreover, the potential of PorBK98Q mutant strain N886 to induce apoptosis in HeLa cells was strongly reduced (Fig. 7C) and infected cells failed to release cytochrome c from their mitochondria (Fig. 7D). When cells were transfected with WT PorBNgo prior to infection with N886, they released cytochrome c (Fig. 7D), confirming that induction of ΔΨm loss and a second unknown signal triggered by N. gonorrhoeae infection are required for the induction of host cell apoptosis.
In this study, we demonstrate that PorB induces the reorganization of mitochondrial cristae and sensitizes mitochondria to release cytochrome c in response to infection and BH3-only protein induced signaling pathways. The initial prerequisite for PorB to sensitize infected cells to apoptosis is its ability to cause ΔΨm loss, probably by bypassing the OMM-SAM complex and accumulating in the intermembrane space.
Our previous data suggested that release of pro-apoptogenic factors by mitochondria during apoptosis induced by gonococcal infection requires two independent steps: The sensitization of mitochondria and the perforation of the OMM. The cooperative effect of PorB at the mitochondria and separate signals elicited by infected cells could be demonstrated in PorB-expressing cells during infection with strain N886, which carries the ATP-binding mutant of PorB, PorBK98Q (Fig. 7D). Our previous data also suggested that the infection-induced perforation of the OMM depends on the BH3-only proteins Bim and Bmf and the pro-apoptotic BH1-3 proteins Bak and Bax, activated by the interaction of the pathogen with host cells [10],[11]. The data presented here now demonstrate that mitochondria-targeted PorB in combination with a BH3-only mimetic compound is sufficient to induce cytochrome c release and the activation of caspases. This statement is supported by our observation that (i) only PorBNgo targets the mitochondria and not the targeting-deficient derivative PorBNmu, and (ii) only PorBNgo recombinant cells and not the non-recombinant neighboring cells responded to BH3I-2 treatment. It is interesting to note that PorB from meningococcal strain H44/76 overexpressed in HeLa cells fails to induce ΔΨm dissipation but instead protects against apoptosis [36], whereas PorB from strain Z2491 has a similar uncoupling activity as gonococcal PorB [6]. Recent data suggest that invasive and carrier strains of N. meningitides from the same clonal complex induce or inhibit apoptosis, respectively. PorB from the invasive isolate promotes apoptosis induction [37], supporting a role for PorB in the life and death decision of Neisseria-infected cells.
As shown before [6], the amounts of PorB are relatively similar in mitochondria isolated from Neisseria-infected cells and cells transfected with PorB. However, the efficiency of transfection is approximately 30–40%, whereas the efficiency of infection is nearly 100%. Therefore, during transfection we estimate that 2- to 3-fold more PorB is present in a cell than during infection. Nevertheless, we observe the similar phenotype, a major reconstruction of cristae structures, in both infected and PorB recombinant cells. In this aspect, the transfection and infection model seem to be comparable.
The reconstitution of cristae structures is required for sensitization of mitochondria for the release of cytochrome c. Previous work has shown that the majority of cytochrome c is sequestered in the closed cristal compartments (85%); as a result structural changes in mitochondria are required to achieve complete and rapid release of cytochrome c [22]. Moreover, ΔΨm loss and matrix condensation contribute to cytochrome c release [24].
In the context of the biogenesis of endogenous mitochondrial proteins, it is not surprising that we find that the uptake of PorB into mitochondria is mediated by the TOM complex. This complex is involved in the import of practically all mitochondrial preproteins from the cytosol [38]–[40]. However, our observation that the import of PorB is independent from the SAM/TOB complex was unexpected. All previously tested endogenous mitochondrial β-barrel proteins, including VDAC, Tom40, Sam50/Tob55, and Mdm10, were transferred to this complex and subsequently inserted into the OMM [25],[41]. Only recently, subunit Sam35 of the SAM/TOB complex was shown to recognize a specific sorting signal in the C-terminal part of mitochondrial β-barrel proteins and to initiate the insertion of these proteins into the OMM [42]. PorB is obviously lacking this sorting signal and is therefore not recognized as a substrate by the SAM complex. Interestingly, C-terminal sequences were also reported to be relevant in the sorting of β-barrel proteins in bacteria [43]. Our first attempt at directing PorB to the OMM by introducing a C-terminal segment of VDAC failed (data not shown), probably due to requirements in the tertiary structure of the protein, as shown previously for PorB [6].
Purified PorB spontaneously integrated into liposomes consisting of a broad variety of different lipids, from both the outer and inner mitochondrial membranes. Insertion of PorB into the lipid phase is accompanied by specific conformational changes (see Protocol S1 and Fig. S7) and requires neither additional assembly factors nor energy sources. A similar finding has been reported for the integration of soluble human VDAC into lipid bilayers [44]. Previously, we found significant amounts of PorB were associated with the OMM upon overexpression of the protein in yeast [6]; however, here we found that highly purified OMM vesicles are devoid of PorB. In HeLa cells, PorB shows the similar distribution pattern as the IMM protein Tim23. Interestingly, the secretin PulD, a β-barrel protein in the Gram-negative bacterium Klebsiella oxytoca, accumulates in the bacterial plasma membrane if assembly in the outer membrane is blocked [45]. We propose a similar mechanism for PorB: PorB trapped in the mitochondrial intermembrane space has the possibility to integrate into the IMM.
Since most proapoptotic factors of the mitochondria are stored in the IMS, recent investigations have concentrated on the question of how their release is triggered by the opening of the OMM. However, we found that during Ngo-induced apoptosis, events at the OMM are preceded by essential modifications at the IMM. Considering a typical size and a simplified volume to surface ratio, a single PorB channel would dissipate ΔΨm in about 0.8 ms (for details see Protocol S1). Thus, a single open PorB pore in the IMM is sufficient to short-circuit a whole mitochondrion. At potentials Vm≥±60 mV, the PorB channel is enclosed in membranes with a lipid composition comparable to that of the IMM, but ATP binding arrests the channel in an open state even at voltages physiological for the IMM of about 150 mV. In previous experiments with time-averaged multichannel recordings, we observed an apparent ATP-dependent decrease in the voltage sensitivity of PorB channels [2]; however, the time-averaged analysis of these multichannel recordings did not allow the resolution of single-gating events. Here, our single-channel analyses unequivocally showed that ATP abolishes the voltage-dependent closure of PorB channels. In summary, following insertion into the IMM, PorB is arrested in an open state by ATP at the prevailing ΔΨm, i.e. by the normal physiological activity of the mitochondria.
The attenuated phenotype of the neisserial strain N886 expressing PorBK98Q provided clear in vivo evidence for a crucial role of ATP binding in dissipating ΔΨm and inducing apoptosis during infection. Our recent data suggest that cell death induced by Ngo requires the close interplay of two interdependent signaling cascades, one leading to the activation of Bak [11] and the other to the sensitization of the mitochondria by translocated PorB (Fig. 8). For the latter, PorB utilizes an existing mitochondrial transport pathway, the TOM complex, but bypasses the SAM complex, and activated by ATP dissipates ΔΨm. This intriguing and so far unique example is a consequence of the evolution of the SAM machinery. SAM has diverted so much from the parental Omp85 machinery that it is now unable to recognize and sort PorB into the OMM. In this way the bacterial effector PorB functions as an integral element of the host's cell death machinery.
N. gonorrhoeae N242 (strain VPI; Opa+, PorBIA) [46] and N920 (strain MS11; Opa+, PorBIA) expressing PorB of N242 [35] have been described. N886 (strain MS11; Opa+, PorBK98Q) was constructed using the same strategy as for N920 but a mutant porB gene of N242 was transformed instead of the wildtype derivative [35]. Gonococci were routinely grown on GC agar base plates (Becton Dickinson, Difco and Remel) supplemented with Proteose Pepton Nr. 3 (Difco) and 1% vitamin mix for 14–20 h at 37°C in 5% CO2 in a humidified atmosphere. Opa phenotypes were monitored by colony morphology under a stereo microscope or by immunoblotting. HeLa cells (human cervix carcinoma) were grown in RPMI 1640 (Gibco) supplemented with 10% heat inactivated fetal calf serum (FCS) in the presence of 5% CO2. Cells were seeded 24 h before infection and washed several times with RPMI without supplements. Infections were routinely performed at a multiplicity of infection (MOI) of 1 without centrifugation. For inhibition of caspases, cells were incubated with 50 µM zVAD-fmk (Bachem) for 15 min prior to infection and throughout the respective infection period.
Cells were seeded on coverslips and transfected with pCMV-Tag-1, containing either the porBNgo (P.IA) or porBNmu gene with an N-terminal FLAG-Tag [3], using Lipofectamine 2000 (Invitrogen) according to the manufacturer's protocol. Twenty four hours post-transfection, cells were stained with 150 nM MitoTracker (Molecular Probes), dissolved in cell culture media for 30 min at 37°C, washed with phosphate buffered saline (PBS) and fixed in 3.7% paraformaldehyde (PFA). Fixed cells were permeabilized using 0.2% Triton X-100, and nonspecific binding was blocked by using 1% goat serum. Samples were stained using anti-FLAG (Sigma) antibody, followed by detection with fluorochrome-coupled secondary antibodies (Jackson Immuno Research). Samples were analyzed under a Leica confocal microscope using TCS software.
5×105 cells per sample were harvested in 100 µl loading buffer and 20 µl of the protein lysates were separated by SDS-PAGE and transferred to nitrocellulose or polyvinylidenfluorid (PVDF) membranes. The following antibodies and sera were used in this study: anti-β-Actin (Sigma); anti-Bak NT (Upstate); anti-Bak (Ab-1) (Millipore); anti-Bax NT (Upstate); anti-Bax (6A7) (BD Pharmingen); anti-cleaved Caspase-3 (Cell Signalling); anti-FLAG (Sigma); anti-Hsp60 (Stressgen Bioreagents); anti-VDAC (Abcam); anti-Tim23 (BD Biosciences); antibodies against human Tom40 and Sam50 were a gift from N. J. Hoogenraad, and against mouse Metaxin 2 (cross-reactive with human Metaxin 2) a gift from P. Bornstein. Antibodies against yeast mitochondrial proteins Tom40 and Tim23 were a gift from N. Pfanner and C. Meisinger. Equal loading was routinely confirmed by appropriate loading controls. Quantitative analysis of immunoblots was performed by using the open source software ImageJ (http://rsbweb.nih.gov/ij/index.html).
Trypsin treatment of mitochondria was performed as previously described [47]. To induce the knockdown by RNA interference, cells were grown for 5 to 7 days in the presence of 1 µg/ml doxycycline as previously described [27]. Efficiency of the knockdown was assessed by western blot. For trancription/translation purposes, PorB was cloned into pGEM-4Z vector (Promega) with two additional methionines at its C-terminus and in vitro transcribed/translated in the presence of 35S-methionine/cysteine (GE Healthcare) using the TnT Quick Coupled System (Promega). Mitochondrial isolation and import of proteins were performed essentially as described previously [27]. A detailed protocol is available as Protocol S1.
Mitochondria were isolated from yeast cells as described previously and used for import of 35S-labelled mitochondrial porin (VDAC), dicarboxylate carrier (DIC), or the ATP-synthase γ-subunit (F1γ) following standard procedures [15]. A detailed protocol is available as Protocol S1.
Purified azolectin (60 mg/ml) in n-decan or a lipid mixture corresponding to the lipid composition of inner mitochondrial membranes (60 m/ml) [32] in n-decan was used to produce stable planar lipid bilayers by using the painting technique [48],[49]. Purified PorB was applied directly below the bilayer in the cis chamber. Buffer conditions were symmetrical with 250 mM KCl, 10 mM Mops-Tris (pH 7.0) in the cis/trans compartment. Two Ag/AgCl electrodes covered by 2 M KCl-agar bridges were inserted into each chamber with the trans chamber electrode connected to the headstage (CV-5-1GU) of a Geneclamp 500 current amplifier (Axon Instruments) and this was used as a reference for reported membrane potentials. Current recordings were carried out using a Digidata 1200 A/D converter. Data analysis was performed by self-written Windows-based SCIP (single-channel investigation program) in combination with Origin 7.0 (Microcal Software). Current recordings were performed at a sampling interval of 0.1 ms, filtered with a low-pass-filter at 2 kHz. Voltage ramps were carried out by continuously increasing the voltage at a rate of 5 mV/s.
After incorporation of single PorB channels into the bilayer by spontaneous insertion, control currents were recorded. Subsequently nucleotides were added either to both sides or separately to the cis or trans side of the membrane. Interactions between the added nucleotides and PorB channels were examined after stirring the aqueous solutions on both sides of the planar lipid bilayer. Binding of ATP by PorB was assayed by chemically crosslinking radiolabelled ATP to neisserial strains expressing different PorB derivatives as previously described [2]. ATP binding was quantified using AIDA Image Analyzer software.
For immuno-EM analysis, the cells were fixed with 3% PFA in stabilizing buffer (1 mM EGTA, 4% PEG 6000 or PEG 8000, 100 mM PIPES pH 6.9) and embedded in 10% Gelatine/PBS. Small blocks of the samples were infiltrated overnight in 2.3 M sucrose/0.1 M Na-phosphate buffer. Ultra-thin sections were cut at −120°C with a diamond knife. The sections were transferred onto carbon-coated pioloform-film on TEM-grids. The sections were then blocked and reacted with the primary antibody against FLAG-tag (Sigma), Tom22 (GeneTex), Tom20 (BD Biosciences), Tim23 (BD Transduction Laboratories) and secondary antibodies coupled with 6 or 12 nm gold particles.
For the analysis of mitochondrial membrane potential cells were harvested by trypsinization and washed with phosphate buffered saline (PBS) before staining with 100 nM tetramethylrhodamine ethyl ester perchlorate (TMRE) (Molecular Probes) in growth media at 37°C, 5% CO2 for 30 min. After staining, cells were washed twice with PBS and immediately analyzed by FACS analysis.
|
10.1371/journal.pcbi.1003697 | Collective Behaviour without Collective Order in Wild Swarms of Midges | Collective behaviour is a widespread phenomenon in biology, cutting through a huge span of scales, from cell colonies up to bird flocks and fish schools. The most prominent trait of collective behaviour is the emergence of global order: individuals synchronize their states, giving the stunning impression that the group behaves as one. In many biological systems, though, it is unclear whether global order is present. A paradigmatic case is that of insect swarms, whose erratic movements seem to suggest that group formation is a mere epiphenomenon of the independent interaction of each individual with an external landmark. In these cases, whether or not the group behaves truly collectively is debated. Here, we experimentally study swarms of midges in the field and measure how much the change of direction of one midge affects that of other individuals. We discover that, despite the lack of collective order, swarms display very strong correlations, totally incompatible with models of non-interacting particles. We find that correlation increases sharply with the swarm's density, indicating that the interaction between midges is based on a metric perception mechanism. By means of numerical simulations we demonstrate that such growing correlation is typical of a system close to an ordering transition. Our findings suggest that correlation, rather than order, is the true hallmark of collective behaviour in biological systems.
| Our perception of collective behaviour in biological systems is closely associated to the emergence of order on a group scale. For example, birds within a flock align their directions of motion, giving the stunning impression that the group is just one organism. Large swarms of midges, mosquitoes and flies, however, look quite chaotic and do not exhibit any group ordering. It is therefore unclear whether these systems are true instances of collective behaviour. Here we perform the three dimensional tracking of large swarms of midges in the field and find that swarms display strong collective behaviour despite the absence of collective order. In fact, we discover that the capability of swarms to collectively respond to perturbations is surprisingly large, comparable to that of highly ordered groups of vertebrates.
| Intuition tells us that a system displays collective behaviour when all individuals spontaneously do the same thing, whatever this thing may be. We surely detect collective behaviour when all birds in a flock fly in the same direction and turn at the same time [1], as well as when all spins in a magnet align, giving rise to a macroscopic magnetization [2], [3]. On the other hand, we would not say that there is any collective behaviour going on in a gas, despite the large number of molecules. The concept of collective behaviour seems therefore closely linked to that of emergent collective order, or synchronization. Indeed, explaining how order spontaneously arises from local inter-individual interactions has been one of the major issues in the field [4]–[6].
The case of insect swarms is tricky in this respect. Several species in the vast taxonomic order Diptera (flies, mosquitoes, midges) form big swarms consisting largely of males, whose purpose is to attract females [7], [8]. Swarming therefore has a key reproductive function and, in some cases, relevant health implications, the obvious, but not unique, example being that of the malaria mosquito, Anopheles gambiae [9]–[11]. It is well-known that swarms form in proximity of some visual marker, like a water puddle, or a street lamp [7]. Swarming insects seem to fly independently around the marker, without paying much attention to each other (see Video S1). For this reason, the question of whether swarms behave as truly collective systems is debated [4], [12]. In fact, it has even been suggested that in Diptera there is no interaction between individuals within the swarm and therefore no collective behaviour at all [13], [14]. Although other studies observed local coordination between nearest neighbours [15], [16], it remains controversial whether and to what extent collective patterns emerge over the scale of the whole group. Clarifying this issue is a central goal in swarms containment [17], [18]. In absence of quantitative evidence telling the contrary, the hypothesis that external factors, as the marker, are the sole cause of swarming and that no genuine collective behaviour is present, is by far the simplest explanation.
We must, however, be careful in identifying collective behaviour with collective order. There are systems displaying important collective effects both in their ordered and in their disordered phase. An example is that of a ferromagnet near the critical temperature i.e. the temperature below which a spontaneous magnetization emerges: the collective response of the system to an external perturbation is as strong in the disordered phase slightly above as it is in the ordered phase slightly below In fact, once below the critical temperature, increasing the amount of order lowers the collective response [2], [3]. Similarly, in animal behaviour it is possible to conceive cases in which individuals coordinate their behavioural reactions to environmental stimuli, rather than the behaviours themselves; conversely we may expect that a group too heavily ordered, i.e. with a very large behavioural polarization, may respond poorly to perturbations, because of an excessive behavioural inertia. Hence, although certainly one of its most visible manifestations, emergent order is not necessarily the best probe of collective behaviour.
The crucial task for living groups is not simply to achieve an ordered state, but to respond collectively to the environmental stimuli. For this to happen, correlation must be large, namely individuals must be able to influence each other's behavioural changes on a group scale. The question then arises of whether correlation in biological systems is a consequence of collective order or whether it can be sustained even in absence of order. The question is relevant because the way individuals in a group synchronize their behavioural fluctuations (correlation) is possibly a more general mechanism than the synchronization of behaviour itself (order). All experimental studies performed up to now, however, concerned highly synchronized groups (as bird flocks, fish shoals and marching locusts [19]–[21]), which displayed both order and correlation. Hence, the question of whether or not order and correlations are two sides of the same phenomenon remained open until now. Here, we attempt to give an answer to this question by experimentally studying large swarms of insects in the field. As we will show, despite the lack of collective order, we do find strong correlations, indicating that in biological systems collective behaviour and group-level coordination do not require order to be sustained.
We perform an experimental study of swarms of wild midges in the field. Midges are small non-biting flies belonging to the order Diptera, suborder Nematocera (Diptera:Chironomidae and Diptera:Ceratopogonidae - see Methods). The body length of the species we study is in the range – Swarms are found at sunset, in the urban parks of Rome, typically near stagnant water. As noted before [7], we find that swarms form above natural or artificial landmarks. Moving the landmark leads to an overall displacement of the swarm. The swarms we studied range in size between and individuals (see Table S1 in Text S1).
To reconstruct the 3d trajectories of individual insects we use three synchronized cameras shooting at frames-per-seconds (trifocal technique – Fig. 1e and Methods). Our apparatus does not perturb the swarms in any way. The technique is similar to the one we used for starling flocks [22], with one notable difference. To reach the desired experimental accuracy we need to know the mutual geometric relations between the three cameras very accurately. In the case of flocks, this could be achieved only by an a priori alignment of the cameras. In the case of swarms, though, we proceed differently. After each swarm acquisition, we pin down the geometry of the camera system by taking multiple images of a calibrated target (Fig. 1f). This procedure is so accurate that the error in the 3d reconstruction is dominated by the image segmentation error due to the pixel resolution. If we assume this to be equal to pixel (typically it is smaller than that because midges occupy many pixels), we make an error of in the determination of the distance between two points apart from each other (a reference value for nearest neighbour distance). The absolute error is the same for more distant points, making the relative precision of our apparatus even higher. This accuracy makes the determination of the correlation functions we study here very reliable.
The -tracking of each midge is performed by using the recursive global optimization method described in [23]. This recursive algorithm dramatically reduces the complexity of the tracking problem, effectively overcoming the limit of other tracking methods [24], [25], and allowing the reconstruction of large swarms, up to midges, for long time, up to frames. Sample reconstructions are shown in Fig. 1b and in Video S2. Compared to previous field [11], [15], [26] and lab [27]–[29] studies, data collected and analysed in the present work have the advantage to span among swarms of different sizes and densities.
Swarms are in a disordered phase. The standard order parameter normally used in collective behaviour is the polarization, where is the number of midges in the swarm and is the velocity of insect The polarization measures the degree of alignment of the directions of motion; it is a positive quantity and its maximum value is The average polarization over all swarms is quite small, (see Fig. 2 and Table S1 in Text S1). As a reference, in starling flocks we find on average [19]. The probability distributions of the polarization fully confirms the swarms' lack of translational order and the stark difference with flocks (Fig. 2). Clearly, swarms are not in a polarized state. Translation is not the only possible collective mode, though. For example, it is well-known that fish schools can produce rotating (milling) configurations. Moreover, a group can expand/contract, giving rise to dilatational (or pulsive) collective modes. For this reason we have defined and measured also a rotational and a dilatational order parameter (see Methods). We find, however, that these quantities too have very small values (Fig. 2). The time series, on the other hand, show that the order parameters can have rare, but strong fluctuations, during which their value may become significantly larger than that of an uncorrelated system (Fig. 2). These large fluctuations are a first hint that non-trivial correlations are present.
The connected correlation function measures to what extent the change in behaviour of individual is correlated to that of individual at distance Correlation is the most accessible sign of the presence of interaction between the members of a group. The absence of interaction implies the absence of correlation. Conversely, the presence of correlation implies the presence of an effective interaction (see Text S1, Section I). Correlation can be measured for different quantities, but in the case of midges, as with birds and other moving animals, the principal quantity of interest is the direction of motion. To compute the connected correlation we first need to introduce the velocity fluctuations, namely the individual velocity subtracted of the overall motion of the group, (see Methods for the detailed definition of and ). This fluctuation is a dimensional quantity, hence it is unsuitable to compare the correlation in natural vs numerical systems, as we shall do later on. We therefore introduce the dimensionless velocity fluctuation,(1)
The connected correlation function is then given by,(2)where if and zero otherwise, and is the space binning factor. The form of in natural swarms is reported in Fig. 3: at short distances there is strong positive correlation, indicating that midges tend to align their velocity fluctuations to that of their neighbours. After some negative correlation at intermediate distances, relaxes to no correlation for large distances. This qualitative form is quite typical of all species analysed (see Fig. 3). The smallest value of the distance where crosses zero is the correlation length, that is an estimate of the length scale over which the velocity fluctuations are correlated [19]. The average value of this correlation length over all analysed swarms is, This value is about times larger than the nearest neighbours distance, whose average over all swarms is (see Fig. 3 and Table S1 in Text S1). Previous works noticed the existence of pairing manoeuvres and flight-path coordination between nearest neighbours insects [4], [15], [16]. Our results, however, indicate that midges within a natural swarm influence each other's motion far beyond their nearest neighbours.
The collective response of the swarm depends crucially on two factors: how distant in space the behavioural change of one insect affects that of another insect (spatial span of the correlation) and how strong this effect is (intensity of the correlation). To combine these two factors in one single observable we calculate the cumulative correlation up to scale ,(3)where is the Heaviside function,
It can be shown (see Text S1, Section II) that this dimensionless function is related to the space integral of the correlation function Hence, reaches a maximum where vanishes, i.e. for (see Fig. 3). This maximum, is a measure of the total amount of correlation present in the system. In statistical physics is exactly equal to the susceptibility, namely the response of the system to an external perturbation [30], [31]. In collective animal behaviour, we do not have a quantitative link between integrated correlation and response, so that calling susceptibility is not strictly correct. However, if the probability distribution of the velocities is stationary, we can follow a maximum entropy approach [32] and still find that the total amount of correlation in the system, is related to the way the group responds collectively to a perturbation (see Text S1, Section II). The value of for midge swarms is reported in Fig. 3.
In order to judge how significant is the correlation function and how large is the susceptibility in natural swarms, we need an effective zero for these quantities, i.e. some null hypothesis baseline. As we have seen in the Introduction, the minimal assumption is that all individuals in the swarm interact with an external landmark independently from each other. Following Okubo [4] (but see also [27] and [16]), we therefore simulate a ‘swarm’ of non-interacting particles performing a random walk in a three-dimensional harmonic potential (see Methods). Visually, the group behaviour of this Non-interacting Harmonic Swarm (NHS) looks remarkably similar to that of a real swarm (see Video S2 and S3): all ‘midges’ fly around the marker and the group lacks collective order.
This similarity, however, is deceptive. In the NHS, the correlation function simply fluctuates around zero, with no spatial span, nor structure (Fig. 3). Moreover, the susceptibility in the NHS is extremely small, whereas the susceptibility in natural swarms is up to times larger than this non-interacting benchmark (Fig. 3). We conclude that swarming behaviour is not the mere epiphenomenon of the independent response of each insect with the marker. Despite the lack of collective order, natural swarms are strongly correlated on large length scales. There are big clusters of midges that move coherently, contributing to the ‘dancing’ visual effect of the swarm. The only way this can happen is that midges interact. What kind of interaction is that?
To understand the nature of the interaction, we study the susceptibility across swarms of different densities. Interestingly, we find that increases when the average nearest neighbour distance, decreases (Fig. 4). Denser swarms are more correlated than sparser ones. This result indicates that midges interact through a metric perceptive apparatus: the strength of the perception decreases with the distance, so that when midges are further apart from each other (larger ) the interaction is weaker and the susceptibility is lower. This is at variance with what happens in starling flocks: starlings interact with a fixed number of neighbours, irrespective of their nearest neighbour distance [33]; such kind of topological interaction does not depend on the group density, hence the susceptibility does not depend on the nearest neighbour distance. Fig. 4, on the other hand, shows that midges interact metrically, namely with all neighbours within a fixed metric range, Hence, in swarms the number of interacting neighbours increases with decreasing (increasing density), and as a consequence of this increased amount interaction, the system becomes also more correlated.
In a system ruled by metric interaction we expect all lengths to be measured in units of the perception range, This implies that the natural variable for the susceptibility is the rescaled nearest neighbour distance, The problem is that we are considering different species of midges, likely to have different metric perception ranges. The simplest hypothesis we can make is that is proportional to the insect body length (which we can measure), so that This hypothesis is confirmed by the data: the susceptibility is significantly more correlated to the variable (P-value ) than to (P-value - see Methods for the definition of P-value). The fact that the natural variable is is a further indication that the interaction in swarms is based on a metric mechanism.
The difference in the nature of the interaction between flocking birds and swarming midges (topological vs. metric) is possibly due to the significant differences between vertebrates and arthropods. Topological interaction, namely tracking a fixed number of neighbours irrespective of their distance, requires a level of cognitive elaboration of the information [33] more sophisticated than a metric interaction, where the decay of the effective force is merely ruled by the physical attenuation of the signal with increasing metric distance. In other words, within a metric mechanism the range of the interaction is fixed by a perceptive cut-off, rather than a cognitive one. Metric interaction is known to be more fragile than topological one against external perturbations [33], and indeed it is far more likely to observe the dispersion of a swarm in the field than that of a flock. This may be the reason why the presence of an external marker is crucial for the swarming behaviour of midges [13].
The experimental observations of a non-trivial connected correlation and of a large susceptibility indicate that midges are effectively interacting with each other by acting on their directions of motion. This does not exclude, of course, that other types of interaction are present. First of all, the empirical observation that the swarm uses a visual marker as a reference for maintaining its mean spatial position, strongly suggests that each individual interacts with the marker. Besides, it is certainly possible that effective positional attraction-repulsion forces between midges, as those described in [34], exist. However, the directional correlations indicate that insects are also effectively interacting by adjusting their velocities. Moreover, the fact that these correlations are positive for short distances means that midges tend to align their direction of motion. This fact may seem surprising, because alignment interactions typically lead to the formation of ordered (polarized) groups, which is clearly not the case for midges. Swarms are disordered, and yet interacting and highly correlated systems. Is this a paradox?
In fact, it is not. An alignment interaction does not per se lead to global order in the group. In all models where imitation of the neighbours is present, the onset of long-range order depends on the value of some key tuning parameter. In a ferromagnet, this parameter is the temperature namely the amount of noise affecting the interaction between the neighbouring spins. At high temperature the system is in a disordered state, whereas by lowering one reaches a critical temperature below which an ordering transition occurs. In models of active matter there is another parameter tuning the transition between disorder and order, that is density or, equivalently, nearest neighbour distance: the system gets ordered once the nearest neighbour distance falls below some transition value. The crucial point is that, in general, the correlation of the system tends to be very large around the transition point, irrespective of whether the system is in the ordered or in the disordered phase. Hence, even a disordered system can display large correlations, provided that it is not too far from an ordering transition. In what follows, we want to show that this is indeed what happens with midge swarms.
The simplest model based on alignment interaction that predicts an order-disorder transition on changing the density is the Vicsek model of collective motion [35]. In this model each individual tends to align its direction of motion to that of the neighbours within a metric perception range, The rescaled nearest neighbour distance, is the control parameter: for low noise, the model predicts a transition from a disordered phase (low polarization) at high values of (low density), to an ordered phase (large polarization) at low values of (high density) [35]–[37]. We numerically study the Vicsek model in three dimensions. As we have seen, real swarms hold their average position with respect to a marker; to reproduce this behavioural trait we introduce an harmonic attraction force that each individual experiences towards the origin (see Methods). Also in central potential the model displays an ordering transition: at large density, for the system is ordered and it has large polarization (Video S4). On the other hand, the polarization is low in the disordered phase, (Fig. 5). However, the correlation function is non-trivial when is sufficiently close to (Fig. 5), indicating the existence of large clusters of correlated individuals, which can be clearly detected in Video S5. We calculate the susceptibility in the same manner as we did for natural swarms, in the disordered phase, and find a clear increase of on lowering (Fig. 5).
This increase of the susceptibility is coherent with the existence of an ordering transition at It has been shown that, unless is much larger than the values analysed here, the transition in the Vicsek model is characterized by a clear second order phenomenology (the nature of the transition for is still debated - see [37]–[39]). As a consequence, the susceptibility is expected to become very large approaching and to follow the usual scaling relation of critical phenomena [38],(4)
A fit to equation (4) of the 3d-Vicsek data is reported in Fig. 5, giving and a transition point, The reason for the growth of approaching in the Vicsek model is quite intuitive. The model is metric, so that at large namely when the nearest neighbour distance is much larger than the interaction range very few individuals interact with each other, and coordination is small. The smaller becomes, the larger the number of particles within the mutual interaction range, thus promoting the correlation of larger and larger clusters of particles. For this reason the correlation length and the susceptibility grow when the nearest neighbour distance decreases. When approaches its critical value, the coordinated clusters become as large as the whole system, so that the groups orders below
The low order parameter, the non-trivial correlation function, and especially the increase of on decreasing the nearest neighbour distance, are phenomenological traits that the metric Vicsek model shares with natural swarms. We conclude that a system based solely on alignment can be in its disordered phase and yet display large correlations, as midge swarms do. It is interesting to note that by approaching the ordering transition a compound amplification of the correlation occurs: when the nearest neighbour distance, decreases, the spatial span of the correlation, increases, so that the effective perception range in units of nearest neighbour distance, is boosted. We emphasize that we are not quantitatively fitting Vicsek model to our data. Our only aim is to demonstrate a general concept: large correlation and lack of global order can coexist even in the simplest model of nearest neighbours alignment, provided that the system is sufficiently close to an ordering transition.
The consistency between our experimental data and the Vicsek model suggests that an underlying ordering transition could be present in swarms as well. An ordering transition as a function of the density has been indeed observed in laboratory experiments on locusts [21], fish [40] and in observations of oceanic fish shoals [41]. In these cases, both sides (low and high density) of the ordering transition were explored. However, midge swarms in the field are always disordered, living in the low-density/high- side of the transition. Locating a transition point having data on just one side of it, is a risky business. The reason why we want to do this here is because it will allow us to give a rough estimate of the metric range of interaction in midges, which can be compared with other experiments.
If a Vicsek-like ordering transition exist, we can use equation (4) to fit the swarms data for (Fig. 4). As we already mentioned, we do not know the value of the metric perception range, in swarms. Therefore, we use as scaling variable where is the body length. Although the fit works reasonably well (Fig. 4), the scatter in the data is quite large; hence, given the non-linear nature of the fit, it would be unwise to pin down just one value for the parameters, and we rather report confidence intervals. The fit gives a transition point in the range, with an exponent in the range, (larger exponents correspond to lower transition points).
Interestingly, there is an alternative way to locate the ordering transition that does not rely on the fit of Let us establish a link between pairs of insects closer than the perception range and calculate the size of the biggest connected cluster in the network. Given a swarm with nearest neighbour distance the larger the larger this cluster. When exceeds the percolation threshold, a giant cluster of the same order as the group size appears [42]. We calculate the percolation threshold in swarms (Fig. 6 and Methods) and find The crucial point is that varying the perception range at fixed nearest neighbour distance is equivalent to varying at fixed Hence, at fixed there is an equivalent percolation threshold of the nearest neighbour distance, such that for a giant cluster appears. Clearly, It is reasonable to hypothesise that the critical nearest neighbour distance is close to the maximal distance compatible with a connected network, given A sparser network would cause the swarm to lose bulk connectivity. Therefore, given a certain perception range the ordering transition occurs at values of the nearest neighbour distance close to its percolation threshold,
At this point we have two independent (and possibly equally unreliable) estimates of the transition point in natural swarms of midges: the first one in units of body-lengths, the second one in units of interaction range, Putting the two together we finally obtain an estimate of the metric interaction range in units of body-lengths, The body length of the species under consideration is in the range, This implies a perception range of a few centimetres, depending on the species. This crude estimate of the midge interaction range is compatible with the hypothesis that midges interact acoustically. In [43] the male-to-male auditory response in Chironomus annularius (Diptera:Chironomidae) was studied and it was found that the range of the response was about not too far from our estimate. Similar measurements in mosquitoes (Diptera:Culicidae) show that the auditory perception range is about [44], which is again compatible with our determination of the interaction range in midge swarms.
We have shown that natural swarms of midges lack collective order and yet display strong correlations. Such correlations extends spatially much beyond the inter-individual distance, indicating the presence of significant cluster of coordinated individuals. This phenomenology is incompatible with a system of non-interacting particles whose swarming behaviour is solely due to the attraction to an external landmark. We conclude that genuine collective behaviour is present in swarms. We stress that the existence of correlation, and therefore of inter-individual interaction, is not in contradiction with the fact that a swarm almost invariably forms in proximity of a marker. The effect of the marker (external force) is merely to keep the swarm at a stationary position with respect to the environment. However, as we have shown in the case of the non-interacting swarm, this stationarity (which superficially would seems the only visible trait of swarming), cannot by itself produce the observed strong correlations. By using Vicsek model as a simple conceptual framework, we have shown that this coexistence of disorder and correlation is a general feature of systems with alignment interaction close to their ordering transition.
We should be careful in interpreting our data as proof that explicit alignment is the main interaction at work in swarms. What we can say is that non-trivial alignment correlation implies effective alignment interaction. However, how this effective alignment interaction is achieved in terms of sensorimotor processes is hard to tell. In fact, as we have already remarked, it is possible that models purely based on repulsion/attraction positional forces, lead to correlations similar to the ones we reported here. Hence, as always when dealing with animal behaviour, it is important to keep in mind the intrinsically effective nature of any interaction. The Vicsek model provides the simplest and most compelling description of collective behaviour when effective alignment is present and this fact is not hindered by the real, non-effective nature of the interaction giving rise to the observed correlations.
Our results suggest that correlation, rather than order, is the most significant experimental signature of collective behaviour. Correlation is a measure of how much and how far the behavioural change of one individual affects that of other individuals not directly interacting with it. Our data show that in swarms correlations are so strong that the effective perception range of each midge is much larger than the actual interaction range. If the change of behaviour is due to some environmental perturbations, such large correlation guarantees that the stimulus is perceived at a collective level.
A notion of collective behaviour based on correlation is more general and unifying than one based on order. For example, bird flocks and insect swarms look like completely different systems as long as we stick to collective order. However, once we switch to correlation, we understand that this big difference may be deceptive: both flocks and swarms are strongly correlated systems, in which the effective perception range, or correlation length, is far larger than the interaction range [19]. In this perspective, the striking difference in emergent order between the two systems, namely the fact that flocks move around the sky, whereas swarms do not, may be related to different ecological factors, rather than to any fundamental qualitative difference in the way these systems interact. Strong correlations similar to those found in bird flocks and midge swarms have also been experimentally measured in neural assemblies [45]. This huge diversity - birds, insects, neurons - is bewildering but fascinating, and it suggests that correlation may be a universal condition for collective behaviour, bridging the gap between vastly different biological systems.
Data were collected in the field (urban parks of Rome), between May and October, in and in We acquired video sequences using a multi-camera system of three synchronized cameras (IDT-M5) shooting at fps. Two cameras (the stereometric pair) were at a distance between and depending on the swarm and on the environmental constraints. A third camera, placed at a distance of from the first camera was used to solve tracking ambiguities. We used Schneider Xenoplan lenses. Typical exposure parameters: aperture , exposure time Recorded events have a time duration between and seconds. No artificial light was used. To reconstruct the 3d positions and velocities of individual midges we used the techniques developed in [23]. Wind speed was recorded. After each acquisition we captured several midges in the recorded swarm for lab analysis. A summary of all swarms data can be found in Table S1 in Text S1.
We recorded swarms of midges belonging to the family Diptera:Ceratopogonidae (Dasyhelea flavifrons) and Diptera:Chironomidae (Corynoneura scutellata and Cladotanytarsus atridorsum). Midges belonging to the family Chironomidae were identified to species according to [46], the ones belonging to the family Ceratopogonidae were identified according to [47] and [48]. Specimens used for identification were captured with a hand net and fixed in alcohol, cleared and prepared according to [49]. Permanent slides were mounted in Canada Balsam and dissected according to [50]. Species identification was based on morphology of the adult male, considering different characters, as wing venation, antennal ratio (length of apical flagellomere divided by the combined length of the more basal flagellomeres) and genitalia, which in Diptera are named hypopygium (a modified ninth abdominal segment together with the copulatory apparatus - see Fig. 1).
Let be the set of coordinates at time and at the next time step. To simplify the notation we set The velocity vector of insect is defined as, To compute the connected correlation function we need to subtract the contribution of all collective modes from the individual velocity. We identify three collective modes: translation, rotation and dilatation (expansion/contraction).
Translation: Let be the position of the centre of mass, and the position of the -th object in the centre of mass reference frame. By subtracting the centre of mass velocity, from the individual velocity, we obtain the translation-subtracted fluctuation,(5)
Rotation: The optimal rotation about the origin is defined [51] as the orthogonal matrix which minimizes the quantity By subtracting the overall translation and rotation, the velocity fluctuation is,(6)
Dilatation: The optimal dilatation is defined [51] as the scalar that minimizes the quantity After subtracting the optimal translation, rotation and dilatation, the velocity fluctuation is finally given by,(7)where with we have indicated the contribution to the velocity of of all three collective modes.
The rotational order parameter is defined as,(8)where is the projection of on the plane orthogonal to the axis of rotation, the operator indicates the cross product, and is a unit vector in the direction of the axis of rotation. In (8), is the angular momentum of midge with respect to the axis In a perfectly coherent rotation, all individuals would have angular momenta parallel to the axis, so that In a non-coherent system, some of the projections of the angular momentum on would be positive and some negative, so Note that is the axis of rotation defined in the previous section, computed using Kabsch algorithm [51].
The dilatational order parameter is defined as,(9)
and it measures the degree of coherent expansion (positive ) and contraction (negative ) of the swarm. In a perfectly coherent expansion/contraction would be parallel to and so the scalar product in equation (9) will be for an expansion and for a contraction.
In the study of flocks [19], we normalized by its limiting value for which is equivalent dividing it by the value in the first bin. In that way the normalized correlation function tends to for so that its value is amplified. In the study of flocks we were only looking at the correlation length, which is not altered by such a normalization. However, here we will be interested in both the range and the intensity of the correlation, so we must not amplify artificially the correlation signal. Normalising the fluctuations as in (1) is equivalent normalising the correlation function by its value at exactly i.e. for which is different from its limit for
The NHS is an elementary model of non-interacting particles performing a random walk in a three-dimensional harmonic potential. The dynamics of each particle is defined by the Langevin equation,(10)where is the position of the -th particle at time is the mass, the friction coefficient, the harmonic constant and is a random vector with zero mean and unit variance, with Clearly, in this model there is no interaction between particles. The parameter tunes the strength of the noise. The equation of motion is integrated with the Euler method [52]. We simulated the NHS in the critically damped regime (), which gives the best similarity to natural swarms. The number of particles is set equal to that of the natural swarm we want to compare it with. Parameters have been tuned to have a ratio between the distance travelled by a particle in one time step (frame) and the nearest neighbour distance comparable to natural swarms,
Let us define a data set as a collection of pairs of variables, with (for example, the susceptibility as a function of the rescaled nearest neighbour distance - Fig. 4). The null hypothesis is that are independent variables. Let us call the Spearman's rank correlation coefficient for a set of data and the probability distribution of in the case of pairs of independent variables. Given the empirical data, we calculate the Spearman's rank correlation coefficient and get a certain value, The P-value is defined as the probability that the statistical test we are using (Spearman) gives a result at least as extreme as the one actually observed, provided that the null hypothesis is true. Hence, the P-value is given by,(11)
Basically, the P-value is telling us how likely it is that the degree of correlation that we observe is just the result of chance. In absence of an a priori model of the noise, we estimate by a permutation test [53], [54]: using the original paired data, we randomly redefine the pairs to create a new data set where the are a permutation of the set we calculate the Spearman's rank correlation coefficient of this new randomized data set; we iterate this permutation times; we compute the fraction of permutations that give This fraction is equal to the P-value of the data set under consideration [53].
We performed numerical simulations of the Vicsek model in 3d [35]–[38], [55]. The direction of particle at time is the average direction of all particles within a sphere of radius around (including itself). The parameter is the metric radius of interaction. The resulting direction of motion is then perturbed with a random rotation (noise). Natural swarms are known to form close to a marker and to keep a stationary position with respect to it [13]. To mimic this behaviour we modified the Vicsek model by adding an external harmonic force equal for all particles. This potential also grants cohesion, without the need to introduce an inter-individual attraction force [4], [16], [27].
The update equation for velocities is therefore given by,(12)where is the spherical neighbourhood of radius centred around is the normalization operator, and performs a random rotation uniformly distributed around the argument vector with maximum amplitude of The term is the harmonic force directed towards the origin. For we recover the standard Vicsek model. The update equation for the positions is, Thanks to the central force we can use open boundary conditions. All particles have fixed velocity modulus Each simulation has a duration of time steps, with initial conditions consisting in uniformly distributed positions in a sphere and uniformly distributed directions in the solid angle. After a transient of time steps, we saved 500 configurations at intervals of 1000 time steps in order to have configurations with velocity fluctuations uncorrelated in time. The control parameter of interest is where is the nearest neighbour distance, which is tuned by The model displays a transition to an ordered phase when We studied the susceptibility for different values of To observe the power-law behaviour of predicted by the model we performed standard finite-size scaling [38]: at each fixed value of the system' size we calculated and worked out the maximum of the susceptibility and its position we finally plotted vs. parametrically in to obtain the function in Fig. 5. The noise, affects the position of the transition point [35]–[37], but this is irrelevant for us, because we do not use any quantitative result from the model to infer the biological parameters of real swarms. The data reported in Fig. 5 have
For each frame we run a clustering algorithm with scale [56]: two points are connected when their distance is lower than For each value of we compute the ratio between the number of objects in the largest cluster and the total number of objects in the swarm (Fig. 6). The percolation threshold, is defined as the point where a giant cluster, i.e. a cluster with size of the same order as the entire system, forms [42]. We define as the point where The percolation threshold scales with the nearest neighbour distance, (Fig. 6). Strictly speaking, the percolation argument only holds at equilibrium, because in a system where particles are self-propelled there may be order even at low density [36]. However, at low values of the noise, we still expect the percolation argument to give a reasonable, albeit crude, estimate of the perception range.
|
10.1371/journal.ppat.1007918 | Lymph node migratory dendritic cells modulate HIV-1 transcription through PD-1 engagement | T-follicular helper (Tfh) cells, co-expressing PD-1 and TIGIT, serve as a major cell reservoir for HIV-1 and are responsible for active and persistent HIV-1 transcription after prolonged antiretroviral therapy (ART). However, the precise mechanisms regulating HIV-1 transcription in lymph nodes (LNs) remain unclear. In the present study, we investigated the potential role of immune checkpoint (IC)/IC-Ligand (IC-L) interactions on HIV-1 transcription in LN-microenvironment. We show that PD-L1 (PD-1-ligand) and CD155 (TIGIT-ligand) are predominantly co-expressed on LN migratory (CD1chighCCR7+CD127+) dendritic cells (DCs), that locate predominantly in extra-follicular areas in ART treated individuals. We demonstrate that TCR-mediated HIV production is suppressed in vitro in the presence of recombinant PD-L1 or CD155 and, more importantly, when LN migratory DCs are co-cultured with PD-1+/Tfh cells. These results indicate that LN migratory DCs expressing IC-Ls may more efficiently restrict HIV-1 transcription in the extra-follicular areas and explain the persistence of HIV transcription in PD-1+/Tfh cells after prolonged ART within germinal centers.
| Increasing number of evidences indicate that B-cell follicles might be anatomical sanctuaries for active transcription in both HIV/SIV viremic controllers and in ART treated aviremic HIV-infected individuals. While multiple mechanisms may be involved in the regulation of HIV transcription, recent studies suggested that immune checkpoint molecule (IC) signaling may contribute to maintain HIV-1 latency in infected CD4 T cells. These observations prompted us to investigate the involvement of IC/IC-L interactions in the regulation of HIV-1 transcription in lymph node (LN) tissues. In the present study, we show that T follicular helper (Tfh) cells predominantly co-expressed PD-1 and TIGIT, which were functionally active. An in-depth mass cytometry analysis revealed that PD-L1, PD-L2 (PD-1 ligands) and CD155 (TIGIT-ligand) were predominantly co-expressed on a specific LN dendritic cell (DC) subpopulation expressing markers of migratory DCs. We subsequently demonstrated that LN migratory DCs, locating predominantly in LN extra-follicular areas, could modulate HIV-1 transcription by a mechanism involving PD-L1/PD-1 interactions. Interestingly, the frequency of LN migratory DCs inversely correlated with HIV-1 transcription from LN memory CD4 T cells, suggesting that IC-L expressing migratory DCs might contribute to control HIV-1 transcription and maintain HIV-1 latency in extra-follicular areas. These findings represent a step forward in our understanding of potential mechanisms contributing to the regulation of HIV persistence in lymphoid tissues.
| One of the major obstacles to HIV-1 eradication resides in the capacity of HIV-1 to rapidly establish a latent reservoir, transcriptionally silent, which is not susceptible to both the host immune response and cART [1–6]. Different cell lineages including CD4 T cells and monocytes/macrophages [7–9] may contribute to the HIV-1 reservoir. Central memory and transitional memory CD4 T cells serve as major cellular compartments of the latent HIV-1 reservoir in blood [6]. More recently, blood memory CD4 T cells with stem-cell like properties [10] or expressing CXCR3 and/or CCR6 were also shown to contain latently HIV-infected cells [11–13]. However, blood contains only 2% of the total lymphocytes that reside predominantly within lymphoid organs[14], and lymphocyte populations within the lymphoid tissues may be phenotypically and functionally distinct from those in blood [15] with differential cell composition and extensive cell heterogeneity regarding T follicular helper (Tfh) cells [16]. Notably, previous studies have demonstrated that lymphoid organs are the major anatomic site for HIV infection, production and spreading and that high concentration of virus particles and CD4 T cells with active virus replication were predominantly restricted to germinal centers (GCs) in viremic individuals [17–19].
Interestingly, recent studies have underscored that while HIV or SIV DNA containing CD4 T cells were consistently detected in cells located in lymph node (LN) follicular and extra-follicular areas in long-term cART treated individuals [11, 20, 21], HIV or SIV RNA detection were mainly restricted to CD4 T cells in LN GC areas of HIV viremic controllers [22], SIV-infected elite controller macaques [23] or cART-treated aviremic HIV-infected individuals [24]. Of note, Tfh cells represent the major cellular compartment for HIV production and replication in viremic individuals [25, 26] and the major CD4 T cell population for persistent HIV-1 transcription in long-term treated individuals [24] as compared to any other blood or LN memory CD4 T cell populations. The limited access of cytotoxic CD8 T cells to GCs[23], the sub-optimal antiretroviral drug penetration in lymphoid tissues [27] and the higher level of activation of Tfh cells [26] are possible explanations for lymphoid organs serving as primary anatomic sites for HIV infection and persistence.
Multiple cellular mechanisms are involved in the establishment and the maintenance of HIV-1 latency including 1) epigenetic silencing induced by histone deacetylation and DNA methylation [28, 29], 2) limiting cellular levels of the essential Tat cofactor P-TEFb and the transcription initiation factors NF-kB and NFAT [30] and 3) condensed chromatin at the viral long terminal repeat [31]. However, under certain circumstances, HIV-1 transcription/production might be reactivated. The parameters associated with HIV-1 reactivation include T-cell receptor (TCR)-mediated signaling via NF-κB [29, 32], cytokine and chemokine stimulations [33] or epigenetic DNA modifications such as acetylation and methylation [34]. Interestingly, Fromentin et al. showed that the expression of immune checkpoint molecules (ICs) such as PD-1, LAG-3 and TIGIT on memory CD4 T cells was associated with HIV-infected cells in distinct blood memory CD4 T cell subsets during ART [35], suggesting that IC signaling may contribute to maintain HIV-1 latency in HIV-1 infected memory CD4 T cells [36, 37].
Given that previous studies have underscored the involvement of IC/IC-L interactions in the functional impairment of Tfh cells in viremic HIV-infected individuals [38–40], we hypothesized that IC/IC-ligand (IC-L) interactions may contribute to modulate HIV latency/virus reactivation in the LN microenvironment. In the present study, we therefore investigated 1) the expression and distribution of ICs and IC-Ls in blood and LN mononuclear cells isolated from cART treated aviremic, viremic and HIV-uninfected subjects using mass cytometry and in situ in lymph node compartments (germinal centers and extra-follicular zones) by immunohistochemistry, and 2) the impact of IC/IC-L interactions on TCR-mediated T-cell proliferation and HIV-1 transcription/production.
We demonstrate that PD-1 and TIGIT, the two major ICs expressed on Tfh cells ex vivo, are functionally active and regulate TCR-mediated HIV-1 transcription and production in vitro. However, PD-L1 (PD-1-ligand) and CD155 (TIGIT-ligand) were predominantly co-expressed on LN migratory (CD1chighCCR7+CD127+) dendritic cells (DCs) which located mainly in the extra-follicular areas of cART treated subjects. These findings suggest that the strength of the inhibitory signal resulting from the IC/IC-L interactions might be selectively reduced in GCs of ART treated subjects, thus creating a microenvironment less constrained for cell activation and HIV transcription. In support of this hypothesis, we demonstrate that LN migratory DCs via IC/IC-L interactions modulate TCR-mediated HIV-1 reactivation and production from LN PD-1+/Tfh cells of cART treated HIV-infected individuals and that the levels of HIV-1 transcription in LN memory CD4 T cells correlated with the reduced frequency of LN migratory DCs.
These findings indicate that the IC-L-mediated modulation of HIV-1 transcription in treated subjects is more efficient in extra-follicular areas and underscore that an imbalance in IC/IC-L interactions is a novel mechanism contributing to HIV-1 persistence in LNs.
We simultaneously collected blood and LN from 10 viremic and 10 cART treated aviremic HIV-1 infected individuals and 7 HIV-uninfected subjects. Mononuclear cells isolated from blood and LNs were then stained with a mass cytometry panel encompassing 38 markers including cell lineage markers, ICs and IC-ligands, i.e. PD-1, CTLA-4, LAG-3, TIM-3, TIGIT, PD-L1, PD-L2 and CD155. The extracellular expression of CTLA-4, LAG-3, TIM-3 and TIGIT was assessed in blood and LN memory CD4 T-cell populations identified on the basis of the expression of PD-1 and/or CXCR5, i.e. CXCR5-PD-1-, CXCR5+PD-1-, CXCR5-PD-1+, CXCR5intPD-1int and CXCR5highPD-1high CD4 T cells (Fig 1A and 1B). Of note, CXCR5highPD-1high CD4 T cells correspond to LN Tfh cells (Fig 1B). Consistent with a previous study [24, 26], LN Tfh cells of viremic untreated HIV-1 infected individuals were increased as compared to HIV-uninfected subjects and their percentage dropped after prolonged cART to levels observed in HIV-uninfected subjects (P<0.05) (Fig 1C).
Co-inhibitory molecule expression was determined extracellularly and the gating strategy defining the positivity of co-inhibitory molecules expression was set on naïve (CD45RA+CD45RO-) CD4 T cells for all co-inhibitory molecules tested (Fig 1D and 1E). The representative examples and cumulative data showed that PD-1 and TIGIT were the two major ICs expressed in CD4 T-cell populations of HIV-uninfected, viremic and aviremic cART treated HIV-1 infected individuals (Fig 1D–1G and S1 Fig). Interestingly, more than 90% of Tfh cells expressed TIGIT, while TIM-3, LAG-3 and/or CTLA-4 were expressed in less than 5% of Tfh cells (Fig 1D–1G and S1 Fig). Notably, TIGIT was also expressed in CXCR5-PD-1-, CXCR5+PD-1-, CXCR5-PD-1+ and CXCR5intPD-1int CD4 T-cell populations isolated from blood and LN of HIV-uninfected, viremic and aviremic ART treated HIV-1 infected individuals, while TIM-3, LAG-3 and/or CTLA-4 were expressed in less than 10% of blood and LN memory CD4 T-cell populations (Fig 1D–1G and S1 Fig).
We subsequently assessed the expression of PD-1 ligands, i.e. PD-L1, PD-L2 and TIGIT ligand, i.e. CD155, in blood and LN mononuclear cells including blood monocytes (CD14+), blood and LN DCs and various blood and LN B-cell populations, including LN GC B cells (CD19+IgD-CD38highCD10+) (S2 Fig). The gating strategy defining the positivity of IC-L expression was set on naïve (CD45RA+CD45RO-) CD4 T cells (S3 Fig).
The representative examples and cumulative data indicated that in blood, PD-L1, PD-L2 and CD155 were predominantly expressed on monocytes and to a lower extent on type 2 conventional DCs (cDC2; HLA-DR+CD11c+CD1c+) and plasmacytoïd DCs (pDC; HLA-DR+CD11c-CD123+), while IC-Ls were poorly expressed on blood B-cell populations of HIV-uninfected, viremic and aviremic ART treated HIV-1 infected individuals (P<0.05) (S3 Fig and Fig 2A). Of note, no correlation was observed between the frequencies of IC-L expressing blood monocytes with HIV viral load in viremic HIV-infected individuals (S4 Fig). However, although not statistically significant, but a trend towards a negative correlation was observed between the frequencies of PD-L-expressing monocytes and the duration of ART in treated individuals (r = -0.58; P = 0.08) (S4 Fig), suggesting that the expression of PD-Ls on monocytes may require longer treatment period to normalize.
In LN, IC-Ls were significantly more expressed on CD1chigh DCs than on pDCs and LN B-cell populations, including GC B cells of HIV-uninfected, viremic and aviremic ART treated HIV-1 infected individuals (P<0.05) (S3 Fig and Fig 2B).
We then further explored the phenotype of IC-L expressing LN CD1chigh DCs. The representative examples and cumulative data indicate that IC-L expressing LN CD1chigh DCs co-expressed CCR7 and CD127 (Fig 2C and S5 Fig), markers of LN migratory DCs [41, 42]. Lymph node CD1chighCCR7+CD127+ DCs were therefore referred to as LN migratory DCs.
We next investigated the influence of HIV-1 infection and treatment initiation on the frequency of LN migratory DCs of untreated viremic and treated aviremic HIV-1 infected individuals. Of note, non-reactive LNs obtained from HIV-uninfected subjects were used as control. The cumulative data indicated that the frequencies of LN migratory DCs of viremic untreated HIV-1 infected individuals were increased as compared to HIV-uninfected subjects, and their percentage dropped after prolonged cART to levels observed in HIV-uninfected subjects (P<0.05) (Fig 2D). In addition, the frequency of both LN migratory DCs and PD-L1/L2 expressing LN migratory DCs of viremic untreated HIV-1 infected individuals directly correlated with HIV-1 viral load (r = 0.875 and P = 0.0017; r = 0.8997 and P = 0.0009; r = 0.8693 and P = 0.0019, respectively) (Fig 2E and S6 Fig), suggesting that quantitative changes in IC-L-expressing LN migratory DCs might be associated with different levels of HIV load. Of note, no statistically significant correlation was observed between LN migratory DC and Tfh frequency or between PD1 and PD-L1 mean signal intensity (MFI) in untreated viremic HIV-infected individuals (P>0.05) (S6 Fig).
We then determined whether active and persistent virus transcription detected in LN PD-1+/Tfh cells [24] may result from reduced IC/IC-L interactions in the GC areas of cART treated HIV-1 infected individuals. For this purpose, the percentage of cells expressing PD-1 and the proportion of PD-L1 positive tissue surface were determined in GC and extra-follicular areas of LN sections collected from untreated viremic and treated aviremic HIV-1 infected individuals using immunohistochemistry staining as previously described[24]. Of note, to determine whether the spatial distribution of PD-1 and PD-L1-expressing cells observed in viremic HIV-infected individuals was associated with non-specific immune activation/inflammation or was specifically associated with HIV infection/replication, LN sections collected from HIV-uninfected individuals suffering from lymphadenopathy (“reactive LNs”) were used as control.
The representative examples and the cumulative data indicate that PD-1 and PD-L1 expressing cells were detected in both GC and extra-follicular areas of HIV-uninfected subjects, viremic and cART treated HIV-1 infected individuals (Fig 3A–3F). Comparison of serial immuno-labeled sections showed that cells expressing PD-1 were predominantly lymphocytes, while PD-L1 expressing cells were predominantly mononucleated histiocytes or dendritic cells. Cells expressing high levels of PD-1 (PD-1high cells) were essentially detected in the GC areas. Consistent with previous study [24], the size of GCs/mm2 and the number of PD-1high cells/mm2 of treated aviremic HIV-1 infected individuals were significantly reduced as compared to viremic untreated HIV-1 infected individuals (P<0.05) (Fig 3G and 3H).
Notably, due to the specific morphology of LN PD-L1-positive cells harboring cytoplasmic prolongations, we estimated the proportion of PD-L1 positive tissue surface as previously described [43]. Interestingly, the proportion of PD-L1 positive tissue surface was significantly lower in GCs than in extra-follicular areas of treated aviremic HIV-1 infected individuals (7.2% versus 2.2%; P<0.05) (Fig 3I). In addition, the proportion of PD-L1 positive tissue surface was significantly reduced in both GCs and extra-follicular areas of ART treated HIV-1 infected individuals as compared to viremic HIV-1 infected individuals (GC, 2.2% versus 9.4%; P<0.05; extra-follicular, 7.2% versus 26.4%; P<0.05) (Fig 3I).
Taken together, these results suggest that due to the different distribution in PD-L1 expression between extra-follicular and GC areas, the strength of the inhibitory signal resulting from PD-1/PD-L1 interactions may be weaker within GCs. The initiation of cART further influences the expression of both PD-1 and PD-L1 being associated with significant reduction in both extra-follicular and GC areas.
We then investigated whether PD-1 was functionally active on LN PD-1+/Tfh cells. To address this issue, PD-1high and PD-1-negative LN memory (CD45RA-) CD4 T cells were sorted from 5 treated aviremic HIV-infected individuals (Fig 4A). Consistent with a previous study, Tfh cells represented about 69% of the sorted PD-1high memory CD4 T cells (Fig 4B) [24]. Sorted cell-populations were labelled with carboxyfluorescein (CFSE) and stimulated with coated anti-CD3/anti-CD28 monoclonal antibodies (MAbs) in the presence or in the absence of PD-L1 recombinant protein. The impact of PD-L1/PD-1 interaction on T-cell proliferation was assessed by flow cytometry based assay, while the reactivation of HIV-1 production was assessed using HIV-1 RNA detection in day 6 culture supernatants in the presence of emtricitabine to prevent de novo HIV-1 infection and virus replication. The cumulative data indicated that PD-L1 recombinant protein did not influence TCR-mediated T-cell proliferation and/or reactivation of HIV-1 production in LN PD-1-negative memory CD4 T cells (P>0.05) (Fig 4C, 4E and 4F). However, PD-L1 recombinant protein significantly reduced both TCR-mediated T-cell proliferation and reactivation of HIV-1 production in LN PD-1+/Tfh cells (P<0.05) (Fig 4D–4F). Similar to PD-L1 recombinant protein, CD155 recombinant protein significantly reduced TCR-mediated reactivation of HIV-1 production from LN memory (CD45RA-) CD4 T cells (P<0.05) (Fig 4G). However, the combination of PD-L1 and CD155 proteins did not further inhibit TCR-mediated reactivation of HIV-1 production as compared to either individual protein (P>0.05) (Fig 4G), suggesting that the negative signal arising upon PD-1 engagement was already sufficient to suppress TCR-mediated HIV production.
These results indicate that PD-1 and TIGIT are functionally active on LN PD-1+/Tfh cells and that the interaction with their ligands reduced TCR-mediated reactivation of HIV-1 production in LN PD-1+/Tfh cells of cART treated HIV-infected individuals in vitro.
We next explored the potential role of LN migratory DCs expressing PD-L1 and/or PD-L2 to modulate TCR-mediated reactivation of HIV-1 production from LN CD4 T cells expressing or not PD-1. To address this issue, PD-1-positive and PD-1-negative LN memory (CD45RA-) CD4 T cells were sorted from 4 treated aviremic cART treated individuals. Sorted cell-populations were stimulated with coated anti-CD3/anti-CD28 MAbs in the presence or in the absence of autologous LN migratory DCs cultured with or without blocking anti-PD-L1/2 MAbs. Reactivation of HIV-1 production was assessed using HIV-1 RNA detection in day 6 culture supernatants in the presence of emtricitabine. Of note, sorted LN migratory DCs expressed more than 80% of PD-L1 and/or PD-L2, respectively (Fig 5A).
As expected, cumulative data indicated that LN migratory DCs did not influence TCR-mediated reactivation of HIV-1 production in LN PD-1-negative memory CD4 T cells (P>0.05) (Fig 5B). However, LN migratory DCs expressing >80% PD-L1/PD-L2 significantly reduced TCR-mediated reactivation of HIV-1 production in LN PD-1+/Tfh cells (P<0.05) (Fig 5B). Notably, treatment of the cultures with anti-PD-L1/2 blocking MAbs partially restored the levels of HIV-1 RNA in the culture supernatants of LN PD-1+/Tfh cells (Fig 5B).
Interestingly, the frequencies of LN migratory DCs of cART treated HIV-infected individuals inversely correlated with the levels of cell-associated HIV-1 gag RNA found in sorted LN memory CD4 T cells (r = -0.828; P = 0.05) (Fig 5C).
Taken together, these data demonstrate that LN migratory DCs could modulate HIV-1 transcription in LN of treated aviremic HIV-infected individuals through a mechanism involving PD-L/PD-1 interactions.
Additional evidence in support of the role of PD-1/PD-L1/2 interactions in regulating TCR-mediated HIV-1 reactivation was obtained evaluating the efficiency of anti-PD-1 MAb (pembrolizumab) to reactivate HIV-1 from latently infected resting memory CD4 T cells isolated from blood using a “modified” quantitative viral outgrowth assay (QVOA)[44]. For this purpose, resting memory CD4 T cells isolated from aviremic long-term treated HIV-1 infected individuals were cultured under different experimental conditions including 1) unstimulated (negative control), 2) stimulated with anti-CD3/anti-CD28 MAbs for 3 days in the presence of IL-2 (positive control), 3) exposed to isotype control, and 4) exposed to clinically relevant concentration of pembrolizumab for 14 days in the presence of autologous irradiated CD8-depleted PBMCs that include IC-L expressing cells. Of note, the use of autologous CD8-depleted PBMCs prevented the increase of transcriptional noise induced by mixed leukocyte reaction generated by the use of heterologous PBMCs and, the irradiation of CD8-depleted PBMCs prevented the reactivation of HIV-1 transcription from the feeder cells[44]. The QVOA was performed in the absence of ART to allow amplification of the reactivated virus and in a multiple replicate/limiting dilution format to allow the estimation of frequencies.
The reactivation of HIV-1 replication induced in the various conditions was then assessed by HIV-1 RNA detection in day 14 culture supernatants using Roche Taqman assay as previously described and by the estimation of the frequencies of infected cells [44]. The levels of HIV-1 RNA induced under the various conditions were generated using the 5 replicates of the highest cell concentration (5.105 cells/condition) of the “modified QVOA”. The cumulative data generated from 10 treated aviremic HIV-1 infected individuals indicated that pembrolizumab induced significantly higher levels of HIV-1 RNA in the culture supernatants as compared to untreated cultures or cells exposed to isotype control (P<0.05) (Fig 6A). The frequencies of cells containing replication competent virus were then evaluated in all conditions by the detection of HIV-1 RNA in QVOA supernatants and expressed as RNA-unit per million (RUPM) [45]. The results indicate that the average RUPM frequency following pembrolizumab exposure was significantly higher as compared to the unstimulated cells or cells exposed to the isotype control and represented about 14 cells containing replication competent virus per million (P<0.05) (Fig 6B).
Taken together, these data demonstrate that anti-PD-1 MAbs can efficiently reverse HIV-1 latency in vitro.
Increasing number of evidences indicate that B-cell follicles might be anatomical sanctuaries for active transcription in both HIV/SIV viremic controllers [22, 23, 26] and in ART treated aviremic HIV-infected individuals [24]. While multiple mechanisms may be involved in the regulation of HIV transcription, recent studies suggested that IC molecule expression may contribute to control HIV-1 transcription and therefore the maintenance of HIV latency in HIV-infected memory CD4 T cells [35–37]. Interestingly, we recently showed that Tfh cells expressing high levels of PD-1 serve as a major site of active and persistent virus transcription in ART treated aviremic individuals [24]. These observations prompted us to investigate the involvement of IC/IC-L interactions in the regulation of HIV-1 transcription in lymph node tissues.
We showed that Tfh cells, predominantly co-expressed PD-1 and TIGIT directly ex vivo in HIV-uninfected, viremic and aviremic ART treated HIV-infected individuals. We subsequently showed that PD-1 and TIGIT expressed on LN CD4 T-cell populations, including PD-1+/Tfh cells, were functionally active. Of note, in the present manuscript, we cannot exclude that LN PD-1+/Tfh cells may also encompass T follicular regulatory (TFR) cells.
Analysis of immunostained LN tissue sections indicated that PD-L1 expressing cells were detected in both extra-follicular and follicular regions of untreated viremic HIV-1 infected individuals and ART initiation was associated with a significant reduction of the proportion of PD-L1 positive tissue surface, even more pronounced in GC areas. An in-depth mass cytometry analysis revealed that PD-L1, PD-L2 and CD155 were predominantly co-expressed on a specific LN CD1chigh DC subpopulation expressing markers of migratory DCs i.e. CCR7 and CD127 [41, 42]. Consistent with previous studies, IC-Ls were also detected at low levels on blood monocytes and on blood and LN B-cell populations including GC B cells [38, 46, 47]. Interestingly, the frequency of migratory DCs directly correlated with HIV-1 viral load and was significantly reduced in aviremic ART treated HIV-infected individuals, thus indicating that HIV-1 replication influences the frequency of migratory DCs.
Migratory DCs differ from tissue to tissue but share the capacity to transport antigens to the draining LNs during both homeostatic conditions and infections [42]. In lymph nodes, migratory DCs preferentially locate at the T cell-B cell border where they contribute to induce Tfh-dependent antibody responses [48, 49].
Multiple cellular parameters such as TCR-mediated signaling, cytokine and chemokine stimulations or epigenetic DNA modifications can trigger reactivation of HIV-transcription [50], which might be in turn modulated by various mechanisms. In the present study, we investigated whether IC-Ls in general and migratory DCs expressing IC-Ls, in particular, could modulate HIV transcription in LN CD4 T cells including Tfh. Consistent with previous study, LN PD-1+/Tfh cells produced significantly higher levels of HIV-1 RNA than PD-1-negative CD4 T cells [24]. In particular, TCR-mediated HIV production of PD-1+/Tfh cells was strongly reduced in vitro in presence of recombinant PD-L1 or CD155.
More importantly, we demonstrated that LN migratory DCs could modulate TCR-induced HIV-1 transcription in LN by a mechanism involving PD-L/PD-1 interactions. Indeed, migratory DCs specifically regulated TCR-mediated HIV production of PD-1-positive CD4 T cells and anti-PD-L1/2 blocking MAb treatment partially restored the levels of HIV-1 RNA detected in PD-1-positive CD4 T-cell supernatants only. These data indicate that LN migratory DCs expressing IC-Ls may more efficiently restrict HIV-1 transcription in the extra-follicular areas versus GCs. In addition, the frequency of migratory DCs inversely correlated with HIV-1 transcription from LN memory CD4 T cells, suggesting that IC-L expressing migratory DCs might contribute to control HIV-1 transcription and maintain HIV-1 latency in extra-follicular areas. Notably, recent studies indicated that LN migratory DCs may harbor tolerogenic properties by a mechanism involving the production of immunosuppressive cytokines such as TGF-β and IL-10 and the induction of regulatory T cells [51–53].
Finally, we postulated that inhibition of PD-1/PD-Ls interactions might reverse HIV-1 latency and evaluated the efficiency of anti-PD-1 MAbs (pembrolizumab) to reactivate HIV-1 from latency using a “modified” QVOA. We demonstrated using primary CD4 T cells isolated from aviremic ART treated HIV-infected individuals and co-cultured with autologous irradiated CD8-depleted PBMCs, that clinically relevant concentration of pembrolizumab could efficiently reverse HIV-1 latency in vitro. While the precise mechanism regulating HIV latency in blood remains to be established, one may postulate that IC-L expressing monocytes may contribute to effects observed upon PD-1 blockade in vitro. These data support the recent observations suggesting the involvement of PD-1 in the establishment of HIV latency [36]. In addition, while one study found no effect [54], two other studies underscored the potential of PD-1 blockade in HIV cure strategies in combination or not with additional latency reversing agents [36, 55]. Therefore, immune checkpoint blocking antibodies (ICBs) represent a novel form of latency reversing agent that are currently being explored in the context of HIV functional cure [56]. Indeed, ICBs may potentially on one hand reverse HIV latency, thereby allowing for the expression of HIV proteins on the cell surface, and on the other hand, rescue the function of exhausted HIV-specific CD8 T cells to facilitate the elimination of reactivated cells [57]. In this regard, a phase II dose-escalation study of anti-PD-L1 antibody therapy in HIV-infected individuals was recently terminated due to safety concerns [58, 59]. Interestingly however, the data collected indicated an increase in Gag-specific CD4 and CD8 T-cell responses in a fraction of individuals with no changes in levels of cell associated HIV RNA or DNA in blood [58]. Unfortunately, no data were collected regarding the impact of ICB treatment on the tissue reservoirs. Taken together, these data indicate that future clinical trials based on ICBs should probably carefully estimate the risk-benefit ratio for HIV-infected individuals on stable suppressive ART and consider a thorough evaluation of the impact of ICB treatment on both blood and tissue reservoirs.
Recent studies suggested that IC signaling may contribute to maintain HIV-1 latency in HIV-1 infected memory CD4 T cells [36, 37, 55], however, these studies were performed exclusively on cells isolated from blood, while the present study provides four lines of evidence on the role of IC/IC-L interactions in regulating HIV transcription in LN tissues that include: 1) PD-1/PD-L1 interactions strongly impact the reactivation of TCR-mediated HIV-1 production from LN memory CD4 T cells 2) the modulation of HIV-1 transcription by LN migratory DCs through a mechanism involving PD-L/PD-1 interactions 3) the relationship between the levels of HIV-1 transcription and the frequency of PD-L1/2 expressing migratory DCs and 4) PD-1 blockade with anti-PD-1 monoclonal antibody treatment can reactivate of HIV-1 replication from latently infected CD4 T cells.
These findings represent a step forward in our understanding of the mechanism regulating the persistence of HIV transcription in lymphoid tissues.
Fifty-seven HIV-1 infected adult volunteers and twelve HIV-uninfected subjects were enrolled in the present study. No statistical method was used to predetermine sample size. The present study was approved by the Institutional Review Board of the Centre Hospitalier Universitaire Vaudois, and all subjects which were adults gave written informed consent. The 57 HIV-1-infected individuals studied had a documented diagnosis of HIV-1 infection between 0.4 and 28.4 years. Treated HIV-1 infected individuals received ART treatment for 0.3 to 23.1 years. No exclusion criteria was implemented except with regards to the reactivation of HIV-1 latency from blood resting memory CD4 T-cell experiments, which were exclusively performed on cells isolated from treated aviremic HIV-1 infected individuals with undetectable viremia (HIV-1 RNA levels <50 copies per ml of plasma) for at least 12 months. Inguinal lymph node biopsies and blood samples from HIV-infected individuals were collected the same day. In addition, LN sections collected from HIV-uninfected individuals suffering from lymphadenopathy were also collected and were referred to as “reactive” LNs.
Blood mononuclear cells were isolated as previously described [24] and lymph node mononuclear cells were isolated by mechanical disruption as previously described [60]. Blood mononuclear cells and lymph node mononuclear cells were cryopreserved in liquid nitrogen for long-term storage.
Cells were cultured in RPMI (Gibco; Life Technologies) containing 10% heat-inactivated FBS (Institut de Biotechnologies Jacques Boy), 100 IU/ml penicillin and 100 μg/ml streptomycin (Bio Concept).
The following antibodies were used for sorting experiments: APC-H7-conjugated anti-CD3 (clone SK7), APC or FITC-conjugated anti-CD4 (clone RPA-T4), ECD-conjugated anti-CD45RA (clone 2H4), V450-conjugated anti-HLA-DR (clone G46-6), PE-Cy7-conjugated anti-CD25 (clone M-A251), PerCP-Cy5.5-conjugated anti-CD69 (clone L78), PE-Cy7-conjugated anti-PD-1 (clone EH12.1) and PE-conjugated anti-CXCR5 (clone MU5UBEE), APC-conjugated CD1c (L161) and PB-conjugated CD127 (HIL-7R-M21). All antibodies including purified coating anti-CD3 (clone UCHT1) and anti-CD28 (clone CD28.2) mAbs were purchased from BD (Becton Dickinson; CA, USA); and ECD-conjugated anti-CD45RA (clone 2H4) from Beckman Coulter (CA, USA) and APC-conjugated CD1c (L161) from Biolegend. Blocking anti-PD-L1 (MIH1) and anti-PD-L2 (MIH18) MAbs were purchased from eBioscience. The following antibodies were used for mass cytometry experiments: 113In-conjugated anti-CD8 (RPA-T8), 115In-conjugated anti-CD4 (RPA-T4), 139La-conjugated anti-CD3 (UCHT1), 141Pr-conjugated anti-CD45 (HI30), 142Nd-conjugated anti-CD19 (HIB19), 143Nd- conjugated anti-ICOS (C398.4A), 145Nd-conjugated anti-CD57 (HCD57), 146Nd-conjugated anti-IgD (IA6-2), 147Sm-conjugated anti-CD7 (CD7-6B7), 148Sm-conjugated anti-PD-L1 (29E.2A3), 149Sm-conjugated anti-CD127 (A019D5), 150Nd-conjugated anti-Lag-3 (11C3C65), 151Eu-conjugated anti-CD123 (6H6), 152Sm-conjugated anti-CD21 (BL13), 153Eu-conjugated anti-Tim-3 (F38-2E2), 154Sm-conjugated anti-TIGIT (MBSA43), 155Gd-conjugated anti-CD27 (L128), 156Gd-conjugated anti-CD10 (HI10α), 158Gd-conjugated anti-CD169 (7–239), 159Tb-conjugated anti-CCR7 (G043H7), 160Gd-conjugated anti-CD14 (M5E2), 161Dy-conjugated anti-CD1c (L161), 162Dy-conjugated anti-CD11c (Bu15), 163Dy-conjugated anti-CXCR3 (G025H7), 164Dy-conjugated anti-CXCR5 (51505), 165Ho-conjugated anti-CD45RO (UCLH1), 166Er-conjugated anti-CD155 (SKII.4), 167Er-conjugated anti-CD38 (HIT2), 168Yb-conjugated anti-CD66b (CD66α-B1.1), 169Yb-conjugated anti-CD45RA (HI100), 170Er-conjugated anti-CTLA-4 (14D3), 171Yb-conjugated anti-CD20 (clone), 172Yb-conjugated anti-PD-L2 (24F.10C12), 173Yb-conjugated anti-CXCR4 (12G5), 174Yb-conjugated anti-HLA-DR (L243), 175Lu-conjugated anti-PD-1(EH12.2H7) and 191Ir was used to label DNA. Antibodies against CD8, CD4, CD3, CD28, CD155 were purchased from Biolegend. Anti-CD57 was purchased from BD. All other antibodies were purchased from Fluidigm/DVS.
Cryopreserved blood mononuclear cells were thawed, enriched using EasySep Human CD4 T Cell Enrichment kit (StemCell Technologies, USA), stained with Aqua LIVE/DEAD stain kit (4 °C; 15 min) and then with anti-CD3 APC-H7, anti-CD4 FITC, anti-CD45RA ECD, anti-HLA-DR V450, anti-CD25 PE-Cy7 and anti-CD69 PerCP-Cy5.5 MAbs. Cryopreserved lymph node mononuclear cells were thawed and then stained with Aqua LIVE/DEAD stain kit (4°C; 15 min) and then with anti-CD3 APC-H7, anti-CD4 APC, anti-CD45RA ECD, anti-PD-1 PE-Cy7 and anti-CXCR5 PE (4°C; 25 min). Viable blood resting memory (CD45RA-HLA-DR-CD25-CD69-) CD4 T cells and viable LN memory (CD45RA-) CD4 T cells, PD-1high and PD-1- CD4 T-cell populations were sorted using FACSAria (Beckton & Dickinson). In all sorting experiments the grade of purity of the sorted cell populations was >98%.
Freshly isolated matched blood and lymph node mononuclear cells were resuspended (106 cells/ml) in complete RPMI medium and incubated (30 min; 4 °C) directly ex vivo (no permeabilization) with metal-conjugated antibodies directed against a panel of 37 parameters (Fluidigm/DVS Science) including lineage markers for T-cell, B-cell and antigen presenting-cell populations and ICs such as PD-1, CTLA-4, TIM-3, LAG-3, TIGIT as well as IC-ligands such as PD-L1, PD-L2 and CD155. Cells were washed and fixed (10 min; room temperature) with 2.4% PFA. Total cells were identified by DNA intercalation (1μM Cell-ID Intercalator, Fluidigm/DVS Science) in 2% PFA at 4 °C overnight. Labeled samples were acquired on a CyTOF1 instrument that was upgraded to CyTOF2 (Fluidigm) using a flow rate of 0.045 ml/min. Data were analyzed using Fluidigm Cytobank software package (Cytobank, Mountain View, CA). At least 100,000 events were acquired for each sample.
Sorted LN PD-1high and PD-1- memory (CD45RA-) CD4 T-cell populations (2 × 105 cells) were washed twice, labeled with CFSE at 37°C for 7 min (Life Technologies) and cultured for 6 days at 37°C and 5% CO2 in 96-well U-bottom plates coated with 10 μg/ml anti-CD3 (BD) and 10 μg/ml anti-CD28 (BD) in presence or not of recombinant IC ligands at 100 μg/mL in complete RPMI. The proliferation of CD4 T-cell populations was assessed by quantifying the percentage of CFSE low cells on a LSR SORP cell analyzer (BD). Of note, all experiments were performed in presence of 75 nM emtricitabine to prevent re-infection of stimulated CD4 T cells.
Lymph nodes were cut into slices and fixed in B-plus or formalin before routine processing and embedding in paraffin blocks. Serial tissue sections (4-μm) were stained according to standard routine protocols by using a Ventana benchmark platform (Roche) with antibodies against PD1 (192106 R&D Systems) and PD-L1 (sp263, Ventana). PD1 and PD-L1-imunostained slides were digitalized using a Hamamatsu Nanozoomer 1.0 scanner (model C9600-01) at 40× with the NDPScan software (v. 2.5.89). Scanning area and focus points were set manually. Image analysis was performed with the Tissue IA–specific module of the Slidepath Software Digital Image Hub (DIH) (version 4.0.7). The surfaces of the entire tissue section and of individual germinal centers (manually circumscribed) were measured. Cell density was estimated using the “measure stained cells algorithm” at 20× to quantify PD-1-positive cells, while PD-L1-positive areas were quantified using “measure stained area algorithm” at 20×.
Sorted LN PD-1high and PD-1- memory (CD45RA-) CD4 T-cell populations (2 × 105 cells) were cultured for 6 days at 37°C and 5% CO2 in 96-well U-bottom plates coated with anti-CD3 (BD) and anti-CD28 (BD) (10 μg/ml) in presence or in absence of recombinant IC ligands at 100 μg/mL in complete RPMI. In some experiments, the combination of two IC-ligands was also tested. Of note, all experiments were performed in presence of 75 nM emtricitabine to prevent re-infection of stimulated CD4 T cells. Supernatants were collected at days 6 and quantification of HIV-1 production was performed by assessing HIV-1 RNA levels by COBAS AmpliPrep/TaqMan HIV-1 Test (Roche; Switzerland) as previously described [44].
Freshly isolated LNMCs were stained with Aqua LIVE/DEAD stain kit (4°C; 15 min) and then with anti-CD1c APC, anti-CD127 PB, anti-CD3-PE, anti-CD4 APC-H7, anti-CD45RA ECD and anti-PD-1 PE-Cy7 (4°C; 25 min). Viable LN memory (CD45RA-) PD-1+ and PD-1- CD4 T-cell populations and CD1chighCD127high DCs were sorted using FACSAria (Beckton & Dickinson). In all sorting experiments the grade of purity of the sorted cell populations was >98%. Sorted CD4 T cell populations (105 cells) were stimulated with anti-CD3/anti-CD28 MAbs (10μg/ml) or co-cultured with autologous DCs (ratio DCs/T 1∶10) in the presence or absence of blocking anti-PD-L1/2 MAbs (10 μg/ml) in 96-well U-bottom plates. All cultures were carried out in the presence of emtricitabine. Supernatants were collected at day 6 and quantification of HIV-1 production was performed by assessing HIV-1 RNA levels by COBAS AmpliPrep/TaqMan HIV-1 Test (Roche; Switzerland) as previously described [44]. PD-L1, PD-L2 and CD155 expression was assessed on sorted migratory DCs using mass cytometry.
Cell-associated HIV-1 RNA (CA-RNA) was assessed as previously described [24]. Briefly, CA-RNA was extracted from sorted CD4 T-cell populations (5 x 104 cells) and subjected to DNase treatment (RNAqueous-4PCR Kit Ambion). RNA standard curves were generated after isolation and quantification of viral RNA from supernatant of ACH2 culture as previously described [61]. One step cDNA synthesis and pre-amplification were performed as previously described [62] using the following primers ULF1: 5'- ATG CCA CGT AAG CGA AAC TCT GGG TCT CTC TDG TTA GAC-3' UR1: 5’- CCA TCT CTC TCC TTC TAG C -3'. Real-time PCR was performed using Roche light Cycler 480II using the following primers: LambdaT: 5'-ATG CCA CGT AAG CGA AAC T -3'; UR2: 5'- CTG AGG GAT CTC TAG TTA CC-3' and probes: 56-FAM 5’-CAC TCA AGG/ ZEN/CAA GCT TTA TTG AGG C-3’ IABkFQ 35.
Different cell concentrations (fivefold limiting dilutions, i.e., 5× 105, 1 × 105, 2 × 104 and 4 × 103 cells of sorted viable blood resting memory (CD45RA-HLA-DR-CD25-CD69-) CD4 T cells from aviremic ART treated HIV-infected individuals were co-cultured with autologous irradiated CD8-depleted blood mononuclear cells (106 cells/ml), as previously described [44], in presence or in absence of blocking anti-PD-1 Mab, Pembrolizumab, at 5 μg/mL or isotype control (Eureka Therapeutics). As a positive control, sorted cells were stimulated for 3 days with anti-CD3 and anti-CD28 mAb-coated plates (10 μg/ml) in presence of IL-2 (50 units/ml). Supernatants were collected at days 0, 5 and 14. Medium was replaced at day 5. Quantification of replication competent HIV-1 was performed at 14 by assessing HIV-1 RNA levels by COBAS AmpliPrep/TaqMan HIV-1 Test (Roche; Switzerland) as previously described [44]. Wells with detectable HIV-1 RNA (≥200 HIV-1 RNA copies/ml) were referred to as HIV-1 RNA-positive wells. The frequency of latently HIV-1 infected CD4 T cells reactivated by IC MAbs was estimated by conventional limiting dilution methods using Extreme Limiting Dilution analysis (http://bioinf.wehi.edu.au/software/elda/) [63] and expressed in RNA-unit per million (RUPM) as previously described [44].
Statistical significance (P values) was obtained using one-way ANOVA (Kruskal-Wallis test) followed by Mann-Whitney test or Wilcoxon Matched-pairs two-tailed Signed Rank test or ratio Paired-t test or Spearman rank test was used for correlations. Finally, Statistical significance (P values) was obtained using Extreme Limiting Dilution analysis (http://bioinf.wehi.edu.au/software/elda/) for comparison of HIV-infected cell frequencies. The analyses of multiple comparisons were taken into account for the calculation of statistical significance.
|
10.1371/journal.pcbi.1004859 | Fast and Accurate Learning When Making Discrete Numerical Estimates | Many everyday estimation tasks have an inherently discrete nature, whether the task is counting objects (e.g., a number of paint buckets) or estimating discretized continuous variables (e.g., the number of paint buckets needed to paint a room). While Bayesian inference is often used for modeling estimates made along continuous scales, discrete numerical estimates have not received as much attention, despite their common everyday occurrence. Using two tasks, a numerosity task and an area estimation task, we invoke Bayesian decision theory to characterize how people learn discrete numerical distributions and make numerical estimates. Across three experiments with novel stimulus distributions we found that participants fell between two common decision functions for converting their uncertain representation into a response: drawing a sample from their posterior distribution and taking the maximum of their posterior distribution. While this was consistent with the decision function found in previous work using continuous estimation tasks, surprisingly the prior distributions learned by participants in our experiments were much more adaptive: When making continuous estimates, participants have required thousands of trials to learn bimodal priors, but in our tasks participants learned discrete bimodal and even discrete quadrimodal priors within a few hundred trials. This makes discrete numerical estimation tasks good testbeds for investigating how people learn and make estimates.
| Studies of human perception and decision making have traditionally focused on scenarios where participants have to make estimates about continuous variables. However discrete variables are also common in our environment, potentially requiring different theoretical models. We describe ways to model such scenarios within the statistical framework of Bayesian inference and explain how aspects of such models can be teased apart experimentally. Using two experimental setups, a numerosity task and an area estimation task, we show that human participants do indeed rely on combinations of specific model components. Specifically we show that human learning in discrete tasks can be surprisingly fast and that participants can use the learned information in a way that is either optimal or near-optimal.
| People are often asked questions that require discrete numerical estimates. When judging from a glance “How many people are in the room?” or “How many dots are on a screen?” the quantity to estimate is discrete and any sensible answer must be a whole number. Discrete numerical estimates are also often required when the underlying quantity is continuous, very commonly when buying items. For example, painters often quickly assess a wall, whose area is a continuous quantity, and then buy a discrete number of paint cans; likewise at parties hosts may have to assess their guests’ hunger and order a discrete number of pizzas.
To understand how discrete numerical estimates are made, a formal framework is needed. Perhaps the most prevalent formal framework for characterizing decision making is Bayesian decision theory, which has provided a normative standard against which to measure behavior in economics [1–3], biology [4, 5] and for a wide variety of tasks in psychology such as categorization, memory, multi-sensory and sensorimotor integration, and reasoning [6–11]. Bayesian decision theory prescribes how to combine prior beliefs about states of the world, the likelihood that a state of the world generated an observation, and the decision function—the function which converts an uncertain representation into the response that maximizes expected reward. The interaction between these components is fixed in Bayesian decision theory, but the prior, likelihood, and decision function can each take many forms. This freedom allows Bayesian decision theory to correspond to a wide range of decision making behavior [12, 13], and in the right experimental design each component can be identified [14–17]. Identifying the prior characterizes how people represent past experience, identifying the likelihood indicates how people represent the evidential value of a new observation, and identifying the decision function characterizes how people convert uncertain beliefs into an estimate.
Despite the prevalence of real-world decisions requiring discrete numerical estimates, not much work has been done to characterize them using Bayesian decision theory to determine the priors, likelihoods, and decision functions that people use (cf. [18]). This stands in contrast to investigations of continuous estimates, such as pointing to a location, which have received much more attention [10, 15, 19–25]. In this paper we address this gap by characterizing how people make discrete numerical estimates. We first describe Bayesian decision theory and how continuous estimates have been characterized. Next we review previous research into discrete numerical estimation, finding it to be sparse but suggestive of differences in how continuous and discrete numerical estimates are made. We then present three new experiments that investigate discrete numerical estimates using bimodal training distributions, which allowed us to identify both the prior and decision function used. Our first experiment uses a numerosity task in which both the ground truth and response are discrete. In our second experiment, we generalize our results to a rectangle area estimation task in which the ground truth is continuous, but participants must give discrete estimates in whole square centimeters. Participants in the first two experiments learn bimodal priors quite quickly, so the third experiment investigates learning of more complex quadrimodal training distribution, which allows us to better distinguish how participants build their priors from experience. Finally we discuss our results, comparing continuous and discrete numerical estimation, exploring the implications for how people learn priors, and discussing how people convert uncertain beliefs into a numerical estimate.
To characterize how people make continuous estimates, first we outline Bayesian decision theory, which prescribes how to maximize expected rewards. Bayesian decision theory is composed of three components that each need to be specified: the prior probability, the likelihood, and the decision function. There are a variety of distributions and functions that can be used for each component, but how the components are combined is fixed by the laws of probability [26].
The decision maker begins with a prior P(S), which gives the prior probability of each state of the world, S. For simplicity, we assume that the states of the world all are arranged along a single dimension and each state has a one-to-one mapping to a response. On each trial, the decision maker observes some data, X, that are noisy or ambiguous, and the likelihood, P(X|S), is the probability of the observed data given each of the possible states of the world. The prior and the likelihood are combined via Bayes’ rule to determine the posterior probability of the states of the world having observed the data,
P ( S | X ) ∝ P ( X | S ) P ( S ) (1)
where equality is achieved if the right-hand side is divided by P(X). In a sequential task, this posterior distribution is used as the prior distribution for the next trial, so the prior reflects a participant’s accumulated experience throughout the task.
The best estimate depends not only on what is believed to be true about the world; because S given X is uncertain, it also depends on what happens if an incorrect response is made. The dependence of rewards on the response is given by the loss function L(R; S), which captures the loss (negative reward) for making response R if the state of the world is S. The decision function, DL, then maps the posterior probabilities onto the response with the smallest expected loss
D L ( P ( S | X ) ) = arg min R ∫ S L ( R ; S ) P ( S | X ) . (2)
Continuous estimates have been modeled using a particular set of priors, likelihoods, and decision functions. Likelihoods are often assumed to be Gaussian because this density is a good match to the perceptual noise in many tasks, and participants have been shown to correctly adapt to the amount of noise in their perception. For example, in multi-sensory integration and sensorimotor tasks, the normative weight applied to each sensory cue depends on the variance of Gaussian-distributed perceptual noise, and participants’ weights come close to matching these normative weights [8, 10, 27].
For the prior, a standard choice for the training distribution in continuous estimation tasks is a Gaussian density because it makes the analysis analytically tractable: combining a Gaussian prior with a Gaussian likelihood results in a Gaussian posterior. However more flexible schemes for learning priors exist, and a common way to introduce flexibility is to use a non-parametric prior which grows in complexity as more data are observed. Kernel density estimation is a standard choice for building a non-parametric prior for continuous training data [28] and this kind of representation has been used in many models of human categorization [29, 30]. In kernel density estimation, the nonparametric prior is constructed from a weighted sum of component parametric densities, one for each previously observed data point. Mixture priors, which have also been used in models of categorization [7, 31, 32], provide another representation that allows more flexibility in the number of parametric components than kernel density estimation. This representation operates between the simple parametric and the kernel density cases, grouping similar data together into the same component, but allowing different components for data that are dissimilar.
In continuous estimation tasks, participants learn Gaussian and other unimodal training distributions quickly: in various continuous estimation tasks, unimodal training distributions have been learned in hundreds of trials [10, 19, 21, 23]. Participants can also learn bimodal training distributions, demonstrating that they do not use just simple parametric priors, but they are slower to do so. Experimenters have had to train participants on bimodal distributions for thousands of trials [10, 22], because fewer training trials do not result in clear evidence of learning [19]. Participants were able to use bimodal distributions when presented with an explicit summary, and this work also showed participants are better described as using a mixture prior rather than kernel density estimate with a very narrow kernel [15].
For the decision function in continuous estimation, a small number of simple functions have been considered, each of which can be motivated by one or more loss functions. An all-or-none loss function leads to choosing the response with the highest posterior probability, a quadratic loss function leads to taking the mean of the posterior, and a linear loss function yields the posterior median as the best response. A fourth decision function, drawing a sample from the posterior, requires a more complex motivation. One route is to speculate that participants assume that the computer is adaptively responding to their input in a competitive fashion, so that a stochastic decision function can increase their expected reward [33, 34]. Another is to assume that participants are maximizing expected reward subject to particular computational costs: if participants draw samples from the posterior and these samples require time or effort to generate, it can be better to make a quick and less accurate decision rather than a slow and effortful accumulation of enough samples to calculate the maximum of the posterior [35, 36].
Comparison of these decision functions in continuous estimation has yielded task-dependent conclusions. Some researchers have found evidence for the mean decision function, finding that participants used a loss function that was quadratic near the correct value but more linear far from the correct value, giving it robustness to outliers [24]. Other work has found evidence for a mean decision function despite feedback which did not encourage this decision function [23], but a later analysis showed that this particular task does not discriminate well between decision functions [14].
Recent research has incentivized the max decision function and then investigated the decision function actually used. Work using Gaussian priors found that instead of the max, participants were performing an interpolation between drawing a single sample and taking the max of the posterior [20]. There are various mechanisms that could produce this interpolation: participants could be drawing a number of samples from the posterior distribution and taking the mean of these samples as their estimate, they could be raising the posterior to a power greater than one and then sampling their estimate from this exponentiated posterior distribution, or perhaps they combine the two by taking the mean of a number of samples from an exponentiated posterior. While these explanations are indistinguishable for Gaussian posteriors [20], work using a bimodal training distribution has successfully tested two of these possibilities, the mean of a number of samples versus a sample from an exponentiated posterior, and found that participants were drawing a single sample from an exponentiated posterior distribution [15].
Though researchers have occasionally investigated discrete numerical estimation, it has not received much attention, possibly because it has been viewed as no different from continuous estimation. However if we look at the work that has been done, there are suggestions that people make these two types of estimates differently. Bayesian decision theory is useful here for cataloguing the similarities and differences.
The likelihood in discrete numerical estimation is similar to that found in continuous estimation. As in continuous tasks, participants find it more difficult to discriminate stimuli that are closer physically even though they are naturally discrete. Indeed, in perceptual numerosity tasks, discrimination performance nearly follows Weber’s law, implying that the standard deviation of the perceptual noise distribution is proportional to its mean [37], which has been modeled as Gaussian noise on the log-transformed numbers [38–41]. Like in continuous tasks, participants appear to use knowledge of their own perceptual noise to set their likelihood in numerosity tasks [38], a point we also address in the S1 Methods.
Investigations into the priors used in discrete numerical estimation have shown that participants can learn unimodal distributions of stimuli quickly, as in continuous estimation. Participants are able to reconstruct the frequency of events from unimodal distributions from just a few trials [42] and their estimations of new events quickly show an influence of the mean in a changing sequence of numbers [43]. In the similar task of absolute identification, in which participants are asked to identify a series of perceptual stimuli with numerical labels [44], participants are also influenced by unimodal distributions of stimuli [45].
However, tasks training bimodal prior distributions point to potential differences between continuous and discrete numerical estimation. The first potential difference is in the speed of learning bimodal priors. In one task, participants asked to reconstruct bimodal prior distributions were able to do so within a few hundred training trials [18], and in another participants could do so for some bimodal distributions after only 12 trials [46]. Though this suggests that participants have a speed advantage in learning priors for discrete numerical estimates, these priors were assessed through reconstruction and it needs to be established whether the same priors are used in estimation.
A potential difference in the decision function was also found in [18], in an experiment in which participants were asked to estimate the revenues associated with trained and novel company names. Participants’ estimates for novel companies were either at the lower edge of the range of trained revenues or were in the middle of the range, results which were modeled as drawing a set of samples from the prior combined with a mixture of two decision functions. One decision function was to use the lowest sample in the set as the estimate (because unknown companies are likely to have low revenue), and the other was to take the mean of the set of samples as the estimate [18]. The use of the mean of a small number of samples as the decision function contrasts with the exponentiated posterior supported by work in continuous estimation, and this is another potential difference. However, revenue estimation is very different from the perceptual tasks used in continuous estimation, so it would help to investigate the decision function in a perceptual discrete numerical estimation task.
Here we investigate these potential differences in two different discrete numerical estimation tasks: estimating the number of dots on a screen and estimating the area of a rectangle. We are particularly interested in whether participants can quickly learn complex multimodal prior distributions and what decision function they use to make their estimates. In exploring this, we go further than previous studies by investigating whether participants use a kernel density estimate, a mixture model, or a categorical distribution as a prior. Through both combinatorial model comparison and fitting of nested models we examine the decision function, investigating whether the mean, max, the mean of a number of samples from the posterior, a sample from an exponentiated version of the posterior distribution, or perhaps a more complex decision function best explains participants’ estimates. We compare our findings to the results from continuous estimation in the discussion as well as explore the implications for what priors people can learn and what decision functions they use.
To characterize how participants make discrete numerical responses, we ran a new experiment on numerosity estimation in which participants were trained on a bimodal distribution. Participants were asked to estimate the number of dots that briefly appeared on a screen in a series of trials, receiving feedback about whether they were correct and what the correct answer was after each trial as shown in Fig 1A. They were not told anything about which numbers to expect in addition to the feedback. On each trial, the number of dots on the screen was drawn from a sharp bimodal distribution with two peaks on either side of a region of lower probability values (e.g., the distribution shown in the top left corner of Fig 2). A sharp bimodal distribution allows us to better identify the prior used. If participants are not generalizing beyond the numbers that were given as feedback, then their prior should eventually match the training distribution and they will not respond outside the range of the stimuli. However, if participants are using a parametric or kernel density prior distribution, then the prior distribution will have some spillover outside the range of stimuli, and participants will respond outside this range even after hundreds of training trials. Using a mixture prior will also result in responses outside the range, but they will likely be fewer in number. Examples of these four possibilities are shown in the top row of Fig 2.
Before and after the main task, we included a separate discrimination task that allowed us to characterize the likelihood distribution for each participant. This removed a degree of freedom from the process of characterizing the prior and decision function. The noise in numerosity judgments is well-known to follow Weber’s law with a standard deviation proportional to the mean [37], and has been modeled in past research as a lognormal distribution [38–41]. We assumed that the likelihood distribution was accurately calibrated and thus equivalent to the noise distribution. The scale of the lognormal distribution σ (i.e., the standard deviation of the natural logarithm of a lognormally distributed variable), which can be determined from the Weber fraction w using the formula σ = log ( w 2 + 1 ), was estimated in a discrimination task in which participants were asked which of two screens contained more dots (shown in Fig 1C). Participants’ discrimination judgments were well fit with a standard deviation that ranged from 0.18 to 0.53, with a median of 0.22. These estimates are in reasonable agreement with previous research that found for numerosity estimates that discriminability was equivalent to σ ≈ 0.16 [37, 47].
After fixing the lognormal standard deviation, we can make predictions for the responses after the training distribution has been learned for each pairing of prior and decision function. Using the median estimate of σ = 0.22, predictions from pairings of possible priors and decision functions are shown in Fig 2 in the form of conditional response distributions (CRDs): for trials on which a particular number of dots are presented, each panel shows the distribution over the responses expected by the combination of prior and decision function. Two qualitative features stand out in these plots. The first is the number of modes in each CRD. If the mean decision function is used or if the prior is a Gaussian distribution then the CRD will be unimodal, otherwise it will be bimodal. The second qualitative feature is whether responses occur outside the range of values that was presented. Responses will always be within the range of presented values for the categorical prior, while for other priors responses can occur outside of the range if participants sample from the posterior distribution.
In the main task, participants were assigned to one of three groups, with each group participating in a series of trials in which the sharp bimodal prior distribution covered a larger or smaller range. This was done to ensure that the results were not strongly dependent on the distance between peaks or on the particular numbers assigned to the modes of the distributions. The average and individual CRDs for each of the three groups are shown in Fig 3, along with the distribution of the training trials given to participants. In order to plot stable performance, the first 300 trials were not included in these plots. The empirical CRDs for each group show a strong bimodality, implying that neither the mean decision function nor the Gaussian prior characterize human data. Responses are not made only at the modes of the training distribution, however, as a large number of responses are found between the two peaks. These middle responses are evidence against the max decision rule (which would be very unlikely to result in an intermediate response), so from qualitative inspection posterior sampling is left as the best characterization of the decision function.
Further inspection of the average data shows very few responses outside of the range of presented values: the narrow group shows no such responses, and the medium and wide groups show few responses of these types. For the medium and wide groups, the responses outside the range of presented values only appear for a small subset of the participants: the third and fifth participant in the medium group, and the third and fourth participant in the wide group. The lack of responses outside of the range of presented values, combined with the identification of participants’ decision function as consistent with posterior sampling, implicates the use of a categorical prior, though this is not easy to distinguish from a mixture prior, as shown in Fig 2.
We fit a set of computational models (see Methods; Model comparison) to provide quantitative evidence that individual participants were using categorical priors and sampling from the posterior. Each model was fit to all of the trials and the prior was updated after each instance of feedback was given. We specifically tested the combinations of prior updating (categorical (Dirichlet) or Gaussian kernel) and decision functions (mean (Average), Max or Sample). We performed a model comparison using the Bayes Information Criterion that adjusts the fit of the model with a penalty for complexity [48]. Eighteen of the twenty participants in this experiment were best described by the categorical prior and a decision function that drew a single sample from the posterior. For the remaining two participants, one was best described by a Gaussian kernel and a max decision function and the other by a categorical prior and a max decision function. The best models for each participant are indicated in Fig 3 and the BIC values transformed into approximate posterior probabilities are shown in the S1 Methods.
To allow for a wider range of possible behaviors, we also fit computational models that allowed for “trembling hand” noise and models that allowed the posterior distribution to be raised to a power before the decision function was applied (see Methods). Once we included this set of models, we found that nineteen of the twenty participants were best described by raising the posterior distribution to an exponent larger than one before the decision function was applied, while the remaining participant was best described with the original model (implying an exponent of one). Once the posterior distribution was raised to a power, behavior was best described as a single sample for ten participants and as the mean of the exponentiated posterior for nine participants. This generalization elaborates on what was found with the first set of computational models: exponentiating the posterior means that participants lie between sampling and the max decision function, and the individual differences in using a single sample or the mean reflect individual differences in the amount of stochasticity in the estimates and in the tendency to sometimes respond near the middle of the presented range of stimuli.
A final generalization was to fit a ‘super-model’ to each participant’s data (see Methods) that allows us to further investigate the individual differences in stochasticity that participants have in their estimates by quantifying the number of samples they use. The individual best fits and an exercise showing that these parameters are identifiable are given in the S1 Methods. Reinforcing the model comparison above, nineteen of the twenty participants used an exponentiated posterior distribution to make their decisions: the exponent was well above 1.0 for all but one participant. This one participant was best fit by a single sample, so there was no evidence that any participants were taking the mean of a small number of samples from the untransformed posterior, instead participants were sampling from an interpolation between the posterior and a distribution that was entirely on the max of the posterior. This analysis allows us to further investigate those participants found to be using the mean of an exponentiated posterior. Half of these participants were best fit by taking the mean of between 2–4 samples, while the remainder were taking the mean of a larger number (i.e., about 30) samples.
In summary, this experiment established that the great majority of participants could learn essentially categorical priors when using discrete numerical responses and tended to respond using a decision function that was either a single sample or the average of multiple samples drawn from an exponentiated posterior distribution. A question then arises about the generality of the results. Are people using a categorical prior distribution because the number of dots is necessarily a discrete quantity? Or are they using a categorical prior distribution because of the discreteness of the responses?
To test whether the results of Experiment 1 were driven by the discreteness of dots or the discreteness of the response, we ran essentially the same experiment, but instead of asking participants to estimate the numbers of dots we asked them to estimate the area of rectangles. Like in the example of buying paint to cover a wall, rectangle area is a continuous quantity but we forced participants to make discrete responses: they were required to estimate the area of rectangles in whole square centimeters.
Two groups of participants were run in this experiment and the results are shown in Fig 4. When we fit their discrimination data, we found a median of σ = 0.41. As in Experiment 1, the average results in the estimation task show a bimodal distribution. Responses often fall in the middle but hardly ever fall outside the range of presented values. This pattern again is qualitatively most consistent with a categorical prior and a decision function that draws a single sample from the posterior.
We used the same analysis approach of successive generalization with the same models as we used in Experiment 1, with all individual results given in the S1 Methods. In the simplest model comparison, we found that every participant was best described by a categorical prior distribution and a decision function that was a single sample from the posterior. This result is given next to each individual in Fig 4. When we generalized the comparison to allow the posterior distribution to be raised to a power before the decision function was applied, we found that every participant was better described by exponentiating their posterior distribution, bringing it closer to a distribution that was entirely on the max. As in Experiment 1, half of participants were best described by taking a sample from this exponentiated posterior, while the other half were best described by taking the mean of the exponentiated posterior reflecting a tendency to sometimes respond near the middle of the presented range of stimuli. The more general ‘super-model’ analysis provided more detail on how many samples were being taken from the exponentiated posterior, and thus the amount of stochasticity in the estimates. All of the participants were best fit by taking the mean of between 3 and 100 samples from an exponentiated posterior distribution, with the participants best described as taking the mean of the exponentiated posterior tending to be on the higher end of this range.
In both Experiments 1 and 2 participants used near-categorical prior distributions and either take a single or multiple samples from an exponentiated posterior when they are asked to make discrete responses, regardless of whether the underlying quantity was discrete or continuous. They clearly were able to learn a bimodal distribution surprisingly quickly, so these results lead to the question of how flexible this representation is. As shown in Fig 2 a mixture of Gaussians prior can quite closely imitate the categorical prior that was best supported by the data. As the mixture model interpolates between a Gaussian prior and kernel density estimation, it is difficult to provide evidence against this model. More generally, allowing mixtures of other types of distributions, such as uniform distributions, makes the problem even more difficult. In order to test whether participants were using mixture models, Experiment 3 investigates whether participants can learn a more complex prior distribution.
In order to further investigate how complex a prior distribution participants can learn within a few hundred trials, participants in this experiment were trained on a quadrimodal prior distribution. As shown in Fig 5, this distribution was designed to test whether participants were using simple mixture models. If participants are assigning all of the trials with 23–25 dots to one mixture component and all of the trials with 29–31 dots to a separate mixture component, then the predictions of the categorical prior and the mixture prior are clearly distinguishable: the mixture model always predicts a peak in response frequency at numbers 24 and 30, while the categorical prior distribution predicts that these numbers will be selected less often than the peaks. These same predictions would also be made if participants are using other distributions in a mixture model, such as uniform distributions, with the same assignments also leads to the prediction that numbers 24 and 30 will be selected at least as often as the other presented numbers.
Three groups of participants were run in this experiment: one group completed a perceptually easier numerosity task, one group a perceptually more difficult numerosity task, and one group completed a rectangle task. When we fit the discrimination data, we found medians of σ = 0.19, σ = 0.14, and σ = 0.14 for the Difficult Numerosity, Easier Numerosity, and Rectangle groups respectively (first estimation task, see Methods) and used the latter value for generating Fig 5. The mean results for each of the groups are shown in Fig 6 and the qualitative results in this experiment again look like a combination of a categorical prior with sampling from the posterior distribution.
We ran the same analysis as was done for Experiments 1 and 2 (model comparison and fitting parameters of the ‘super-model’ with individual results given in the S1 Methods) to determine which model explained participants responses best. Using the simplest model comparison, 15 of 21 participants were best described by a categorical prior and by sampling their estimates from the posterior. The remaining six participants were better described by using a Gaussian kernel prior, with four of them taking the max of the posterior and two sampling. The correspondence of these results to the individual data is shown in Fig 6. The less restrictive model comparison again showed that a categorical prior and a single sample from an exponentiated posterior distribution was the best description of the largest number of participants (12 of 21). The other participants were best fit by a variety of models. For the ‘super-model’ analysis, which allows us to better investigate the stochasticity in the decision function, 11 out of the 21 participants drew a single sample from a posterior distribution that had an exponent clearly above 1.0, while the remainder either drew a single sample from a posterior distribution with an exponent not much different from 1.0 or took the mean of a larger number of samples from an exponentiated posterior distribution. No participants were best fit by averaging multiple samples from the unexponentiated posterior distribution.
To test whether participants were using a simple mixture model which assigned the trials with dots 23–25 to one mixture component and trials with dots 29–31 to separate mixture component, we looked at whether participants produced fewer responses of 24 and 30 compared to the peaks of the distribution. If participants were equally likely to respond with any of the presented numbers (after trial 300 and ignoring what the actual presented value was), then participants should have picked numbers 24 or 30 on at least 1/3 of trials. Using 1/3 of trials as a null hypothesis we ran binomial tests to determine if the actual number of responses was significantly lower than this value for each participant. Overall, 14 of 21 participants produced significantly fewer responses than the null hypothesis predicted (p < 0.05). The participant showing significant differences are marked with stars in Fig 6. Clearly a number of participants were not using this simple mixture model as their prior distribution.
Mixture models that are closer to the categorical prior are harder to rule out. For example, mixture model components might consist of separate components for every number, except for a single pair of numbers that are represented with the same component. For the prior distribution trained in this experiment, this would be a mixture model consisting of five component densities to represent the six presented responses. We simulated how often responses just outside the presented range would appear if there were separate mixture components for every number except for one single pair of adjacent numbers for a variety of choices of the adjacent pairs and values of σ and found that participants would be expected to produce a response just outside the range on perhaps as few as 0.6% of trials. This low rate means that it is not possible to say that any individual participant produced significantly fewer responses: the probability of producing zero of these responses on 200 trials assuming a 0.6% probability is 0.3. However, we do note that 11 of 21 participants did not produce any of these just-outside-the-range responses (and four additional participants produced only one). If all participants were consistently grouping adjacent numbers together the probability of observing this many participants with zero responses of this type is low (p = .027).
Overall, Experiment 3 demonstrates that participants are accurately learning very complex quadrimodal prior distribution within a few hundred trials. The complexity of the prior learned allowed us to even rule out for most participants a simple mixture model that could have explained behavior in the first two experiments.
In the preceding pages we have characterized the components necessary for optimal Bayesian decision making with discrete numerical stimuli and explained how model comparison and model fitting allows us to tease apart these components given the right experimental setup.
The results from three experiments show that most participants were better described by a perfect match to the training distribution, rather than a parametric distribution, a kernel density estimate, or some forms of mixture model. Generally speaking participants were best characterized as raising their posterior distribution to a power, which interpolates between the original posterior distribution and a distribution entirely on the max on the posterior. Using this transformed distribution, participants’ estimates were best described as averaging one or more of samples from an exponentiated posterior distribution, with the number of samples reflecting individual differences in the stochasticity of their estimates and in the probability of responding near the middle of the presented range of stimuli. As participants tended to either draw a single sample or use a large exponent to transform their posterior distribution, this ruled out the mean or the mean of a number of samples as viable decision functions. Experiments 1 and 2 established this pattern across both the numerosity and area estimation tasks, pointing toward the discreteness of the response rather than the discreteness of the stimuli as the driver of this behavior. Experiment 3 expanded this to even more complicated stimulus distributions, providing evidence against other mechanisms for updating the prior. We now compare our results to continuous and discrete estimation in previous experiments, and discuss the conclusions we can draw about priors and decision functions.
Previous investigations had suggested that the decision function used in discrete numerical estimation might be different from that used in continuous estimation tasks. In tasks in which the max decision function was incentivized, work with continuous estimation tasks has shown that participants exponentiated the posterior distribution and drew a sample from this exponentiated distribution [15, 20]. This result contrasts with the findings from discrete numerical tasks showing that even when incentivized to use the max rule, participants still appeared to use the mean of a small number of samples [18].
Our results for discrete numerical estimation were different. We encouraged the max decision rule (by giving participants feedback of ‘correct’ or ‘incorrect’), and we found that participants were using an exponentiated version of the posterior distribution. Our ‘super-model’ analysis allowed for both exponentiation of the posterior and for the mean to be taken of a number of samples, and we found strong evidence for exponentiation in a large majority of participants and individual differences in whether a single or multiple samples were used. Despite the individual differences, no participants were best fit by the mean of a small number of samples from the untransformed posterior.
The divergent results between our task and the revenue estimation task of [18] need explanation. One potential key difference is that the likelihood was characterized in our task but not in the revenue estimation task. Pronounceable company names induce different expectations about company stock performance than non-pronounceable names [49], and these expectations could be reflected in a variety of likelihoods that cause the resulting estimates to resemble those coming from a mixture of decision functions. This is of course speculative, and the divergence may be due to other task differences, but it highlights the importance of characterizing all of the components of Bayesian decision theory.
Overall, our findings for the decision function roughly correspond with those found in continuous estimation, so there seems to be no strong dividing line between the decision function used to make these two types of estimates. Of course decision functions have usually not been characterized in much detail nor have been characterized across a range of tasks, so later investigations could reveal subtle differences in how the task shapes the decision function used.
In terms of the prior, across three experiments we found the novel result that participants were better characterized as using a categorical prior than by a simple parametric distribution or by a kernel density estimate with any appreciable width. Mixture priors could possibly explain the results of Experiments 1 and 2, but Experiment 3 showed that a simple implementation of a mixture prior did not match the data as well for most participants.
The use of a categorical prior was supported by participants’ ability to learn complex multimodal distributions very quickly. The speed and flexibility of participants’ prior learning stands in contrast to work in continuous tasks, where it is difficult to find evidence for quick learning of bimodal priors. One task required 4,000 feedback training trials to teach participants a bimodal distribution [10], and another required 1,700 trials [22]. However, when [19] used 1,500 training trials in an interval timing task, there was some suggestion of bimodality if the peaks were well-separated, but the data could also be explained by a uniform prior.
The only example of equally fast learning of a bimodal prior without giving participants hints comes from other experiments using discrete numerical responses. In the revenue estimation experiments of [18], it was found that a bimodal prior distribution could be reconstructed within 400 training trials. More impressive was the demonstration that bimodal priors could be reconstructed after as little as 12 trials by individual participants [46]. However, these demonstrations come from tasks in which participants are asked to reconstruct the distribution rather than make an estimate, so this work is the first to show that participants do use bimodal priors when making estimates, and can learn to do so more quickly than in continuous tasks. In addition the priors that participants learned were impressively accurate: we showed that quadrimodal prior distributions could be learned, and that their priors were better described by a categorical distribution rather than a kernel density estimate or some forms of mixture models.
Given these differences in speed of learning, it is interesting to speculate whether there are particular properties of tasks that require discrete numerical responses that make it easier to learn a complex prior. One difference between discrete numerical and continuous estimates is that it is easier to provide clear feedback for discrete numerical estimates. Both [19] and [10] used visual position as feedback in their sensorimotor and interval timing tasks, and noise in vision and memory makes this feedback less certain. In the orientation estimation task of [22] participants were told their average deviation every 20 trials, while this feedback is digital it does not provide as much information as the true orientation used on each trial. It is difficult to see how the feedback could be improved for tasks that require continuous responses: feedback either needs to be susceptible to noise (perhaps both sensory noise and noise in encoding and remembering the feedback) or it is not directly mapped to the responses. Participants cannot perfectly be shown what response they should have made.
In contrast, our experiments and the experiments of [18] showed participants the correct response after every trial in essentially a noise-free fashion. This is a real advantage of using discrete numerical responses and feedback because feedback can be given uncorrupted by sensory noise after every trial and it is easily mapped to the responses than participants make. In fact, [46] explicitly showed this difference when participants were asked to reproduce a distribution: for experiments in which numbered stimuli were replaced by circles of various sizes, participants required more trials and greater separation between the modes to learn the bimodal distributions. It is possible that the differences in clarity of feedback explain the rates at which participants learn bimodal priors in different tasks. If participants are using a form of Occam’s razor when constructing their prior distribution, then the more informative trials would more quickly convince them to abandon a simpler prior in favor of a more complex representation.
The priors learned in our experiments, especially Experiment 3, were much more complex than those taught to participants in other estimation tasks [10, 15, 18, 19, 22, 46]. In addition to having four modes, our prior had a pattern of low-probability and no-probability responses that participants’ responses matched. Participants were not just representing the prior as a mixture of two parametric components, but were learning the prior probabilities associated with individual responses.
Work using other tasks has demonstrated fairly complex prior learning, but in other tasks it is generally not clear whether participants are learning a prior or a mapping. For example using a categorization task, a subset of participants learned to discriminate a multidimensional quadrimodal distribution from a multidimensional mixture of two Gaussian distributions [50]. While participants were able learn these complex discriminations and their behaviour could be described by a model that approximates Bayesian inference, this work did not rule out a complex decision bound model (i.e., a mapping) as an alternative [50, 51].
In Experiment 3 we ran additional trials to test whether participants were actually learning and using a prior or if instead they were learning a mapping from the stimuli to the responses. As discussed in the S1 Methods, we found that the responses of more than half of participants were best explained by a prior rather than a mapping. Use of a prior is also supported by recent work that demonstrated that participants take into account the reliability of various senses in a multisensory numerosity task [38].
Our results contrast with other work showing that participants do not learn a categorical prior. In a continuous estimation task with a wide range of possible responses, a categorical prior did not explain the data as well as a mixture [15]. Likewise in a numerosity task that showed participants a much wider range of numbers than our experiments, a mixture model provided a better fit to their participants’ data than just using the trained examples as a prior [52].
The key difference is likely the variety of correct responses in each experiment. As the number of potential responses increases it is hard to imagine that participants would precisely track the frequency with which every single number appeared. For example, if every number from 100 to 200 appeared in a random order with the exception of number 134, it is implausible that participants would notice.
This contrast raises questions about where the transition between a categorical prior and a mixture model occurs, and even if there is a distinction between the two. It is possible that participants represent a small set of numbers symbolically and use a categorical prior, but represent a large set of numbers as a mixture prior over a continuous variable. Alternatively, it could be that our categorical prior is simply a mixture with a separate component for each response. In this case, there would likely be a smoother transition between representations of the prior for small and large sets of responses.
Very few of our participants were best fit by a simple decision function: the max or mean of the untransformed posterior distribution. Instead it appeared that the large majority of participants were performing some kind of approximate inference by drawing one or more samples. Previous work has put forward mechanisms with which this could be done. For example, [20] showed that participants’ responses were consistent with either taking the mean of a number of samples in a continuous estimation task or drawing a single sample from an exponentiated posterior distribution. Later, [15] disambiguated these two operations with a bimodal prior, showing that raising the posterior distribution to an exponent was the better description.
Both of these mechanisms have been touted as tradeoffs between effort and accuracy, and possibly a rational use of cognitive resources [35], though there is always the possibility that participants have particularly complex hypotheses about the computer’s behavior instead. Drawing a sample may take time or effort, and a small number of samples may provide the best tradeoff between effort and accuracy to yield the highest overall reward [36]. Similarly, raising a posterior distribution to a power has also been cast as a tradeoff between effort and accuracy, but one that assumes effort is required to perform the exponentiation that transforms the belief distribution into a response distribution [53].
While this makes for a nice contrast, the picture is complicated by two additional mechanisms that are essentially indistinguishable from exponentiating the posterior distribution, even for bimodal priors: taking the maximum of a number of samples drawn from the posterior distribution, and taking the maximum of a posterior distribution that has been corrupted with noise [15]. This last mechanism may well differ from the others if it is assumed that the amount of noise in the posterior is not under the control of the participant; in this case sampling-like behavior would not be a tradeoff between effort and accuracy.
We add to this literature by showing that while an exponentiated posterior distribution is necessary to explain the data as in [15], additionally a large number of participants appear to be taking the mean of a number of samples drawn from this exponentiated posterior distribution. Despite the fact that the maximum of the posterior was asked for by identifying only exactly correct responses as ‘correct’, participants still showed some tendency to produce some responses near the mean of the posterior.
It is interesting to speculate what sort of mechanism could support both a tendency to respond with the max and a tendency to respond with the mean. Our best fitting combination of the mean of samples from an exponentiated posterior distribution is one possibility. It could be that participants are using an exponentiated posterior helps to emphasize the mode which is most likely under the posterior distribution. The later sampling operation helps to select the best response in that mode, trading off the need to pick the highest posterior with the uncertainty introduced by having several highly likely responses in close proximity. It may even be that responses near the mean of the posterior are an accidental byproduct of this two-stage process.
However, the difficult-to-distinguish alternatives to an exponentiated posterior point toward alternative combinations. One of these is a pure sampling approach: participants draw samples from the posterior distribution and sometimes take the maximum of the samples and sometimes take the mean. Another alternative combination is to ascribe all the variability to noise in the posterior distribution: using a noisy posterior distribution, sometimes participants take the maximum and sometimes they take the mean.
To gain additional purchase on this question, we correlated the average response times of participants with the model parameters. It might be expected that if participants were using any of tradeoffs between effort and accuracy that there would be correlations between each participant’s average response time and the number of samples or the exponent that the super-model recovered. This kind of correlation has been found in previous work when looking two-alternative responses [36]. However, both within each experiment and across experiments, we found no reliable relationship between the either of these model parameters and the response times of participants (see S1 Methods for details).
On the surface, this null result could be considered evidence that participants use the maximum or mean of a noisy posterior distribution to produce their estimates and that the amount of noise in the posterior does not depend on participant effort. However, it could also be that participants have such differing goals for effort / accuracy tradeoffs that this washes out whatever correlations there are between response time and model parameters. Future work would provide stronger tests of these mechanisms using within-participant designs that manipulate rewards and time-pressure, along with emphasizing that the computer is not responding to participant behavior.
There are many examples of discrete numerical tasks in everyday life, such as the examples of painters quickly assessing the size of a wall in order to buy the right number of paint cans or of the party hosts assessing the hunger of their guests when buying a discrete number of pizzas. In our experiments, we used a numerosity task and an area estimation tasks because both of these tasks are well studied and the likelihood distributions have been well characterized. This allowed us to quickly measure the standard deviation of the likelihood for each participant. If we had used less controlled stimuli, then we might have had to measure the full distribution of responses that each individual stimulus evoked in order to characterize the likelihood.
Our laboratory tasks are similar to some everyday tasks. The numerosity task we used is similar to estimating the number of visible stars in the sky (which does vary depending on the time of day and light pollution), and estimating the size of a rectangle shares some similarities to the example of the painter who needs to assess the area of wall. However, there are differences as well: the stars in the sky do not differ in size from night to night and painters can view a wall from many distances and angles before producing an estimate. With the right stimuli, it would be interesting to investigate real-life performance in discrete numerical estimation tasks.
Our results demonstrate that people represent a surprising amount of complexity in their prior distribution with relatively few training trials and use this complex prior when making new estimates. Training complex priors has multiple benefits: we can more easily observe how people represent priors and we can investigate some of the more complex schemes describing how people convert the posterior distribution into a single estimate.
This work raises many questions about how prior and posterior distributions are represented and how estimates are made. Discrete numerical estimation tasks, which are simple to implement and quick to train, are well suited for future work in this area.
Twenty-one University of Warwick students participated in this experiment for course credit. Participants gave written informed consent and the experiment was approved by the University of Warwick Humanities and Social Sciences Research Ethics Committee. Participants were divided into three groups and each participated in one version of the experiment, as outlined below. One participant was excluded because of computer error and second was only given one block of discrimination trials but was included in the analysis.
The stimuli consisted of displays of a number of identical dots. Each display of dots consisted of white dots on a black background, visible for 500 ms. Dot radius and dot density were randomized for each display to encourage participants to make numerosity judgments instead of judging the amount of light produced by the display, the density of the display, or the area occupied by the dots. In a single display all dots had a single common radius of between 3 and 9 pixels that was chosen randomly with equal probability on each trial. Dots were positioned randomly within a circular available region which was centered on the display, subject to the constraint that no dot could lay within one-dot-diameter of another dot. The available region randomly varied between 150 and 450 pixels in radius. A uniform draw was made over the possible values of dot density (where density equaled the maximum number of dots that could appear in that block divided by the area of the available region), which determined the radius of the available region on a trial.
The experiment consisted of a single session with three blocks. The estimation trials, in which participants saw a single display of dots and responded with their estimate of the number of dots in each display, were presented in the second block. The estimation trials differed for the three groups of participants: the narrow group, the medium group, and the wide group. The narrow group consisted of eight participants who saw 800 estimation trials in which the number of dots varied between 23 and 29. For this group, displays with 23 and 29 dots each appeared with probability 0.3 and the displays with the remaining numbers appeared with probability 0.08. The medium group consisted of six participants who saw 700 estimation trials in which the number of dots varied between 23 and 32. For this group, displays with 23 and 32 dots each appeared with probability 0.3 and the displays with the remaining numbers appeared with probability 0.05. The wide group consisted of six participants who saw 700 estimation trials in which the number of dots varied between 23 and 35. For this group, displays with 23 and 35 dots each appeared with probability 0.28 and the displays with the remaining numbers appeared with probability 0.04. Every participant saw 10 practice estimation trials displaying between one and four dots before beginning the main phase of the experiment.
The first and third block consisted of 128 discrimination trials each (always proceeded by 4 practice trials), in which participants saw two sequential displays of dots and picked the display that contained the larger number of dots. On every discrimination trial, one of the displays had a specific high or low number of dots. These anchor numbers were set to be either 11 dots below or above the lowest number seen in the estimation trials. The other display consisted of a number of dots that was equal to the anchor plus an offset. The offset was randomly chosen with equal probability from the set of {-8, -4, -2, -1, 1, 2, 4, 8}. Because of computer error one participant, in the group that had a range of 23 to 29 in the estimation trials, was given anchor trials of 18 and 54, in his or her first block. This participant received the correct anchor trials (12 and 40) in the third block.
On estimation trials, after the dot display disappeared, participants were asked to enter the number of dots that they saw. After entering their response, participants received feedback about whether they were correct and the actual number of dots that were shown. On discrimination trials, participants only received feedback about whether they were correct or not.
Twelve University of Warwick students participated in this experiment for £6 apiece. Participants gave written informed consent and the experiment was approved by the University of Warwick Humanities and Social Sciences Research Ethics Committee. Participants were divided into two groups and each participated in one version of the experiment, as outlined below.
The stimuli consisted of displays of rectangles of particular areas. For each display, the width of the rectangle and its position were randomized to encourage participants to judge the area of the rectangles without exclusively relying on its length, width, or position on the screen. Rectangle width was chosen from a continuous uniform distribution between 2cm and 10cm, with length chosen to achieve the desired area given the width. A fixed positional jitter was chosen for each trial uniformly from a 6cm square. On estimation trials the rectangle was on average in the center of the screen and appeared for 500ms. On discrimination trials, the first rectangle appeared 10cm left of center plus positional jitter for 500ms, and after a 500ms delay the second rectangle appeared 10cm right of center plus positional jitter for 500ms.
The procedure was identical to Experiment 1 with the following exceptions. Participants responded in the estimation trials with the area of the rectangle in square cm. A narrow group and a medium group was run in this experiment consisting of six participants each, with equivalent numbers and probabilities to the groups with the same names in Experiment 1. Every participant saw 700 estimation trials in this experiment.
Twenty-four University of Warwick students participated in this experiment for £6 apiece. Participants gave written informed consent and the experiment was approved by the University of Warwick Humanities and Social Sciences Research Ethics Committee. Participants were divided into three groups and each participated in one version of the experiment, as outlined below. Two participants were excluded from the rectangle group and one from the easier numerosity group because of computer error. Two additional participants from the easier numerosity group saw between ten and fifteen additional non-feedback trials at the beginning of the second estimation block and their data (excluding these trials) were included in the analyses.
For all groups the experiment consisted of four blocks: the first discrimination block of 128 trials, the first estimation block of 500 trials, the second estimation block of 200 trials, and finally the second discrimination block of 128 trials. Feedback was given for all blocks except for the second estimation block which served as a test of whether a prior had been learned. The first discrimination and estimation blocks used easier-to-see displays than the second discrimination and estimation blocks. The three groups of participants differed in the details of the displays they were shown. The difficult numerosity group were run with the same display parameters as in Experiment 1 during the first discrimination and estimation blocks. During the second estimation and discrimination blocks, the first numerosity group was shown each dots display for 50ms and at a much reduced luminance. The easier numerosity group was given different display parameters during the first discrimination and estimation blocks: the range of the common radius of dots was from 6 to 9 pixels, while the available region randomly varied between 225 and 375 pixels. The easier numerosity group had a shorter display (50ms) as well as more variability on the second discrimination and estimation blocks: the range of the common radius of dots was from 1 to 11 pixels, while the available region randomly varied between 150 and 450 pixels. The third group, the rectangle group, was given rectangles that randomly varied in width between 4.8 and 6.5cm during the first discrimination and estimation blocks. In the second discrimination and estimation blocks, this group was given a shorter (50ms) and slightly dimmer (gray instead of white) display with rectangles that randomly varied in width between 2 and 10cm.
All participants in this experiment were given the same trial structure during the estimation trials. The distribution that generated the trials was quadrimodal with a 20% chance of drawing each of the numbers 23, 25, 29, or 31. In addition, there was a 10% chance of drawing each of the numbers 24 and 30.
To estimate the variability in participants’ internal estimates, X, we analyzed the discrimination data in order to utilize fitted parameters for the estimation task. Specifically we assumed that the internal estimate was distributed according to a log-normal distribution (in accordance with Weber’s law) log(X)∼N(log(X); log(S), σ2) where X and S are positive integers. Participants were presented with two stimuli, S1 and S2 (as in the 2AFC discrimination trials) and had to estimate which one was larger.
In order to fit the variable σ we maximized the log-likelihood across trials (or rather minimized the negative log-likelihood) using Matlab’s fminbnd function σ ^ = arg max Σ i log P ( R i | S i , 1 : 2 , σ ). The model likelihood P(Ri|Si,1: 2, σ) was estimated numerically for each trial and condition by sampling X1 and X2 10,000 times and for each set generating a fictitious response. P(Ri|Si,1: 2, σ) = (1/10000)Σl H(Xl,1 − Xl,2) where H is the Heaviside function and log(Xl,1)∼N(log(Xl,1); log(S1), σ2) and log(Xl,2)∼N(log(Xl,2); log(S2), σ2) are samples from the generative model above. This analysis assumed that participants chose the most likely response on each trial, which is what was found in a recent analysis of 2AFC choice tasks [54].
The purpose of our analysis for the Estimation data is to compare and rule out different models of human decision making (see Experiments 1–3 above). One common way of comparing perceptual models of different number of parameters (see e.g. [55]) is to fit the parameters of each model through maximum likelihood and compensate for differences in model complexity by calculating the Bayesian Information Criterion (which penalizes models with large number of parameters).
Our secondary analysis instead created a single model that encompasses all of the candidate models as special cases, that is for certain parameter sets the larger model is equivalent to each of the nested candidate models. The best fit of the parameters therefore shows which of the models best describes the data.
We will first describe the generative model, how to perform inference over it, how to perform model comparison, and lastly we describe the ‘super-model’ and explain what specific models it encompasses.
A traditional way of comparing models is by maximizing the likelihood of each model and correcting for number of parameters using the Bayesian Information Criterion [48]. We factorially combined two different priors (Dirichlet or Gaussian kernel), three types of decision function (mean, max or sampling) and three types of noise models (none, trembling hand or softmax) to produce 16 different models for each participant (note that softmax is not defined for a max decision function removing two of the 2x3x3 = 18 combinations).
The priors were updated using either a variable width Gaussian kernel or “zero width” Dirichlet updating. Combined with the likelihood, this creates the posterior distribution P(St|Xt) upon which the subject bases their decision.
A softmax noise model performs a transformation of the posterior
P n ( S t = i | X t ) = P ( S t = i | X t ) β Σ j P ( S t = j | X t ) β (6)
where β < 1 leads to a widening (or flattening) of the posterior, while β > 1 leads to a sharpening of the posterior. For noise models none or trembling hand we set β = 1.
Choices (as discussed above) are then made based on either max, mean (average), or sampling of the exponentiated posterior. Finally, the trembling hand noise model, included as an alternative to the softmax model, states that participants have a small probability ϵ, of performing a random choice. I.e. S t ^ ∼ U [ 1 : 100 ] if e < ϵ, where e is randomly sampled from U[0: 1]. While the trembling hand noise model is structurally implemented after the decision is made, it has a similar effects as the softmax noise model in increasing the variability of the responses.
As the internal variable Xt, was unknown to the experimenters we used ancestral sampling [28], drawing 10,000 samples from the generative process for Xt ∼ P(Xt|St) followed by the inference by the subject as described above (based on any model parameters). This generates 10,000 independent estimates of S ^ which we can use to numerically approximate the probability of response S t ^. As they are essentially discrete counts we describe them as a discrete categorical distribution which provides us with the model likelihood P m ( S t ^ | S t , p a r ) for trial t and any model m and its associated parameters par.
For each model and parameter set for each subject the log-likelihood was thus calculated as log L = Σ t log P ( S t ^ = R t | S t , p a r ). Note that the parameter σ had been fit independently for each subject through the discrimination experiment above and was thus not a free parameter. Any trials with subject responses of less than 5 dots (Rt < 5) were ignored as erroneous key presses given that the true number of dots presented were at least 23. Furthermore to avoid any singularities in calculating the likelihood we allowed for a small probability (0.001) that participants would make a random response in the range [1:100] (similar to a slight Trembling Hand).
Any parameters par = (β, ϵ, or ψ) were fit through maximum likelihood (Matlab’s fminsearch). For model comparison we calculated BIC for each subject and each model:
B I C = - 2 * Σ t log P ( S t ^ = R t | S t , p a r ) + N * log ( M ) (7)
where N is the number of model parameters in par (0, 1 or 2) and M is the total number of trials.
As an alternative to the model comparison we can compare models factorially based on how they update the prior (given by parameter ψ), how many samples are drawn (parameter n), and how exponentiated the posterior is (parameter β). The fitted parameter set for each subject encapsulates which model aspects best explain the subject’s behavior. Thus the specific models compared above become special cases within this larger parameter space, allowing us to extrapolate between the models.
We put all of these into a unified framework, which we refer as the ‘Super-model’ (as other models are sub-sets of it). Given the posterior P(St|Xt) we assume that participants choose their response by averaging over samples:
S t ^ = 1 n Σ i n s k (8)
where the samples sk are given by
s k ∼ P n ( S t = i | X t ) = P ( S t = i | X t ) β Σ j P ( S t = j | X t ) β (9)
where the parameters n and β are fitted for each subject (see below).
The summation over samples allows us to approximate properties of specific decision functions. For n = 1 a single sample is drawn, equivalent to the sampling decision function. The averaging approximates the mean of the distribution for very large n (thus approximating the mean decision function).
The sampling of sk is from a softmax function (also known as the exponentiated Luce choice rule [56, 57]) which causes all the probability density to be sharpened at the peaks of the posterior for larger values of β. For large values of β the number of samples (n) becomes of little consequence (for example for β > 2.7, with Xt = 23 and σ = 0.22 after learning for 300 trials, more than 95 percent of the probability is at the maximum a posteriori).
In this way specific parameters emulate the mean (average, n = 10000), max (β = 1000) and sampling (n = 1) decision functions.
In order to fit the variables (ψ, n, β) we performed log-likelihood maximization on ψ, β using Matlab’s fminsearch function (on −logL with 5 random initializations), for each of n = [1, 2, 3, 4, 5, 10, 30, 100, 1000]. For each subject this allowed us to find the parameter set with the maximum likelihood and, given that the models of interest are nested models of this model-parameter set, indirectly find the model that best describes the data.
|
10.1371/journal.pntd.0006201 | Rhinoscleroma pathogenesis: The type K3 capsule of Klebsiella rhinoscleromatis is a virulence factor not involved in Mikulicz cells formation | Rhinoscleroma is a human specific chronic granulomatous infection of the nose and upper airways caused by the Gram-negative bacterium Klebsiella pneumoniae subsp. rhinoscleromatis. Although considered a rare disease, it is endemic in low-income countries where hygienic conditions are poor. A hallmark of this pathology is the appearance of atypical foamy monocytes called Mikulicz cells. However, the pathogenesis of rhinoscleroma remains poorly investigated. Capsule polysaccharide (CPS) is a prominent virulence factor in bacteria. All K. rhinoscleromatis strains are of K3 serotype, suggesting that CPS can be an important driver of rhinoscleroma disease. In this study, we describe the creation of the first mutant of K. rhinoscleromatis, inactivated in its capsule export machinery. Using a murine model recapitulating the formation of Mikulicz cells in lungs, we observed that a K. rhinoscleromatis CPS mutant (KR cps-) is strongly attenuated and that mice infected with a high dose of KR cps- are still able to induce Mikulicz cells formation, unlike a K. pneumoniae capsule mutant, and to partially recapitulate the characteristic strong production of IL-10. Altogether, the results of this study show that CPS is a virulence factor of K. rhinoscleromatis not involved in the specific appearance of Mikulicz cells.
| Rhinoscleroma is a human specific chronic infection characterized by the formation of granuloma in the nose and upper airways. It is a rare disease endemic in low-income countries where hygienic conditions are poor and caused by the Gram-negative bacterium Klebsiella pneumoniae subsp. rhinoscleromatis. A hallmark of this pathology is the appearance of atypical foamy monocytes called Mikulicz cells. Very little is known about the cellular and molecular mechanisms underlying this disease and the bacterial virulence factors of K. rhinoscleromatis are unknown. In this study, we created the first mutant made in K. rhinoscleromatis and inactivated the production of capsule, an outer-membrane-anchored polysaccharide. Using a murine model recapitulating the formation of Mikulicz cells and this bacterial capsule mutant, we observed that capsule is a virulence factor for K. rhinoscleromatis which is not required for the formation of Mikulicz cells, indicating that other specific virulence factors are present in the genome of the bacterium. This works opens the way to further genetic analysis of K. rhinoscleromatis and identification of new specific virulence factors.
| Rhinoscleroma is a chronic granulomatous infectious disease that affects the nose and other parts of the respiratory tract down to the trachea [1]. Although few sporadic cases are typically described in Western Europe and in the USA, this disease is still endemic in impoverished areas of the Middle East, Eastern Europe, tropical Africa, South East Asia, Central and South America. A delay in the diagnosis can lead to complications such as physical deformity, upper airway obstruction and, rarely, sepsis. Treatment can be challenging and includes surgery and prolonged course of antibiotics to avoid relapses. The bacterium implicated as the causative agent of rhinoscleroma is Klebsiella pneumoniae subsp. rhinoscleromatis (hereafter mentioned as K. rhinoscleromatis or KR), a subspecies of Klebsiella pneumoniae. Despite being geographically broadly distributed, K. rhinoscleromatis has been isolated mainly in human [2] although three recent reports mention the identification of K. rhinoscleromatis in cockroaches [3,4] or chickens [5] in low hygiene settings. K. rhinoscleromatis is very closely related to Klebsiella pneumoniae subsp. pneumoniae but can be distinguished from K. pneumoniae sensu stricto by biochemical properties and multilocus sequence typing [6].
Rhinoscleroma development is typically described clinically and pathologically into three overlapping stages: catarrhal stage, proliferative stage, and sclerotic stage [7]. The catarrhal stage is marked by purulent rhinorroea and nasal obstruction, which persists for months. Histological examination shows evidence of squamous metaplasia with a subepithelial infiltrate of polymorphonuclear cells. However, in the subepithelial layer, bacteria are incompletely digested and further released into tissues. The proliferative stage is characterized by symptoms of epistaxis, nasal deformity and other problems depending on the other areas affected. In addition, histology shows the appearance of Mikulicz cells, a hallmark of rhinoscleroma [8]. These cells are large foamy macrophages with numerous enlarged vacuoles containing viable or non-viable bacteria. Finally, the sclerotic stage is characterized by increasing deformity, granulomatous areas and scar formation. Most patients are diagnosed in the proliferative stage, when the lesion appears as a bluish-red, rubbery granuloma and the typical Mikulicz cells can be observed.
Mikulicz cells are only documented in rhinoscleroma and have been described as atypical inflammatory monocytes specifically recruited from the bone-marrow upon K. rhinoscleromatis infection [9]. These cells represent a peculiar state of highly vacuolated inflammatory monocytes unable to digest bacteria. Moreover, it has been shown that IL-10, an anti-inflammatory cytokine, is essential in the establishment of a proper environment leading to the phenotypic maturation of Mikulicz cells [9].
Different virulence factors have been implicated in the pathogenesis of K. pneumoniae. Capsule polysaccharide (CPS) is recognized as one of the most important virulence determinants of this pathogen. The presence of CPS inhibits the deposition of complement components onto the bacterium [10–12], impedes adhesion and reduces phagocytosis of the bacterium by macrophages and epithelial cells [10,12–17]. Using in vivo models of colonization and pathogenesis, CPS mutants have been shown to be unable to colonize either pulmonary or systemic tissues [13,18,19]. Clearly, CPS plays an important role in the interplay between K. pneumoniae and the innate immune system.
K. pneumoniae and K. rhinoscleromatis are heavily capsulated bacteria. K. pneumoniae express 134 different capsular serotypes that they are easily transferred via homologous recombination [20,21]. Interestingly, despite their scattered geographical distribution, all K. rhinoscleromatis isolates belong to capsular type 3 (K3) [6]. This is raising the question of whether the K3 serotype capsule composition plays any specific role in rhinoscleroma pathology. Indeed the K3 capsule repeated unit is rich in mannose residues, its repeated unit being composed of →2-[(4,6-(S)-pyruvate)-α-D-Man-(1→4)]-α-D-GalA-(1→3)-α-D-Man-(1→2)-α-D-Man-(1→3)-ß-D-Gal-(1→ [22]. This is also suggestive of possible interaction of the bacteria with mannose receptors mainly carried by macrophages and dendritic cells. Indeed, the K3 capsule has been shown to be one of the few Klebsiella K types able to bind to the mannose receptor [23]. The complete sequence of the genomic region comprising the capsule polysaccharide synthesis gene cluster was determined [24]. However, to date, the link between CPS and K. rhinoscleromatis virulence remains to be elucidated. The role of the K. rhinoscleromatis CPS has never been tested in vivo since, currently, there are no K. rhinoscleromatis CPS mutants available.
As CPS is a prominent factor in other bacteria, here we explored the possibility that K. rhinoscleromatis CPS is implicated in the peculiar pathophysiological aspects of rhinoscleroma. We have previously established an intranasal mouse model of K. rhinoscleromatis infection recapitulating the formation of Mikulicz cells, the major histological feature of the disease [9]. In this work, we successfully constructed a K. rhinoscleromatis CPS mutant strain, representing the first report of the use of genetic tools in K. rhinoscleromatis. Further, using our mouse model, we compared the host responses to wild-type and K. rhinoscleromatis CPS mutant infections by examining cytokine production and pulmonary histology. We report that the K. rhinoscleromatis CPS mutant is attenuated in vivo but also that Mikulicz cells are observed upon infection with high dose of K. rhinoscleromatis CPS mutant. Our data indicate that capsule is a virulence factor of K. rhinoscleromatis but is not involved in the specific appearance of Mikulicz cells.
All protocols involving animal experiments were carried out in accordance with the ethical guidelines of Pasteur Institute, Paris and approved by the Comité d'Ethique de l'Institut Pasteur (CETEA) (comité d'éthique en expérimentation animale n°89) under the protocol license number: 2013–0031. All mice had free access to food and water and were under controlled light/dark cycle, temperature and humidity. Animals were handled with regard for pain alleviation of suffering. Animals were anesthetized using ketamine and xylazine, and euthanized with CO2.
Bacterial strains and plasmids used in this study are listed in Table 1. The K. pneumoniae subsp. rhinoscleromatis SB3432 strain (KR WT) was isolated in 2004 at the Avicenne hospital, Bobigny, France, from a biopsy of the left nasal cavity of an 11-years old patient diagnosed with rhinoscleroma. The K. pneumoniae subsp. pneumoniae Kp52145 strain is a previously described clinical isolate (serotype O1:K2) [25]. The Escherichia coli strains used in the cloning experiments were DH5α λpir (Invitrogen) and ß2163, kind gift from Didier Mazel (Institut Pasteur, France). pGEM-T (Promega) is TA cloning vector used for cloning PCR products. pDS132 was a kind gift from Dominique Schneider (Université Joseph Fourier, France). A kanamycin cassette was PCR amplified from the plasmid pKD4 [26] and recombineering plasmid pSIM6 expressing Red system was used to create mutant in Kp52145 [27]. The plasmid pAT881 carrying the luxABCDE operon was used to make bioluminescent strains [28].
Bacteria were grown in Lysogeny Broth (LB) medium at 37°C with shaking. When appropriate, antibiotics were added at the following concentrations: ampicillin (Amp) 100 μg/ml; chloramphenicol (Cm) 30 μg/ml; kanamycin (Kan) 50 μg/ml. When necessary, DAP was supplemented to a final concentration of 0,3 mM. For selection against sacB, LB medium was supplemented with sucrose to a final concentration of 5% (wt/vol).
Inocula were prepared from overnight bacterial cultures grown on a loan on LB plates at 37°C resuspended in physiological saline.
Capsule K. rhinoscleromatis mutant (KR cps-) was obtained by insertion of the plasmid pAM2 in the wzc gene. Briefly, a kanamycin cassette flanked by 1kb of upstream (wzb) and 1 kb of downstream (wbaP) sequences of wzc using a three-step PCR method [29] was cloned into pGEM-T and then subcloned into pDS132 suicide vector. The resulting plasmid was introduced in the E. coli ß2163 donor strain (DAP-) and the recombinant strain was used for conjugation with K. rhinoscleromatis. KR cps- mutants were selected onto Kan/DAP- plates.
K. pneumoniae 52Δwzc (Kp52Δwzc) was generated using the λ RED recombination technique [26]. Briefly, a kanamycin cassette was amplified by PCR from the pKD4 plasmid using primers Kp52WzcUpKan (5’-ATCAGTGTTCAAACTTATTGAGCAATCTGCACTGTTATGGGCTGAGAAATTAAAAGCTTAGAAATTCAGGAAATAATGCATGATTGAACAAGATGGATTG -3’) and Kp52WzcDownKan (5’- CGATATGGATGACGTTCATTATTATCCTTTTATTATATATTTTAAAAAAGGGGATTCTTCGTCCCCTTCTTGAGTAACTCAGAAGAACTCGTCAAGAAGG -3’). The PCR product was purified onto a column, digested with DpnI, repurified and electroporated into K. pneumoniae carrying pSIM6, which encodes the λ RED recombinase. Kan-resistant clones were screened for successful genomic replacement of the entire wzc. Deletion of wzc on the K. pneumoniae 52145 chromosome was confirmed by PCR and sequencing.
Female BALB/cJ mice were purchased from Janvier (Le Genest-Saint-Isle, France).
Inocula of WT and mutant bacteria used in this study are 2.107 bacteria for KR WT, 2.107, 4.108 and 109 for KR cps- and 109 for Kp52Δwzc. When appropriate, similar inocula of the respective bioluminescent strains were used.
Bacterial counts were determined as colony forming units (CFU) by plating serial dilutions of lung homogenates in 3 ml ice-cold PBS supplemented with 0,5% Triton X-100 and EDTA-free protease inhibitors (Fisher Scientific).
For survival studies, mice received either 2.107 KR WT or 2.107, 4.108, 109 KR cps- by the intranasal route. Following infection, animals were returned to standard housing and observed for 14 days. A census of survivors was taken daily.
In order to maintain the plasmid conferring luciferase expression, mice were injected intraperitoneally twice daily from 1 day post-infection with 20 mg/kg spectinomycin (Spectam). Following isoflurane anesthesia, bioluminescence imaging was performed using an IVIS Spectrum (Perkin Elmer). Analysis and quantification of bioluminescence were done using Living Image (Perkin Elmer).
At 96h post-infection lungs were inflated with 4% PFA and fixed overnight at 4°C. Paraffin-embedded tissue blocks were cut into 7 μm sections and stained with hematoxylin-eosin (HE). Images were acquired with the AxioScan.Z1 (Zeiss) using the Zeiss Zen2 software.
FISH staining was performed as follows. Paraffin lung sections were deparaffinized, rehydrated in PBS and covered with a solution of lysozyme at 10 mg/ml in PBS during 30 min at 37°C. Slides were then washed twice in PBS, preincubated 30 min at 42°C in hybridization buffer (20 mM Tris-HCl [pH 8], 0.9 M NaCl, 0.01% SDS, 30% formamide) and incubated overnight at 55°C in hybridization buffer containing 50 nM of the pan-bacteria probe Eub338-Alexa555 5′-GCTGCCTCCCGTAGGAGT-3 [30]. After washing in 1X SSC (1 SSC is 0.15 M NaCl plus 0.015 M sodium citrate), slides were covered for 1 min with DAPI to visualize the nuclei, washed in PBS and mounted in Prolong Gold reagent. Images were acquired on an upright fluorescence microscope equipped with the Apotome technology (Zeiss AxioImager with Apotome2, Carl Zeiss Jena).
The number of Mikulicz cells was estimated from HE stained sections by manually segmenting region containing high number of Mikulicz cells in Zen Blue software (Zeiss). Regions containing Mikulicz cells within dense infiltrate of inflammatory cells were not included. Mikulicz cells-containing region was quantified as % area of the total lung area.
The number of bacteria present in the tissue section was quantified from fluorescence images using the Fiji plugin TrackMate [31]. Bacteria were defined as spots of 1.5 μm after Laplace Gaussian fitting.
Capsule was quantified as the concentration of uronic acid in the samples from a standard curve of D-glucuronic acid as described by Favre-Bonte et al [14]. The uronic acid content was expressed in nanograms per 106 CFU.
At various time post-infection the five pulmonary lobes were removed and collected in ceramic-beads containing tubes (Precellys lysing kit CK28) with 2,5 ml of ice cold PBS supplemented with 0,5% Triton X-100 and EDTA-free protease inhibitors (Fisher Scientific). Samples were then crushed using the Precellys homogenizer with the following program: 3 cycles of 15 sec at 5.000 × g with 10 sec pause. Twenty microliters were removed to determine the number of CFU/lung. After adding 10 μl of Pen/Strep (100X, Sigma), samples were centrifuged at 300 × g for 10 min and left on ice for 30 min. The supernatants were frozen rapidly in dry-ice ethanol bath and stored at -80°C. The following cytokines were measured: IL1ß, IL-10, IL-17, TNFα (Duoset, all from R&D Systems). Assays were performed according to the manufacturer’s instructions.
Correlation between bioluminescence signal and CFU number was analyzed by Pearson correlation using GraphPad Prism 5.
To investigate the role of capsule in K. rhinoscleromatis virulence, we constructed a KR capsule mutant (KR cps-) from the K. rhinoscleromatis wild-type strain SB3432 (KR WT) by insertion of a suicide plasmid. It has been shown that inactivation of wzc gene, whose product is involved in capsular polysaccharide export machinery, leads to a capsule-minus phenotype in K. pneumoniae [13]. We decided thus to mutate the capsular operon in SB3432 by replacing the wzc gene by a kanamycin cassette by using the suicide plasmid pAM2. Although this suicide plasmid can be normally excised following double crossover using sacB counter-selection, we did not manage to obtain the desired gene replacement, possibly because KR does not grow on media without salt which is required for sacB counter-selection. Nevertheless, sequencing of KR cps- confirmed the integration of the suicide plasmid in the wzb gene leading to a polar effect and one base deletion in the sacB gene leading to the production of a truncated SacB protein, hence explaining the selection of this mutant during the counter-selection step. A schematic representation of the wild-type KR and the capsule mutant KR cps- capsule export portion of cps operon is shown in Fig 1A. As expected, colonies of KR cps- did not show the slimy and mucoid phenotype characteristic of surface polysaccharide-producing KR colonies (Fig 1B). We also quantified the amount of capsule produced and observed a drastic reduction from 329±59 to 10±9 ng uronic acid / 106 bacteria for KR WT and KR cps- respectively. Altogether, these results indicated that this KR mutant is an effective capsule mutant.
Capsule is a well-characterized virulence factor of K. pneumoniae. K. pneumoniae capsule mutants are avirulent and they are not able to cause pneumonia or urinary tract infections [13,19,34]. We sought to analyze whether the KR cps- strain was attenuated in vivo. Anticipating that at identical inoculum of KR WT and KR cps- this would be the case, we wondered whether we could recapitulate part of the disease by increasing the infectious dose of the KR cps-. BALB/c mice were thus infected intranasally with 2.107 KR WT or 2.107, 4.108 or 109 KR cps- and survival was monitored over 14 days (Fig 1C). While all mice infected with 2.107 bacteria of KR WT strain succumbed within 6 days post-infection, mice infected with 2.107 or 4.108 KR cps- bacteria recovered from the infection and survived. However, a 50% death rate was observed with the highest dose of 109 KR cps-. Altogether, these findings show that the KR cps-strain is attenuated in vivo, confirming the crucial role of capsule in KR virulence.
In order to compare KR WT and KR cps- infections, we tested the capacity of bioluminescent bacteria to colonize the lungs after intranasal instillation. Mice were infected with either 2.107 bioluminescent KR WT or 2.107, 4.108 or 109 bioluminescent KR cps-, and bioluminescence imaging was performed and quantified 6, 24, 48, 72, 96 hours post-infection (Fig 2A and 2B). Mice infected with bioluminescent KR WT showed a gradual increase in lungs bioluminescence with a 430 fold signal increase at 4 days post-infection as compared after 6 hours. On the other hand, the bioluminescence signal started to decrease from 6 hours post-infection with 2.107 KR cps- and reached background level at day 1. A similar but less pronounced decrease was observed in mice infected with 4.108 or 109 KR cps- indicating a higher persistence of the mutant bacteria in the lungs. Moreover, because of a more viscous inoculum at high infection doses leading to difficulties to achieve proper intranasal infection, some mice swallowed part of the inoculum and showed a bioluminescent signal in the gut that disappeared in most of the animals at day 4, indicating that the bacteria transited in the gut before being eliminated. To correlate the bioluminescent signal with the bacterial load, we quantified the number of CFU in the lungs after bioluminescence imaging 96 hours post-infection. After subtraction of the background signal, we observed a significant correlation between bioluminescence and CFU in mice infected with 2.107 KR WT and 4.108 or 109 KR cps- (S1 Fig), allowing a good estimate of CFU greater than 5.105 bacteria in the lungs from the bioluminescence signal.
We also directly monitored the lungs bacterial load during the same time course in mice infected with inocula of 2.107 KR WT or 2.107, 4.108 KR cps- (Fig 2C). While the number of bacteria in mice infected with 2.107 KR WT gradually increased from 4.107 bacteria per lungs 6 hours post-infection to reach 4.109 bacteria at 96 hours, the number of bacteria in animals infected with the same inoculum of KR cps- decreased gradually until the bacteria were being completely cleared from the organ in 72 hours. However, lungs from mice infected with a higher inoculum of 4.108 KR cps- presented a still significant amount of bacteria in the organ 4 days post-infection, providing a more relevant comparison to the wild-type infection. By 96 hours after infection with 4.108 KR cps-, 33% of mice successfully cleared the infection while the others were still being colonized and had between 5.105 and 109 bacteria in their lungs. These results indicated that the KR cps- mutant is strongly attenuated but that at a higher inoculum, after a certain threshold, KR cps- is able to persist and proliferate within the host.
To examine the pathology induced by KR cps-, lungs of mice were also examined histologically at 4 days post-infection (Fig 3A). Animals infected by 2.107 KR WT presented the classical extensive but moderately destructive inflammation of the lungs characterized by the recruitment and formation of large Mikulicz cells filling alveoli. By contrast, mice infected with 2.107 KR cps- showed localized dense inflammatory lesions with signs of hemorrhages and recruitment of monocytic and polymorphonuclear cells. No classical Mikulicz cells could be observed. This phenotype reflects the inflammatory response that was required to eradicate the bacteria. Interestingly, when mice were challenged with 4.108 or 109 KR cps-, many alveoli were filled almost exclusively with Mikulicz cells, similarly to what is observed with KR WT, although the alveolar lining was more often disrupted. Regions with dense and localized inflammatory regions, characterized by infiltration of numerous polymorphonuclear cells, were also observed (Fig 3A, highlighted zone). Of note, all mice infected with 4.108 or 109 KR cps- out of 9 examined histologically presented Mikulicz cells. Altogether, these observations suggested that while capsule is a virulence factor in KR, it is not required to induce the formation of Mikulicz cells in KR pathogenesis.
As mice infected with the KR cps- strain showed variations in the intensity of the Mikulicz cells infiltrate observed by histology, we wondered whether this variation was correlated to the bacterial burden. Because we cannot directly quantify total CFU and perform an histological analysis on the same sample, we estimated the number of bacteria by fluorescence in situ hybridization and quantified the Mikulicz cells infiltrate by manually segmenting regions containing highly visible Mikulicz cells on adjacent lungs sections (S2 Fig). Mice infected with KR cps- showed different number of bacteria spots and extend of Mikulicz cells infiltrate in the lung section (Fig 3B). Both parameters were significantly correlated, suggesting that the local bacterial load drives the intensity of recruitment of Mikulicz cells.
Cytokines are key mediators of immune responses and the anti-inflammatory cytokine IL-10 has been shown to be highly produced after K. rhinoscleromatis infection and to play a crucial role in the establishment of a proper environment leading to Mikulicz cells maturation [9]. Therefore, we characterized the production of some major cytokines in mouse lung extracts upon KR cps- infection. When BALB/c mice were infected with 2.107 KR WT or 4.108 KR cps-, the pro-inflammatory cytokines IL-1β, IL-17 and TNF-α were produced in high amounts from 6 hours post-infection onwards (Fig 4A and S3 Fig). However, although produced in similar amounts at the beginning of the infection in mice infected with 2.107 KR cps-, the level of these cytokines diminished overtime because bacteria were progressively cleared from the organ. As previously shown, the anti-inflammatory cytokine IL-10 was highly produced upon infection with 2.107 KR WT but not in mice infected with 2.107 KR cps-. IL-10 was also produced in mice infected with 4.108 KR cps-, but to a lower extend and in more variable manner as compared to KR WT (Fig 4B). These observations indicate that a high inoculum of KR cps- allows recapitulating a high production of IL-10, thereby suggesting that capsule does not have a direct role in IL-10 production upon KR infection.
Because we observed a high variability in the production of IL-10 in mice infected with 4.108 KR cps-, we wondered whether it was correlated with the burden of the infection. We thus compared the production of IL-1β, IL-17, TNF-α and IL-10 to the number of CFU in the lungs at 96 hours post-infection for each animal. While a high production of IL-1β (> 3.104 pg/ml) is indicative of the presence of bacteria in the lungs (mainly ranging from 105−109 bacteria), IL-1β is expressed at intermediate levels (500–3.000 pg/ml) when mice managed to clear the infection (Fig 4C). Similar observation was made for IL-17 and TNF-α (S4 Fig). On the other hand, this is different for IL-10 (Fig 4D). A first group of mice mildly colonized (between 5.105 and 2.107 CFU) showed intermediate level of IL-10 (between 70 and 200 pg/ml) while a second group of mice that were unable to control the infection (> 2.107 CFU) were characterized by an intense production of IL-10 (> 103 pg/ml) suggesting that KR is able to induce an intense production of IL-10 only above a certain threshold of KR bacteria in the lungs.
To establish that the occurrence of Mikulicz cells observed with 4.108 and 109 KR cps- was not due to the higher inoculum of KR cps- as compared to KR, we measured bacterial loads and cytokines expression in animals inoculated with the same high inoculum (109 bacteria) of Kp52Δwzc 4 days post-instillation. The Kp52Δwzc strain is a similar capsule mutant from K. pneumoniae strain Kp52145 obtained after deletion of the wzc gene showing a drastic reduction of capsule expression from 256±22 ng uronic acid / 106 bacteria for Kp52145 to 34±13 ng uronic acid / 106 bacteria (S5 Fig). We observed that the bacterial load of mice infected with 109 Kp52Δwzc was around 105 bacteria per organ and was lower than the bacterial load of mice infected with 109 KR cps- indicating that a wzc mutant in Kp52145 is less virulent that its counterpart in KR (Fig 5A).
We then measured the cytokines levels in lungs of infected animals (Fig 5B and S6 Fig). Pro-inflammatory cytokines IL-1β, IL-17 and TNF-α were expressed in similar amounts in mice infected with Kp52Δwzc or KR cps-. However, IL-10 was expressed at low level after 109 Kp52Δwzc infection (53–63 pg/ml), contrasting the higher amount observed after 109 KR cps- infection in some mice. By histology, we observed an intense and dense inflammation characterized by a strong recruitment of monocytes and polymorphonuclear cells and an absence of Mikulicz cells formation (Fig 5C). Altogether, and combined with the histological data, these observations suggest that when present in high concentration in the lungs from 3 days of infection without being lethal, KR or its capsule mutant are able to induce the recruitment and maturation of Mikulicz cells and drive a strong production of IL-10.
The diversity of capsule types in Klebsiella pneumoniae species is strikingly very large, as 134 different capsule loci have been identified up to now [21]. This tends to indicate that K. pneumoniae species is under strong selection pressure to diversify its capsule. However, and strikingly, all K. rhinoscleromatis strains isolated so far are of the KL3 (K3) serotype despite having been isolated from diverse geographical locations [6]. Because of this homogeneity, we speculated that this specific K3 serotype could be an important factor driving the rhinoscleroma disease. By creating a capsule mutant in K. rhinoscleromatis, we showed that if capsule is an important virulence factor for this species, it is not necessary to induce the formation of Mikulicz cells, the hallmark of rhinoscleroma, as these cells have been observed when using high inocula of this mutant.
The saccharide composition of the capsule has been linked to some extent to K. pneumoniae virulence. K1 and K2 serotypes have been suggested to be major determinants in liver abscess-causing K. pneumoniae [35,36]. Strains from other serotypes, including K5, K16, K20, K54 and K57, have also been described as highly virulent [37]. In addition, switching the capsular serotype of a highly virulent K2 strain to a weakly virulent K21a strain has been shown to lead to a decrease in virulence in mouse and in survival in blood and to an increased binding to macrophages. Conversely, switching the capsule serotype of the K21a strain to virulent K2 resulted in an increased virulence in mouse and in survival in blood and to a lower binding to macrophages [38,39]. In addition, switching highly conserved genes of the capsule cluster involved in capsule export from K1 into K20 hypervirulent strain strongly reduced its bacterial virulence in mice while increasing its neutrophil phagocytosis and survival in macrophages, although it is still not known whether this is due to a change in capsule expression[40]. However, a recent pan-genomic analysis did not reveal any correlation between capsule serotype and strains responsible of invasive community-acquired infection but rather suggested that the presence of one or several siderophores explains bacterial virulence [41]. Thus the exact role of capsule composition in virulence still remains to be clearly determined.
Capsule plays an important role in immune cells evasion by preventing binding of complement and antibodies to the bacteria thereby decreasing opsono-phagocytosis and complement-mediated killing [10–13,42–45]. Moreover, Klebsiella capsule composition has been shown to influence the binding of the bacteria to macrophages. K3, K46 and K64 K. pneumoniae capsule are binding more to the mannose receptor, which is highly expressed on macrophages, than other serotypes, in a mannose-dependent manner, while other serotypes presented no binding [23]. A common feature of these three different serotypes is that they have two or three mannose residues in their repeated unit. Though, other K serotypes that present also two mannose residues did not show any binding to the mannose receptor, suggesting that binding of mannose-bearing capsule to the mannose receptor is influenced by other factors than its mannose composition. However, as all K. rhinoscleromatis strains are of the K3 serotype, and even though K3 capsule interacts with mannose receptor, our results obtained with a high infection dose of KR cps- suggest that this step is not important in driving the development of Mikulicz cells.
Some results obtained with the high inocula of KR cps- were heterogeneous: the bacterial load 4 days post-infection was spread over 4 logs, IL-10 levels in the lungs were quite variable and some mice showed some bacteria in the digestive tract by bioluminescence. This variability is a consequence of the use of higher inocula, which are thicker and more viscous than lower inocula used for the KR WT and of KR cps- strains that are more fluid. As a consequence, part of the inoculum is swallowed by mice and passes into the digestive tract. This is also suggesting that above a certain threshold of cps- bacteria delivered to the lungs, the animal cannot control the infection and the bacteria are able to multiply and maintain themselves in high number, although to a lower burden than WT bacteria. We wondered whether there was a correlation between the number of bacteria in the lungs and the level of IL-10 produced. Indeed we observed that IL-10 was produced in high amount when the bacterial load was high, raising the possibility that high IL-10 expression was the result of a high bacterial burden and not specific to K. rhinoscleromatis. To verify this one needs to compare IL-10 production upon similar bacterial burdens, greater than 108 bacteria, at 4 days post-infection with different bacteria. We first thought to use a high dose of a K. pneumoniae mutant inactivated in the same gene as the KR cps- strain, but showed that the bacterial load was lower (104−105 bacteria) than the lowest ones obtained with high KR cps- (106 to 108 bacteria) and that IL-10 levels were also quite low. This showed that this Kp52Δwzc mutant was actually more attenuated than KR cps- and suggested that K. rhinoscleromatis is better adapted to surviving in lungs. Some virulent K. pneumoniae strains can cause intense and severe and acute pneumonia in mice with high burden. We had previously observed that a variable bacterial load can be achieved 3 and 5 days post-infection with a low dose of the virulent strain Kp52145 [9] and that about 30% of mice were presenting a high bacteria burden 3 and 5 days post-infection. By measuring CFU loads and cytokines in mice infected with Kp52145 we observed that mice that had a high bacterial load were producing IL-10 in amount similar to those that were less colonized. Comparable high bacterial burden were obtained with the widely used K. pneumoniae strain 43816 [18,19,46] and IL-10 was produced in similar low amounts 3 days post-infection [18]. Hence these observations indicate that the intense IL-10 production observed upon infection with KR WT or KR cps- is specific of K. rhinoscleromatis and does not result from a global high bacterial load.
Moreover, all high dose KR cps—infected mice out of 9 observed by histology show the presence of Mikulicz cells in their lungs, although to various extent. We also observed that the density of the Mikulicz cells infiltrate is correlated to the number of bacteria. We also tried to see whether there was a similar correlation with the amount of IL-10 on a mouse to mouse basis, but were unable to detect directly this cytokine by immunohistochemistry. Nevertheless, the variation in the host response to KR cps- infection is likely correlated to the amount of IL-10 produced: lower number of bacteria lead to fewer Mikulicz cells and low amounts of IL-10 whereas an intense IL-10 production is accompanied by high number of bacteria and Mikulicz cells and less destructive inflammation.
Recently, IL-10 has been shown to regulate metabolic processes in activated macrophages and thus control the inflammatory response. IL-10 impedes glycolysis and promotes oxidative phosphorylation maintaining mitochondrial fitness. This metabolic reprogramming of macrophages is controlled by IL-10 through inhibition of mechanistic target of rapamycin (mTOR) signaling pathway [47]. Interestingly, deregulation of mTOR signaling, such as prolonged mTORC1 activation, leads to metabolic changes, hyperproliferation of macrophages and granuloma formation, contributing to disease progression in human granulomatous sarcoidosis [48]. These mechanisms might be associated with formation of granulomas in rhinoscleroma, where Mikulicz cells could undergo similar metabolic remodeling mediated by IL-10.
The fact that the capsule is not required for Mikulicz cells recruitment and formation indicates that the factors responsible of this process are still unknown and remain to be identified. Current in vivo screening approaches, such as signature tagged mutagenesis, cannot be used as they identify mutants unable to grow in specific experimental conditions, but not those that are required for the expression of a particular phenotype, such as the appearance of Mikulicz cells. Therefore an in vitro screening assay has to be developed. However, in vivo phagocytosis assays can often be difficult to set up and standardize due to the high expression of capsule in K pneumoniae species and its strong anti-phagocytic effect. Our results show that capsule is not required for the formation of Mikulicz cells, opening the way to in vitro assays of Mikulicz cells formation and to in vitro screening of factors that are driving this maturation in vivo.
|
10.1371/journal.pgen.1002368 | PBX1 Genomic Pioneer Function Drives ERα Signaling Underlying Progression in Breast Cancer | Altered transcriptional programs are a hallmark of diseases, yet how these are established is still ill-defined. PBX1 is a TALE homeodomain protein involved in the development of different types of cancers. The estrogen receptor alpha (ERα) is central to the development of two-thirds of all breast cancers. Here we demonstrate that PBX1 acts as a pioneer factor and is essential for the ERα-mediated transcriptional response driving aggressive tumors in breast cancer. Indeed, PBX1 expression correlates with ERα in primary breast tumors, and breast cancer cells depleted of PBX1 no longer proliferate following estrogen stimulation. Profiling PBX1 recruitment and chromatin accessibility across the genome of breast cancer cells through ChIP-seq and FAIRE-seq reveals that PBX1 is loaded and promotes chromatin openness at specific genomic locations through its capacity to read specific epigenetic signatures. Accordingly, PBX1 guides ERα recruitment to a specific subset of sites. Expression profiling studies demonstrate that PBX1 controls over 70% of the estrogen response. More importantly, the PBX1-dependent transcriptional program is associated with poor-outcome in breast cancer patients. Correspondingly, PBX1 expression alone can discriminate a priori the outcome in ERα-positive breast cancer patients. These features are markedly different from the previously characterized ERα-associated pioneer factor FoxA1. Indeed, PBX1 is the only pioneer factor identified to date that discriminates outcome such as metastasis in ERα-positive breast cancer patients. Together our results reveal that PBX1 is a novel pioneer factor defining aggressive ERα-positive breast tumors, as it guides ERα genomic activity to unique genomic regions promoting a transcriptional program favorable to breast cancer progression.
| Approximately two-thirds of breast cancers depend on the estrogen receptor alpha (ERα) for their growth. Its capacity to act as a transcription factor binding DNA following estrogen stimulation is central to promote a pro-tumorigenic transcriptional response. Importantly, different classes of ERα-positive breast tumors can be discriminated based on outcome. However, the underlying mechanisms driving these differences are unknown. Here we demonstrate that PBX1 acts as a pioneer factor recognizing a specific epigenetic modification to remodel chromatin and guide ERα genomic activity. This translates in a specific transcriptional program associated with poor-outcome in breast cancer patients. Even more, PBX1 expression alone is sufficient to identify a priori ERα-positive breast cancer patients at risk of developing metastasis. Overall, this study defines the mechanisms dependent on the pioneer factor PBX1 that drives an aggressive response in a subset of ERα-positive breast cancers. These features highlight the uniqueness of PBX1 and demonstrate its potential prognostic value.
| The implementation of transcriptional programs is central to the commitment of pluripotent cells occurring throughout development [1], [2]. Likewise, diseases commonly arise from altered transcriptional programs. This requires active reprogramming characterized by chromatin remodeling and altered epigenetic signature at lineage-specific functional genomic elements [2]–[5]. The estrogen receptor alpha (ERα) is a nuclear receptor central to breast cancer development. Upon estrogen stimulation, it binds at thousand of genomic loci defining its cistrome to promote a pro-proliferative transcriptional program [6]–[9]. Its genomic actions are in part dependent on the pioneer factor FoxA1 [6], [7], [8], [10], [11], [12], [13], [14]. Pioneer factors are an emerging class of DNA binding proteins. They play a central role in defining transcriptional programs as they can integrate and remodel condensed chromatin rendering it competent for transcription factor binding [6], [15], [16], [17], [18], [19]. Their recruitment to the chromatin is sequence specific and can be facilitated by an epigenetic signature dependent on histone methylation [6], [20].
PBX1 (Pre-B-cell leukemia homeobox 1) is a member of the Three Amino acid Loop Extension (TALE)-class homeodomain family required for diverse developmental processes including hematopoiesis [21], skeleton patterning [22], pancreas [23], and urogenital systems organogenesis [24], [25]. While it is best known as an oncoprotein when fused to E2A in pre-B-cell leukemia [26], it also contributes to prostate, ovarian and esophageal cancer [27]–[30]. It is also highly expressed in breast cancer [31]. PBX1 is a cofactor for homeobox (HOX) transcription factors as it increases their affinity and specificity to chromatin [32], [33]. However, recent interactome studies have revealed that 12% of PBX1 putative partners are non-homeodomain transcription factors [34], [35]. In agreement, PBX1 modulates the transcriptional activity of nuclear receptors such as the thyroid and glucocorticoid receptors and was recently proposed to act as a pioneer factor for the bHLH factor MyoD [36]–[38]. However, the contribution of PBX1 to chromatin structure and epigenetic signatures regulating transcription in ERα-positive breast cancer cells is unknown. In the present study, we have investigated the pioneer function of PBX1 towards ERα genomic activity in breast cancer.
Condensed chromatin constitutes a barrier for the recruitment of transcription factors to the DNA. FoxA1 binding at specific genomic regions allows for chromatin remodeling favorable to ERα recruitment at a subset of its cistrome [6], [8], [13], [19], [39]. However, ERα is recruited to thousands of FoxA1-independent sites across the genome [6]. To identify candidate pioneer factors guiding ERα recruitment to the chromatin at these sites we performed seeded motif analyzes using the Cistrome-web application (http://cistrome.dfci.harvard.edu/ap/). This revealed that over 85% of the ERα cistrome harbors the DNA motif recognized by PBX1 (Figure 1A and 1B). Noteworthy, the presence of the PBX1 motif in ERα binding sites was significantly different from another similar size cistrome (androgen receptor (AR) cistrome from LNCaP cells, p<1e-99) (Figure S1A).
Analyzing expression profiles from the NCI60 panel of cancer cells compiled on bioGPS (http://biogps.gnf.org) [40], [41] reveals that PBX1 is significantly co-expressed with ERα (co-expression coefficient 0.7784 using probe 205253_at) (Table S1). This was also revealed by comparing PBX1 mRNA expression across 47 distinct ERα-positive and negative breast cancer cells (p = 8.98e-7) (Figure 1C). ERα mRNA expression was also significantly correlated with ERα-histological status of breast cancer cells (p = 1.71e-8) (Figure 1C). These results are further supported by RT-qPCR, immunofluorescence and western blot analyzes in ERα-positive MCF7 and ERα-negative MDA-MB231 breast cancer cells demonstrating co-expression of ERα and PBX1 at the mRNA and protein level (Figure 1D). PBX1 is one of four PBX family members [33]. RT-qPCR against other PBX1 genes demonstrates that PBX1 is the predominant family member expressed in ERα-positive breast cancer cells (Figure S1B). Analyses of 41 independent breast cancer expression profile studies, such as van de Vijver study, demonstrate that PBX1 and ERα are also co-expressed in primary breast tumors (p = 2.72e-13 for the van De Vijver study and p≤1e-4 for all other studies) (Figure 1E) [42]. The correlation between ERα mRNA expression and ERα-histological status is also reported for the van de Vijver study (p = 2.27e-74) (Figure 1E).
To address the functional relation between PBX1 and ERα we assessed the role of PBX1 on estrogen-induced growth in the ERα-positive MCF7 breast cancer cells. PBX1 mRNA and protein levels were significantly depleted (∼70%) in MCF7 breast cancer cells transfected with one of two independent siRNA against PBX1 (Figure 2A and 2B). In agreement with a role for PBX1 in breast cancer [27], PBX1 depletion completely prevented the estrogen-induced proliferation of MCF7 breast cancer cells (Figure 2C and S2A-B). Importantly, PBX1 depletion in MCF7 breast cancer cells did not affect ERα or FoxA1 expression both at the mRNA and protein level (Figure 2D). Overall these results support a functional role for PBX1 in mediating the response to estrogen in ERα-positive breast cancer.
Estrogen signaling involves ERα activation and subsequent recruitment to the chromatin. Pioneer factors can therefore be identified through their role at the chromatin prior to estrogen treatment. Immunofluorescence assays against PBX1 in MCF7 breast cancer cells deprived of estrogen demonstrate its localization to the nucleus (Figure 3A). While PBX1 and FoxA1 have a similar nuclear distribution, confocal immunofluorescence analysis against FoxA1 reveals that it only partially overlaps with PBX1 (Figure 3A and Figure S3A and S3B). To demonstrate that PBX1 occupies the chromatin in MCF7 breast cancer cells we performed a ChIP-seq assay in cells maintained in full media. This identified 24254 high-confidence PBX1 sites (p≤1e-5) predominantly localized a distant regulatory elements (Figure 3B and Figures S4A and S4B, S5, S6, S7, S8). Directed ChIP-qPCR assays on 37 randomly selected PBX1 bound sites identified by ChIP-seq demonstrates that it is loaded to the chromatin in absence of estrogen (Figure S4B). Approximately 50% of the estrogen-induced ERα cistrome overlaps with PBX1 bound sites (Figure 3B). A significant overlap between ERα and PBX1 is also observed for all publically available ERα cistromes (Figure S9) [6], [7], [9], [43], [44], [45], [46], [47], [48], [49], [50], [51]. FoxA1 is loaded to the majority of these sites (Figure 3B). In fact, ChIP-reChIP assays in MCF7 breast cancer cells maintained in estrogen free media demonstrates that both pioneer factors co-localize on the chromatin at shared sites (Figure S11). Importantly, over 37% of the FoxA1-independent ERα binding sites overlap with PBX1 (Figure 3B). Expression profile analysis in MCF7 breast cancer depleted of PBX1 reveals that a 71% of estrogen-induced target genes are dependent on PBX1 (Table S2 and Figure S12). Importantly, the estrogen signature identified by this expression profile was highly enriched for genes defining ERα-positive primary breast tumors (p = 5.75e-10) [52].
To assess the relation between genome-wide binding and expression profiles we cross-examined the estrogen responsive gene lists (all estrogen responsive genes and PBX1-dependent estrogen responsive genes) defined in MCF7 breast cancer cells against the binding profiles for ERα, PBX1 and FoxA1. This was accomplished by determining the number of estrogen responsive genes (all or PBX1-dependent) harboring at least one binding sites shared or unique to a given factor within ±20 kb from their transcription start site (TSS). This was repeated for the null list consisting of all genes from the refseq gene list not regulated upon estrogen stimulation in MCF7 breast cancer cells. The ratio of estrogen responsive genes associated with binding events within ±20 kb of their TSS over the number of genes from the null list associated with binding events within ±20 kb of their TSS was then plotted in a radar format. Estrogen target genes were significantly associated with PBX1-ERα shared sites (7% of total estrogen-responsive genes) and PBX1-FoxA1-ERα shared sites (12% of total estrogen-responsive genes) (blue line, Figure 3C). FoxA1-ERα shared sites did not preferentially associate with estrogen regulated genes (Figure 3C). Remarkably, PBX1-dependent estrogen target genes were specifically associated with PBX1 unique and PBX1-ERα shared sites (red line, Figure 3C). This was validated through RT-qPCR against estrogen target genes dependent on PBX1, FoxA1 or both. Indeed, PBX1 depletion disrupted only the regulation of shared or PBX1-dependent estrogen target genes in MCF7 breast cancer (Figure 3D and Figure S13). Conversely, FoxA1 silencing impacted only the regulation of shared and FoxA1-dependent estrogen target genes (Figure 3D and Figure S13). Collectively, these data support the notion that PBX1 is required to regulate a specific subset of estrogen responsive genes. Moreover, they suggest that PBX1 is required for the implementation of an estrogen regulated transcriptional program distinct from FoxA1.
ERα-dependent transcriptional response is dependent on its recruitment to the chromatin following estrogen stimulation. To test if PBX1 directly impacts ERα genomic activity we first assessed PBX1 occupancy through ChIP-qPCR assays at known ERα binding sites in MCF7 breast cancer cells treated or not with estrogen. Focusing on both FoxA1-dependent and independent ERα binding sites overlapping with PBX1 (Figure S4C), our results demonstrate that PBX1 is pre-loaded on the chromatin prior to estrogen treatment and remains bound following estrogen treatment (Figure 4A). These sites were chosen from our genome-wide analysis since they are proximal to genes fundamental for breast cancer proliferation and ERα biology. For instance, Myc, CCND1, FOS and EGR3 are well-studied ERα targets promoting breast cancer growth and progression [53], [54], [55]. TFF1 (also known as PS2) is the prototypical estrogen target gene [56]. Sequential ChIP assays (ChIP-reChIP) against ERα and PBX1 in both estrogen treated and untreated MCF7 breast cancer cells demonstrates that both factors co-occupy the same sites following ERα recruitment (Figure 4B).
ChIP-qPCR assays against ERα in PBX1 depleted MCF7 breast cancer cells demonstrate that ERα recruitment following estrogen treatment is dependent on PBX1 (Figure 4C). Importantly, ERα recruitment is disrupted selectively at sites with pre-loaded PBX1 but not at PBX1-independent sites (Figure 4D and Figure S4D) thus ruling out the possibility of a widespread non-specific impact on ERα ability to bind DNA in cells depleted of PBX1. Overall these results demonstrate that PBX1 can occupy the chromatin prior to ERα recruitment and is required for its genomic activity driving estrogen target gene expression. This is in agreement with a role for PBX1 as a novel pioneer factor in breast cancer.
Chromatin structure inherently represents an obstacle for transcription factor activity. Through their ability to integrate and open condensed chromatin, pioneer factors act as molecular beacons for other transcription factors. Using FAIRE (Formaldehyde Assisted Isolation of Regulatory Elements) assays [39], [57] to measure chromatin condensation/openness prior to estrogen stimulation, we demonstrate that PBX1 acts as a pioneer factor. Indeed, genome-wide FAIRE-seq assays in MCF7 breast cancer cells [44] reveals that PBX1 occupied chromatin is already highly accessible (Figure 5A and Figure S14). Interestingly, the pioneering activity of PBX1 and FoxA1 is synergistic on shared sites (Figure 5A). Sites only bound by FoxA1 are the least accessible (Figure 5A). Comparing FAIRE signal in estrogen starved MCF7 breast cancer cells depleted or not of PBX1 through siRNA revealed a significant decrease in chromatin openness in PBX1-depleted compared to control cells at the majority of tested sites (Figure 5B). In agreement, we demonstrate that PBX1 depletion in MCF7 breast cancer cells seen at the mRNA and protein level (Figure 2A and 2B) also significantly decreases its occupancy on the chromatin (Figure S10B). These results suggest that PBX1 plays a central role in increasing chromatin accessibility essential for transcription factor recruitment further supporting its role as a pioneer factor in breast cancer cells.
Immunofluorescence, ChIP-seq assays and ChIP-reChIP against PBX1 and FoxA1 suggests that they co-occupy genomic regions in MCF7 breast cancer cells (Figure 3A and 3B, Figures S3A and S3B, S9, S10, and S11). To determine if they collaborate with each other at these genomic regions or if they are part of a common complex we profiled FoxA1 binding following PBX1 depletion in estrogen starved MCF7 breast cancer cells. In agreement with both pioneer factors acting independently of each other, FoxA1 depletion did not alter PBX1 binding to the chromatin (Figure 5C). Similarly, PBX1 depletion did not affect FoxA1 recruitment to the chromatin (Figure 5D). Overall, these results reveal that PBX1 acts as a pioneer factor guiding ERα genomic activity independently of FoxA1 in breast cancer.
Covalent modifications are a main staple of epigenetic regulation. Previous reports have demonstrated that methylation of histone H3 on lysine 4 (H3K4me) can define functional regulatory element [58]–[61]. Furthermore, cell type-specific distribution of the mono and di-methylated H3K4 (H3K4me1 and me2) epigenetic modifications are central to cell type-specific transcriptional responses [6], [59], [60]. In cancer cells, depletion of H3K4me2 interferes with FoxA1 binding to chromatin [6], [39]. However, the relationship between FoxA1 and H3K4me2 may not be unidirectional, recent evidence suggesting that FoxA1 can favor H3K4me2 deposition [62]. Genome-wide analysis revealed that H3K4me2 is present on approximately 50% of the PBX1 cistrome (Figure 5E). A similar proportion of FoxA1 cistrome overlaps with the H3K4me2 distribution in MCF7 breast cancer cells (Figure 5E). To test if H3K4me2 favors PBX1 binding to the chromatin we overexpressed H3K4me2 demethylase KDM1 (LSD1/BCH110) and determined PBX1 chromatin occupancy through ChIP-qPCR assays. KDM1 over-expression led to a significant reduction of bound PBX1 in estrogen starved MCF7 cells (Figure 5F). In contrast, PBX1 depletion had no effect on H3K4me2 levels and did not affect KDM1 expression (Figure S15A and S15B). Hence, similarly to FoxA1, the H3K4me2 epigenetic signature favors PBX1 binding.
ERα drives proliferation in over 70% of all breast cancers. Accordingly it serves both as a therapeutic target and prognostic factor [63]. In addition, ERα is to date the most exploited marker in the clinic and generally associates with good outcome [64]. FoxA1 does not appear to provide any additional power to discriminate breast cancer subtypes in comparison to ERα profiling [65]–[67]. To assess the prognostic value of PBX1 in breast cancer we performed a meta-analysis using breast tumor expression studies with follow-up data available through Oncomine (Compendia Bioscience, Ann Arbor, MI). We differentiated breast cancer patients according to high (top 10%) or low (bottom 10%) PBX1 mRNA levels and then generated Kaplan-Meier curves according to the metastasis-free survival status of breast cancer patients. In addition, we independently generated Kaplan-Meier curves using the KMplot web application [68]. Results derived from this analysis performed against FoxA1 confirmed previous reports limiting its prognostic value to identify ERα-positive breast cancers within all breast cancer subtypes. PBX1 expression did not discriminate outcome in these same patients (Figure 6A and 6B and B) Interestingly, while FoxA1 mRNA levels where predictive of ERα status, PBX1 levels were evenly distributed in the ERα-positive breast cancer subgroups or all-cases (Figure S17). By focusing our analysis on ERα-positive breast cancer patients (as defined by pathological staining) we revealed the prognostic value of PBX1. Indeed, ERα-positive breast tumors with high PBX1 expression levels are associated with a reduced metastasis-free survival compared to ERα-positive breast tumors with low PBX1 expression (p<0.002) (Figure 6C and Figure S16C and S16D). FoxA1 expression could not stratify metastasis-free survival within ERα-positive breast cancer patients (Figure 6D and Figure S16C and S16D) in agreement with the redundant prognostic value of FoxA1 and ERα [67].
These results are further supported by comparing the PBX1-dependent estrogen induced transcription (Table S2 and Figure S12) against expression profiled from breast tumors using Oncomine (Compendia Bioscience, Ann Arbor, MI). This reveals the strong correlation between PBX1-dependent estrogen target genes and twenty-two expression signatures typical of poor-outcome in breast cancer patients (ex: metastasis, mortality, recurrence and high grade) (p<0.01, O.R. >2) (Figure 6E). In contrast, the FoxA1-dependent estrogen target genes [44] are significantly associated with only one poor-outcome expression signature (mortality) from breast cancer (Figure 6E). Taken together, this suggests that PBX1 drives a very specific transcriptional response underlying progression in ERα-positive breast cancer and reveal the potential prognostic potential for PBX1 within this breast cancer subtype to predict outcome.
Accurate regulation of complex transcriptional programs is central to normal organ development. This is dependent on several layers of controls including DNA sequence, epigenetic signatures and chromatin structure. However, how these different elements are integrated to generate lineage-specific transcriptional programs and how they are affected in the course of disease development is ill defined. In particular, we still misunderstand how epigenetic signatures and chromatin structure affect the transcriptional response to estrogen stimulation in breast cancer. Here we demonstrate that PBX1 acts as a pioneer factor guiding ERα genomic activity in breast cancer (Figure 7). Indeed, PBX1 translates the H3K4me2-based epigenetic signature to remodel specific genomic domains rendering them accessible for ERα. PBX1 was show to be crucial for histone H4 acetylation [69] and previous reports focusing on the recruitment of MyoD and PDX1 to the chromatin in myeloid and pancreatic islet cells, respectively, were suggestive of the pioneering role of PBX1 [36], [70]. Considering that PBX1 plays a fundamental role in the development of diverse organs [21], [24], [25] and contributes to various types of cancers, namely leukemia, prostate, ovarian and esophageal cancers [26]–[30], its pioneering functions are likely to apply beyond breast cancer. Similarly, the genomic activity of a wide-range of transcription factors including both homeodomain (HOX, MEIS, etc) and non-homeodomain protein (MyoD, GR, TR, etc) is promoted by PBX1 [32], [33], [36], [37], [38], [71], [72]. Hence, PBX1 pioneering functions are expected to affect additional transcriptional programs.
Finally, we reveal that PBX1 and FoxA1 can co-occupy specific genomic regions in breast cancer cells. While co-occupancy of specific genomic region by pioneer factors, such as PU.1 and GATA1 has previously been reported [73], our results demonstrates that this translates into greater chromatin accessibility. Furthermore, we reveal that FoxA1-independent PBX1 bound sites are more accessible than PBX1-independent FoxA1 sites. In agreement, the estrogen induced transcriptional response is preferentially associated with ERα binding at PBX1 or PBX1-FoxA1 shared sites. This also relates to a distinct prognostic value for FoxA1 and PBX1. Indeed, while FoxA1 expression in ERα-positive primary breast tumors does not discriminate their metastasis-free outcome, elevated PBX1 expression has significant prognostic potential towards metastasis. Gene signatures such as the Oncotype DX or MammaPrint have been successfully employed in the clinic to discriminate outcome in breast cancer based mostly on their ability to identify specific breast cancer subtypes [74], [75]. However they do not perform as well when restricted to ERα-positive patients [76], [77]. Our study introduces PBX1 as a potential clinical tool with additive prognostic value to ERα. Indeed, all patients with ERα-positive metastatic breast cancer and half or more of ERα-positive early stage breast cancers develop resistance to endocrine therapies leading to a poor outcome [78]. Hence, it is fascinating to speculate a role for PBX1 in the development of drug resistance in breast cancer.
Taken together, these results reveal the intricate interplay between distinct pioneer factors required for the implementation of specific transcriptional response to estrogen in breast cancer and distinguishes PBX1 as a prognostic marker.
FoxA1-independent ERα binding sites across the genome were identified by subtracting the False Discovery Rate (FDR) 20% FoxA1 cistrome from the FDR1% estrogen-induced ERα cistrome from MCF7 breast cancer cells. This was accomplished using the bedfiles that specifies the genomic coordinates for the FoxA1 cistrome called by MAT available through the Cistrome website (http://cistrome.dfci.harvard.edu/ap/) using a cutoff based on the FDR 20% and the bedfile that specifies the genomic coordinates for the ERα cistrome called by MAT using a cutoff based on FDR 1%. These files were loaded on the Cistrome website and the FoxA1 bedfile was subtracted from the ERα bedfile using the “Operate on Genomic Intervals - subtract” [79]. To define the proportion of the ERα cistrome overlapping or not with FoxA1 harboring the PBX1 DNA recognition motif (Transfac M01017) we used the default settings of the “Integrative Analysis – Screen motif” function available on the Cistrome website.
Expression correlation between ERα and PBX1 from the NCI60 cancer cell panel using BioGPS (http://biogps.gnf.org). Expression correlation analysis between ERα and PBX1 in breast cancer cells or primary tumors was achieved using Oncomine (https://www.oncomine.com).
Venn diagrams were generated by defining the proportion of sites shared and unique between different bedfiles using the functions found under “Operate on Genomic Intervals” within the Cistrome website. Overlapping binding sites were defined by having at least one base pair in common. Genome structure correction (GSC) [80] was run to establish the significance of the overlap between datasets. The software was run with the following setting: (region fraction) -R = 0.2, (sub-region fraction) –S = 0.4 and basepair_overlap_marginal (-bm) as statistic text. P values for results presented on Figure S6A and S6B have been corrected using the Bonferroni post-test based on 12 comparisons.
For immunofluorescence, MCF7 cells were treated as previously described [81]. PBX1 was stained using PBX1 monoclonal antibody (Abnova Corporation). FoxA1 was stained using FoxA1 polyclonal antibody (Abcam). Secondary antibodies Alexa 488 and 555 were purchased from Invitrogen. Digital images were analyzed with ImageJ (http://rsbweb.nih.gov/ij/index.html).
MCF7 cells were maintained in phenol red-free medium (Invitrogen) supplemented with 10% CDT-FBS as described previously (Lupien et al. 2008) [6] prior to transfection. Following two days of estrogen starvation cells were transfected with siPBX1 #1 (Darmachon) or siPBX1 #2 (Invitrogen). Small-interfering RNA against Luciferase was used as a negative control [8]. Transfection was performed using Lipofectamine2000 according to manufacturer's instructions (Invitrogen). For cell proliferation assays, cell number or O.D. (450 nm) (WST-1 assay, Takara Bio Inc) was determined every 24 h after estrogen (E2) addition (1×10−8 M final). For expression assays, RNA was extracted 3 h following E2 stimulation.
RNA samples from siControl or siPBX1 treated MCF7 in the presence or absence of estrogen were hybridized on HT12 human beads array (Illumina Inc.). Analyses were performed using BRB-Array Tools Version 3.8.1. Raw intensity data were log2 transformed, median normalized and filtered to remove non-detected spots as determined by Illumina Software. The normalization was performed by computing a gene-by-gene difference between each array and the median (reference) array, and subtracting the median difference from the log intensities on that array, so that the gene-by-gene difference between the normalized array and the reference array is zero. Two class non-paired comparison analyses were performed by computing a t-test for each gene using normalized log-intensities. Differentially expressed genes were determined at a significance level of p less than 0.01. A four class ANOVA at p less than 0.01 was also performed to identify genes expressed differentially across the four groups.
Hierarchical clustering was employed using a Euclidean distance measure to generate heat maps for subsets of significant genes using the open source software Cluster/Treeview. The data can be accessed in GEObrowser under superSeries GSE28008
FoxA1 dependent gene-signature was obtained from previously published microarray data [44].
ChIP qPCR was performed as described previously [82]. Antibodies against PBX1 (Abnova) FoxA1, H3K4me2 (Abcam) and ERα (Santa cruz biotechnology) were used in these assays. ChIP–reChIP was performed as described previously [83]. Statistically significant differences were established using a Student's t-test comparison for unpaired data versus an internal negative control. Primer sequences used in this assay are found in Table S3.
ChIP assay were conducted as described above. Library preparation for next-generation sequencing was performed according to manufacturer's instruction starting with 5 ng of material (Illumina Inc.). Single paired libraries were sequenced using the GAIIx (Illumina Inc). Over 28 and 31 million reads were generated through the GAIIx for the PBX1 ChIP and Input samples, respectively. Of those, 88% and 96%, respectively, were aligned to the human reference genome. These reads were aligned using the ELAND software. The MACS peak-calling algorithm was used to call significantly enriched peaks using default settings (P<10−5) and specifying the peak size = 200 bp. The data is accessible on the GEObrowser (accession number: PBX1:GSE28008 and H3K4me2:GSE31151).
FAIRE analysis was performed as previously described [39], [84]. FAIRE-seq data were already published [44].
MCF7 cells were maintained in DMEM (Invitrogen) supplemented with 10% FBS as described previously (Lupien et al. 2008) [6] prior to transfection. MCF7 cells were transfected with the pCMX-KDM1construct or the control empty vectors (10 µg per well in 6 well plates) using Lipofectamine 2000 DNA transfection reagent according to the manufacturer's instructions (Invitrogen). ChIP assays against PBX1 were performed 48 h post-transfection.
Several expression profiles [42], [63], [85], [86], [87], [88], [89], [90], [91] compiled in Oncomine (https://www.oncomine.com) were used to define PBX1 and FoxA1 mRNA expression levels. ERα stratification was based on protein levels provided in each independent expression study employed in this analysis. Samples were ranked according to processed probe signal provided by each independent expression study (Max to Min) and top and bottom 10% were classified as high and low expression respectively. Each sample was then matched with its associated outcome with a 1, 3 and 5 years follow-up provided by each independent study (metastasis-free survival: alive or dead). Statistical analyses were performed using Fisher exact test.
PBX1-dependent or FoxA1-dependent estrogen (E2) upregulated gene signatures [44] were analyzed against several expression profiles previously shown to be significantly associated with breast cancer outcome using Oncomine. [86], [87], [88], [90], [92], [93], [94], [95], [96], [97], [98], [99], [100], [101], [102], [103] Significant association was established at a pValue of at least <0.01 and an Odds Ratio >2.
|
10.1371/journal.ppat.1006447 | Rapid identification of genes controlling virulence and immunity in malaria parasites | Identifying the genetic determinants of phenotypes that impact disease severity is of fundamental importance for the design of new interventions against malaria. Here we present a rapid genome-wide approach capable of identifying multiple genetic drivers of medically relevant phenotypes within malaria parasites via a single experiment at single gene or allele resolution. In a proof of principle study, we found that a previously undescribed single nucleotide polymorphism in the binding domain of the erythrocyte binding like protein (EBL) conferred a dramatic change in red blood cell invasion in mutant rodent malaria parasites Plasmodium yoelii. In the same experiment, we implicated merozoite surface protein 1 (MSP1) and other polymorphic proteins, as the major targets of strain-specific immunity. Using allelic replacement, we provide functional validation of the substitution in the EBL gene controlling the growth rate in the blood stages of the parasites.
| Developing a greater understanding of malaria genetics is a key step in combating the threat posed by the disease. Here we use a novel approach to study two important properties of the parasite; the rate at which parasites grow within a single host, and the means by which parasites are affected by the host immune system. Two malaria strains with different biological properties were crossed in mosquitoes to produce a hybrid population, which was then grown in naïve and vaccinated mice. Parasites with genes conveying increased growth or immune evasion are favoured under natural selection, leaving a signature on the genetic composition of the cross population. We describe a novel mathematical approach to interpret this signature, identifying selected genes within the parasite population. We discover new genetic variants conveying increased within-host growth and resistance to host immunity in a mouse malaria strain. Experimental validation highlights the ability of this rapid experimental process for generating insights into malaria biology.
| Malaria parasite strains are genotypically polymorphic, leading to a diversity of phenotypic characteristics that impact on disease severity. Discovering the genetic basis for such phenotypic traits can inform the design of new drugs and vaccines. Both association mapping and linkage analyses approaches have been adopted to understand the genetic mechanisms behind various phenotypes of malaria parasites [1–5] and with the application of whole genome sequencing (WGS), the resolution of these methodologies has been dramatically improved, allowing the discovery of selective sweeps as they arise in the field [6]. However, both approaches suffer from drawbacks when working with malaria parasites: linkage mapping requires the cloning of individual recombinant offspring, a process that is both laborious and time-consuming, and association studies require the collection of a large number of individual parasites (usually in the thousands) from diverse geographical origins and over periods of several months or years to produce enough resolution for the detection of selective sweeps.
Linkage Group Selection (LGS), like linkage mapping, relies on the generation of genetic crosses, but bypasses the need for extracting and phenotyping individual recombinant clones. Instead, it relies on quantitative molecular markers to measure allele frequencies in the recombinant progeny and identify loci under selection [7, 8]. This approach bears similarity to Bulked Segregant Analysis (BSA) [9], a technique developed to study disease resistance in plants. In BSA, individuals from a population are segregated based upon their phenotype (e.g. disease resistance), following which the frequencies of genetic markers in each population are analysed, identifying loci at which different alleles are found for the differently phenotypes populations. Segregating individuals by phenotype, while relatively straight forward for large organisms such as plants, is not feasible for unicellular pathogens such as malaria parasites. Instead, in LGS, the segregating population is grown both in the presence or absence of a selection pressure (e.g. drug treatment, immune pressure, etc.). Selection removes susceptible individuals in the selected “pool”, while leaving both susceptible and resistant individuals in the unselected “pool”. In its original implementation, LGS was successfully applied in studying strain-specific immunity (SSI) [10, 11], drug resistance [7, 12] and growth rate [8] in malaria and SSI in Eimeria tenella [13]. LGS is essentially identical to the extreme QTL approach (xQTL) that was independently developed by yeast researchers based on BSA [14].
In both the original implementations of BSA and LGS a limiting factor is the availability of molecular markers differentiating the two populations. One step in increasing the number of molecular markers was through the use of array hybridisation that allowed the identification of thousands of SNPs as molecular markers in Arabidopsis thaliana [15]. BSA (still using pre-selected pools) was also combined with tiling microarray hybridisation and used probe intensities to detect a gene underlying xylose utilisation in yeast [16]. The xQTL method increased the power and rapidity of the approach by making use of available yeast microarray data as well as Next Generation Sequencing (NGS) of DNA hybridised to microarray probes to identify a large number of markers across the genome, this time comparing selected and unselected populations, rather then generating pools based on phenotype [14, 17]. In the absence of microarray databases, an alternative approach was to use NGS short reads to identify genome-wide SNPs between two parents and then use these SNPs as molecular markers to identify target genes in the selected progeny population compared against the unselected population, as done to study chloroquine resistance in malaria [18].
In this study, we apply an improved LGS approach for the identification of genes controlling two independent and naturally occurring phenotypic differences between two strains of the rodent malaria parasite Plasmodium yoelii; growth rate, and strain-specific immunity. A mathematical model, built upon methodological improvements in the analysis of genetically crossed populations [19, 20], was developed to analyze the data. This modified LGS approach relies on the generation and selection of at least two independent crosses between the strains. The progeny from both crosses pre- and post-selection are then subjected to high-throughput WGS, and SNP marker movement analyzed using best fitting modeling (Fig 1).
Our novel statistical framework both accounts for the influence of clonal growth in the cross population, and allows for a locally variable recombination rate in the parasite population, unlike previous analyses applied to comparable data [21]. Applying this framework to crosses between two strains of P. yoelii that induce SSI, and which differ in their growth rates, we were able to identify three genomic regions and alleles controlling both phenotypes, demonstrating that the approach can be used to analyze multiple complex phenotypes concomitantly with high genomic resolution within a short space of time.
The difference in blood-stage parasite growth rate between the two clones was followed in vivo for nine days in CBA mice. A likelihood ratio test using general linear mixed models indicated a more pronounced growth rate for 17X1.1pp compared to CU clone by time interaction term, L = 88.60, df = 21, p<0.0001, Fig 2A). To verify that the two malaria clones could also be used to generate protective SSI, groups of mice were immunized with 17X1.1pp, CU or mock immunized, prior to challenge with a mixture of the two clones (S1 Fig). The relative proportions of the two clones were measured on day four of the infection by real time quantitative PCR (Q-RT-PCR) targeting the polymorphic msp1 locus [22]. A strong, statistically significant SSI was induced by both parasite strains in CBA mice (Fig 2B).
Two kinds of selection pressure were applied in this study: growth rate driven selection and SSI. Two independent genetic crosses between 17X1.1pp and CU were produced, and both these crosses were subjected to immune selection (in which the progeny were grown in mice made immune to either of the two parental clones), and grown in non-immune mice. Progeny were harvested from mice four days after challenge, at which point strain-specific immune selection in the immunized mice, and selection of faster growing parasites in the non-immune mice had occurred. Using deep sequencing by Illumina technology, a total of 29,053 high confidence genome-wide SNPs that distinguish the parental strains were produced by read mapping with custom-made Python scripts. SNP frequencies from these loci from each population were filtered using a likelihood ratio test to remove sites where alleles had been erroneously mapped to the wrong genome location.
A hidden Markov model was applied to the data to identify allele frequency changes (Table 1) that were likely to have arisen from the clonal growth of individuals within the cross population or possible incorrect assembly of the reference genome, as described in the Materials and Methods section and in more detail in the supplementary mathematical methods (S1 Appendix). In a genetic cross population, an especially high fitness clone generated by random recombination events can grow to substantial frequency, this being manifested as sudden jumps in allele frequency occurring at the recombination points in this individual [23]. Jumps of this type were primarily identified in the 17X-immunized population, where the increased virulence of the 17X strain had less of an effect in driving alleles to high frequency, and in the first replica experiment; the data in the first experiment seemed to have been more affected by clonal growth in the population. The consistency of identified jumps between treatment conditions reflects the common origin of the differently treated populations; the jump at the end of chromosome XIV inferred in both replicas may be artefactual.
Based upon an analytical evolutionary model describing patterns of allele frequencies following selection, a maximum likelihood approach was used to define confidence intervals for the positions of alleles under selection in each of the genetic cross populations. In the absence of selection acting for a variant in a region of the genome, the allele frequencies in that region are expected to be locally constant. In common with a previous approach to identifying selected alleles [21], a search was therefore made for regions of the genome in which allele frequencies varied substantially according to their position in the genome. Next, wherever deviations of this form were consistently identified in both replica experiments a model of selection was applied to the data, inferring for each set of replica data the position in that region of the genome that was most likely to be under selection; this model was based upon expected changes in allele frequency under a constant local rate of recombination and is described further in the Methods section. Regions of the genome in which this inference of selection produced consistent results across replica datasets were then identified (Table 2). Of a total of 11 genomic regions suggesting evidence of non-neutrality, six showed sufficient evidence of consistent selection.
For each of these regions of the genome, a more sophisticated evolutionary model, accounting for variation in the local recombination rate, was then applied to the data, refining the position of the putatively selected allele. At this point, a putative selected allele in chromosome IV was removed from consideration, leaving five cases of potential alleles under selection in three regions of the genomes; confidence intervals for the positions of the selected loci are given in Table 3. Optimal positions of variant loci derived from each replicate are detailed in S1 Table; results of the variable recombination rate model are shown in S2 Table, with inferred recombination rates in S3 Table.
Of the final three putative loci, two were detected under multiple experimental conditions (Fig 3). When considering the combined largest intervals, a selective sweep was inferred at position 1,436–1,529 kb on Chromosome (Chr) XIII in replicate crosses grown in both non-immunized mice and 17X1.1pp-immunized mice, resulting from selection against CU-specific alleles at the target locus. A second sweep was inferred at position 1,229–1,364 kb on Chr VIII, detected in the parasite crosses grown in both CU and 17X1.1pp immunized mice, though not in the non-immunized mice. Here, selection pressure acted against different alleles according to the strain against which mice were immunized. The third sweep was detected at a locus between positions 725–814 kb on Chr VII. This event was only detected in mice replicates immunized with the 17X1.1pp strain, albeit that a consistent change in allele frequencies was also observed between replicas grown under these conditions (Fig 3B). The remaining loci (on Chrs VIII and XIII) were not consistently detected between replicates (S1 Table) and were thus considered to be non-significant.
All the genes in the combined conservative intervals of the three main loci under selection are listed in S4–S6 Tables, along with annotation pertaining to function, structure, orthology with P. falciparum genes and Non-synonymous/Synonomous SNP (NS/S) ratio in the P. falciparum orthologue, which is calculated by the PlasmoDB website (6.2) based on SNP data from 202 individual strains. These include both laboratory strains and field isolates obtained from six collections (see Methods for more details). The locus associated with SSI on Chr VIII contains 41 genes. We considered the presence of either transmembrane (TM) domains or a signal peptides as necessary features of potential antigen-encoding genes. Only 16 genes met these criteria. Functional annotation indicated 10 likely candidates among these; eight genes described as “conserved Plasmodium proteins”, and two encoding RhopH2 and merozoite surface protein 1 (MSP1). Of these genes, the P. falciparum orthologue of msp1 had the highest NS/S SNP ratio (8.43). MSP1 is a well characterized major antigen of malaria parasites that has formed the basis of several vaccine studies [24] and has been previously linked to SSI in Plasmodium chabaudi [10–12].
The locus under selection on Chr VII consists of 21 genes. Only seven contained TM domains and/or a signal peptide motif. Based on functional annotation, four of these could be potential targets for SSI. One of these genes, PY17X_0721800, encodes an apical membrane protein orthologous to Pf34 in P. falciparum. This protein has recently been described as a surface antigen that can elicit an immune response [25]. Three conserved proteins of unknown function (PY17X_0720100, PY17X_0721500 and PY17X_0721600) also displayed potential signatures as target antigens.
The growth rate associated selected locus on Chr XIII contains 29 genes. In this case, the presence of TM domains or signal peptide motifs were not considered informative criteria. Only eight genes contained NS SNPs between the parental strains 17X1.1pp and CU according to the WGS data. Among these was a duffy binding protein, Pyebl. Pyebl, is a gene that has been previously implicated in growth rate differences between strains of P. yoelii [8, 26]. A single NS SNP was predicted from the WGS data in this gene. Due to the very high likelihood of its involvement based on previous work, this gene was considered for further analysis.
Examining the Pyebl gene, Sanger capillary sequencing re-confirmed the existence in 17X1.1pp of an amino acid substitution (Cys >Tyr) at position 351 within region 2 of the encoded protein. When aligned against other P. yoelii strains and other Plasmodium species, this cysteine residue is highly conserved, and the substitution observed in 17X1.1pp was novel (Fig 4A). Crucially, no other polymorphisms were detected in the coding sequence of the gene, including in region 6, the location of the SNP previously implicated in parasite virulence in other strains of P. yoelii [8]. Structural modeling of the EBL protein in both wild-type and 17x1.1pp (C351Y) mutants predicted the abolition of a a disulphide bond between C351 and C420 in the mutant parasites that alters the tertiary structure of the receptor binding region of the ligand in these parasites (Fig 4B and 4C).
The functional role of this polymorphism was verified by experimental means. In order to study the functional consequences of the polymorphism, the Pyebl alleles of slow growing CU and faster growing 17X1.1pp clones were replaced with the alternative allele (i.e. CU-EBL-351C>Y and 7x1.1pp-EBL-351Y>C), as well as with the homologous allele (i.e. CU-EBL-351C>C and 17x1.1pp-EBL-351Y>Y). The latter served as a control for the actual allelic swap, as the insertion of the plasmid for allelic substitution could potentially affect parasite fitness independently of the allele being inserted. To establish whether the C351Y substitution affected EBL localization, as was shown for the previously described region 6 mutation, Immunoflurescence Analysis (IFA) was performed. This revealed that, unlike the known mutation in region 6 [8], the EBL proteins of 17X1.1pp and CU were both found to be located in the micronemes (Fig 5 and S2 Fig).
Transgenic clones were grown in mice for 10 days alongside wild-type clones. Pair-wise comparisons between transgenic clones with the parental allele against transgenic clones with the alternative allele (that is CU-EBL-351C>C vs CU-EBL-351C>Y and 17x1.1pp-EBL-351Y>Y vs 17x1.1pp-EBL-351Y>C) showed that allele substitution could switch growth phenotypes in both strains (Fig 6A and 6B). This confirmed the role of the C351Y mutation as underlying the observed growth rate difference.
RNA-seq analysis revealed that transfected EBL gene alleles were expressed normally, (S3 Fig), thus indicating a structural effect of the polymorphism on parasite fitness, rather than an alteration in protein expression.
The development of LGS has facilitated functional genomic analysis of malaria parasites over the past decade. In particular, it has simplified and accelerated the detection of loci underlying selectable phenotypes such as drug resistance, SSI and growth rate [7, 8, 10]. Here we present a radically modified LGS approach that utilizes deep, quantitative WGS of parasite progenies and the respective parental populations, multiple crossing and mathematical modeling to identify loci under selection at ultra-high resolution. This enables the accurate definition of loci under selection and the identification of multiple genes driving selectable phenotypes within a very short space of time. This modified approach allows the simultaneous detection of genes or alleles underlying multiple phenotypes, including those with a multigenic basis.
Applying this modified LGS approach to study SSI and growth rate in P. yoelii, we identified three loci under selection that contained three strong candidate genes controlling both phenotypes. Two loci were implicated in SSI; the first time LGS has identified multigenic drivers of phenotypic differences in malaria parasites in a single experimental set-up. The strong locus under selection in Chr VIII, associated with the gene encoding MSP1, is consistent with existing knowledge of malaria immunity. The Chr VII locus, which includes the orthologue of Pf34 as well as other potential unannotated antigens, underscores the power for hypothesis generation and gene detection of the LGS approach using multiple crosses.
Our approach also provided a genetic rationale for the difference in growth rate of the parental clones CU and 17X1.1pp. Phenotypically, this occurs due to the ability of 17X1.1pp to invade both reticulocytes and normocytes, while CU is restricted to reticulocytes [22]. Previously, differences in growth rates between strains of P. yoelii have been linked to a polymorphism in Region 6 of the Pyebl gene that alters its trafficking so that the protein locates in the dense granules rather than the micronemes [8, 26]. In the case of 17x1.1pp however, direct sequencing of the Pyebl gene revealed a previously unknown SNP in region 2, the predicted receptor-binding region of the protein, with no polymorphism in region 6. Consistent with this, the EBL protein of 17X1.1pp was shown to be located in the micronemes, indicating that protein trafficking was unaffected by the region 2 substitution. Allelic replacement of the parasite strains with the alternative allele resulted in a switching of the growth rate to that of the other clone, thus confirming the role of the substitution.
Region 2 of the Pyebl orthologues of P. falciparum and Plasmodium vivax [27–29] are known to interact with receptors on the red blood cell (RBC) surface. Furthermore, the substitution falls within the central portion of the region, which has been previously described as being the principal site of receptor recognition in P. vivax [29]. Wild-type strains of P. yoelii (such as CU) preferentially invade reticulocytes but not mature RBCs, whereas highly virulent strains are known to invade a broader repertoire of RBCs [30]. Further structural and functional studies are required to elucidate how the polymorphism described here enables mutant parasites to invade a larger repertoire of erythrocytes than wild type parasites. We show that the cysteine residue at position 351 in EBL forms a disulphide bond with a cysteine at position 420, and that this is abolished following the C351Y substitution, altering the tertiary structure of the binding region. This leads to the possibility that such an alteration of the shape of the binding domain may enable the ligand to bind to a larger repertoire of receptors.
LGS with multiple crosses offers a powerful and rapid methodology for identifying genes or non-coding regions controlling important phenotypes in malaria parasites and, potentially, in other apicomplexan parasites. Through bypassing the need to clone and type hundreds of individual progeny, and by harnessing the power of genetics, genomics and mathematical modeling, genes can be linked to phenotypes with high precision in a matter of a few months, rather than years. Here we have demonstrated the ability of LGS to identify multiple genetic polymorphisms underlying two independent phenotypic differences between a pair of malaria parasite strains; growth rate and SSI. This methodology has the potential power to identify the genetic components controlling a broad range of selectable phenotypes, and can be applied to studies of drug resistance, transmissibility, virulence, host preference, etc., in a range of apicomplexan parasites that are amenable to genetic crossing.
The applicability of the approach to human malaria species has been recently demonstrated: the original LGS approach was successfully applied to study P. falciparum immune evasion in mosquitoes in vivo [31], while we recently tested its applicability in vitro to detect loci under selection following antifolate drug treatment and in vitro growth rate competition. With the advent of humanized mice that are able to support the complete malaria life cycle, the generation of new genetic crosses between strains of human malaria has become more feasible, as recently demonstrated [32]. With the ability to maintain these crosses without the need of simian hosts, application of a broader range of selection pressures (excluding, for now, selection mediated by the presence of a complete immune response) is now more feasible in vivo, thus extending the application of the LGS approach to medically relevant malaria species.
Full and complete details of the mathematical methods are given in S1 Appendix, Supplementary Mathematical Methods.
Laboratory animal experimentation was performed in strict accordance with the Japanese Humane Treatment and Management of Animals Law (Law No. 105 dated 19 October 1973 modified on 2 June 2006), and the Regulation on Animal Experimentation at Nagasaki University, Japan. The protocol was approved by the Institutional Animal Research Committee of Nagasaki University (permit: 1207261005–2).
Plasmodium yoelii CU (with slow growth rate phenotype) and 17X1.1pp (with intermediate growth rate phenotype) strains [33] were maintained in CBA mice (SLC Inc., Shizuoka, Japan) housed at 23°C and fed on maintenance diet with 0.05% para-aminobenzoic acid (PABA)-supplemented water to assist with parasite growth. Anopheles stephensi mosquitoes were housed in a temperature and humidity controlled insectary at 24°C and 70% humidity, adult flies were maintained on 10% glucose solution supplemented with 0.05% PABA.
Plasmodium yoelii parasite strains were typed for growth rate in groups of mice following the intravenous inoculation of 1 × 106 iRBCs of either CU, 17X1.1pp or transfected clones per mouse and measuring parasitaemia over 8–9 days. In order to verify the existence of SSI between the CU and 17X1.1pp strains, groups of five mice were inoculated intravenously with 1 × 106 iRBCs of either CU or 17X1.1pp parasite strains. After four days, mice were treated with mefloquine (20mg/kg/per day, orally) for four days to remove infections. Three weeks post immunization, mice were then challenged intravenously with 1 × 106 iRBCs of a mixed infection of 17X1.1pp and CU parasites. A group of five naïve control mice was simultaneously infected with the same material. After four days of growth 10 μl of blood were sampled from each mouse and DNA extracted.
Strain proportions were then measured by Quantitative Real Time PCR using primers designed to amplify the msp1 gene [34]. All measurements were plotted and standard errors calculated using the Graphpad Prism software (v6.01) (http://www.graphpad.com/scientific-software/prism/). Wilcoxon rank sum tests with continuity corrections were used to measure the SSI effect, and were performed in R [35]. Linear mixed model analyses and likelihood ratio tests to test parasite strain differences in growth rate were performed on log-transformed parasitaemia by choosing parasitaemia and strain as fixed factors and mouse nested in strain as a random factor, as described previously [22]. Pair-wise comparisons of samples for the transfection experiments were performed using multiple 2-way ANOVA tests and corrected with a Tukey’s post-test in Graphpad Prism software (v6.01).
SNP frequencies were processed to filter potential misalignment events. We note that, during the cross, a set of individual recombinant genomes are generated. Considering the individual genome g, we define the function ag(i) as being equal to 1 if the genome has the CU allele at locus i, and equal to 0 if the genome has the 17X1.1pp allele at this locus. In any subsequent population of N individuals, the allele frequency q(i) at locus i can then be expressed as
q ( i ) = 1 N ∑ g n g a g ( i ) (1)
for some set of values ng, where ng is the number of copies of genome g in the population.
To filter the allele frequencies, we note that each function ag(i) changes only at recombination points in the genome g. As such, q(i) should change relatively smoothly with respect to i. Using an adapted version of code developed for the inference of subclones in populations [39], we therefore modeled the reported frequencies q(i) as being (beta-binomially distributed) emissions from an underlying diffusion process (denoted by x(i)) along each chromosome, plus uniformly distributed errors, using a hidden Markov model to infer the variance of the diffusion process, the emission parameters, and an error rate. A likelihood ratio test was then applied to identify reported frequencies that were inconsistent with having been emitted from the inferred frequency x(i) at locus i relative to having been emitted from an inferred global frequency distribution fitted using the Mathematica package via Gaussian kernel estimation to the complete set of values {x(i)}; this test filters out reported frequencies potentially arising from elsewhere in the genome.
Next, the above logic was extended to filter out clonal growth. In the event that a specific genome g is highly beneficial, this genome may grow rapidly in the population, such that ng becomes large. Under such circumstances the allele frequency q(i) gains a step-like quality, mirroring the pattern of ag(i). Such steps may potentially mimic selection valleys, confounding any analysis. As such, a jump-diffusion variant of the above hidden Markov model was applied, in which the allele frequency can change either through a diffusion process or via sudden jumps in allele frequency, modeled as random emissions from a uniform distribution on the interval [0, 1]. For each interval (i,i + 1) the probability that a jump in allele frequency had occurred was estimated. Where potential jumps were identified, the allele frequency data were split, such that analyses of the allele frequencies did not span sets of alleles containing such jumps. The resulting segments of genome were then analyzed under the assumption that they were free of allele frequency change due to clonal behavior.
Inference of the presence of selected alleles was performed using a series of methods. In the absence of selection in a chromosome, the allele frequency is likely to remain relatively constant across each chromosome. A ‘non-neutrality’ likelihood ratio test was applied to each contiguous section of genome, calculating the likelihood difference between a model of constant frequency x(i) and the variable frequency function x(i) inferred using the jump-diffusion model. Next, an inference was made of the position of the allele potentially under selection in each region. Under the assumptions that selection acts for an allele at locus i, and that the rate of recombination is constant within a region of the genome, previous work on the evolution of cross populations [19, 20] can be extended to show that the allele frequencies within that region of the genome at the time of sequencing are given by
x ( i ) = x + Δ x (2) x ( j ) = [ X + 1 2 ( 1 - X ) ( 1 + e - ρ Δ i j ] x + [ 1 2 X ( 1 - e - ρ Δ i j ) ] ( 1 - x ) + Δ x (3)
for each locus j not equal to i, where X is the CU allele frequency at the time of the cross, ρ is the local recombination rate, Δij is the distance between the loci i and j, x is an allele frequency, and Δx describes the effect of selection acting upon alleles in other regions of the genome. A likelihood-based inference was used to identify the locus at which selection was most likely to act. In regions for which the ‘non-neutrality’ test produced a positive result for data from both replica crosses, and for which both the inferred locus under selection, and the direction of selection acting at that locus were consistent between replicas, an inference of selection was made.
For regions in which an inference of selection was made, an extended version of the above model was applied, in which the assumption of locally constant recombination rate was relaxed. Successive models, including an increasing number of step-wise changes in the recombination rate, were applied, using the Bayesian Information Criterion [40] for model selection. A model of selection at two loci within a region of the genome was also examined. Given an inference of selection, a likelihood-based model was used to derive confidence intervals for the position of the locus under selection.
For each combined conservative interval of relevant loci under selection, genes were listed based on the annotation available in version 6.2 of PlasmoDB and verified against the current annotation (release 26). For each gene, information on predicted transmembrane domains, signal peptides and P. falciparum orthologues. For the P. falciparum orthologues, the NS/S SNP ratios were obtained from PlasmoDB, based on the count of synonymous and non-synonymous SNPs found in 202 individual strains collected from 6 data sets stored on the website. More details on the data sets can be found at the following link: https://goo.gl/lUwKn1.
All primer sequences are given in Supplementary S7 Table. Plasmids were constructed using MultiSite Gateway cloning system (Invitrogen).
To assess the course of infection of wild type and transgenic parasite lines, 1 × 106 pRBCs were injected intravenously into five 8-week old female CBA mice for each parasite line. Since the 17X1.1p and CU-recipient strains were transfected on separate occasions, the transgenic lines were tested separately. Thin blood smears were made daily, stained with Giemsa’s solution, and parasitaemias were examined microscopically.
Since the atomic structures of EBL protein of P. yoelii Wild Type: (Py17X-WT) and its mutant P. yoelii (C351Y): (Py17X1.1pp) are not known, homology models were generated. The homology models were generated using P. vivax Duffy Binding Protein (PvDBP) atomic structure (PDB ID: 3RRC, [46] with the Swiss-Model server (https://swissmodel.expasy.org) [47–50]. The homology models showed maximum amino acid sequence homology of 32% with Py17X-WT EBL, compared to another homologous protein P. falciparum Erythrocyte Binding Antigen 140 (PfEBA-140/BAEBL) (PDB ID: 4GF2, [51], that had 26% sequence homology. These models were then subsequently stabilized by minimizing their energies for at least 10 times each, to attain reasonably well equilibrated structures using the YASARA server (www.yasara.org).
The prediction of disulfide bonds in our homology models were performed using DISULFIND (http://disulfind.dsi.unifi.it) [52–55]. Our analysis showed high probability of disulfide bond formation by this Cys351 residue. Confirming that C351 is a potential residue for forming a disulfide bond, the energy minimized stable homology models were subjected to Disulfide bond visualization to check whether the Cys351 is involved in any disulfide bond formation with any other Cys and what is the effect of the C351Y substitution.
The homology models along with their disulfide bonds were visualized (Fig 4B and 4C) and the images were obtained using the “Disulfide by Design 2.0” server (http://cptweb.cpt.wayne.edu) [56].
Code used in this project is available online from https://github.com/cjri/LGSmalaria
|
10.1371/journal.pmed.1002155 | Prophylactic Oral Dextrose Gel for Newborn Babies at Risk of Neonatal Hypoglycaemia: A Randomised Controlled Dose-Finding Trial (the Pre-hPOD Study) | Neonatal hypoglycaemia is common, affecting up to 15% of newborns, and can cause brain damage. Currently, there are no strategies, beyond early feeding, to prevent neonatal hypoglycaemia. Our aim was to determine a dose of 40% oral dextrose gel that will prevent neonatal hypoglycaemia in newborn babies at risk.
We conducted a randomised, double-blind, placebo-controlled dose-finding trial of buccal dextrose gel to prevent neonatal hypoglycaemia at two hospitals in New Zealand. Babies at risk of hypoglycaemia (infant of a mother with diabetes, late preterm delivery, small or large birthweight, or other risk factors) but without indication for admission to a neonatal intensive care unit (NICU) were randomly allocated either to one of four treatment groups: 40% dextrose at one of two doses (0.5 ml/kg = 200 mg/kg, or 1 ml/kg = 400 mg/kg), either once at 1 h of age or followed by three additional doses of dextrose (0.5 ml/kg before feeds in the first 12 h); or to one of four corresponding placebo groups. Treatments were administered by massaging gel into the buccal mucosa. The primary outcome was hypoglycaemia (<2.6 mM) in the first 48 h. Secondary outcomes included admission to a NICU, admission for hypoglycaemia, and breastfeeding at discharge and at 6 wk. Prespecified potential dose limitations were tolerance of gel, time taken to administer, messiness, and acceptability to parents. From August 2013 to November 2014, 416 babies were randomised. Compared to babies randomised to placebo, the risk of hypoglycaemia was lowest in babies randomised to a single dose of 200 mg/kg dextrose gel (relative risk [RR] 0.68; 95% confidence interval [CI] 0.47–0.99, p = 0.04) but was not significantly different between dose groups (p = 0.21). Compared to multiple doses, single doses of gel were better tolerated, quicker to administer, and less messy, but these limitations were not different between dextrose and placebo gel groups. Babies who received any dose of dextrose gel were less likely to develop hypoglycaemia than those who received placebo (RR 0.79; 95% CI 0.64–0.98, p = 0.03; number needed to treat = 10, 95% CI 5–115). Rates of NICU admission were similar (RR 0.64; 95% CI 0.33–1.25, p = 0.19), but admission for hypoglycaemia was less common in babies randomised to dextrose gel (RR 0.46; 95% CI 0.21–1.01, p = 0.05). Rates of breastfeeding were similar in both groups. Adverse effects were uncommon and not different between groups. A limitation of this study was that most of the babies in the trial were infants of mothers with diabetes (73%), which may reduce the applicability of the results to babies from other risk groups.
The incidence of neonatal hypoglycaemia can be reduced with a single dose of buccal 40% dextrose gel 200 mg/kg. A large randomised trial (Hypoglycaemia Prevention with Oral Dextrose [hPOD]) is under way to determine the effects on NICU admission and later outcomes.
Australian New Zealand Clinical Trials Registry ACTRN12613000322730
| Neonatal hypoglycaemia is common, with 30% of babies born at risk for hypoglycaemia and half of these developing low blood-glucose concentrations.
Babies who develop hypoglycaemia are at increased risk of neurodevelopmental impairment.
There are no strategies, other than feeding the babies formula, for preventing hypoglycaemia in at-risk babies,
Oral dextrose gel has been shown to be effective at treating neonatal hypoglycaemia.
The authors conducted a randomised controlled trial of prophylactic oral dextrose gel in babies at risk of developing neonatal hypoglycaemia (infants of mothers with diabetes, small, large, and preterm) to determine the most effective dose of oral dextrose gel to reduce the incidence of hypoglycaemia.
Four hundred and sixteen at-risk babies were randomised to receive either a standard (200 mg/kg) or high (400 mg/kg) dose of dextrose gel or placebo, either once or followed by three more standard doses before feeds.
The authors found that 200 mg/kg (0.5 ml/kg of 40% dextrose gel) was effective at reducing the incidence of hypoglycaemia (relative risk [RR] 0.68; 95% confidence interval [CI] 0.47–0.99, p = 0.04).
A single dose of prophylactic oral dextrose gel reduces the incidence of neonatal hypoglycaemia in babies born at risk.
Further research is needed to determine if oral dextrose gel also reduces rates of neonatal intensive care unit (NICU) admission and neurodevelopmental impairment.
Most of the babies in the trial were babies of mothers with diabetes, so these results may not be as applicable to babies from other risk groups.
| Approximately 30% of newborn babies require multiple blood tests for screening for neonatal hypoglycaemia under current guidelines. Half of these will develop hypoglycaemia [1], and an unknown proportion will experience brain damage and developmental delay as a result. Despite recommendations in clinical guidelines that prophylactic measures should be taken in babies at risk of neonatal hypoglycaemia [2–4], there currently are no strategies beyond early feeding for prevention [5]. Dextrose gel has been shown to be effective in treating neonatal hypoglycaemia, without detrimental effect on breastfeeding [6]. We therefore considered that it might also be effective as prophylaxis against neonatal hypoglycaemia. However, we first needed to determine an effective dose of dextrose gel to prevent neonatal hypoglycaemia.
A dose of 200 mg/kg glucose is the standard treatment dose of intravenous glucose administered as a “mini bolus” to babies with hypoglycaemia [7] and is also the dose demonstrated to be effective in treatment of neonatal hypoglycaemia with dextrose gel [6]. However, we considered that a single dose of 200 mg/kg glucose might not be adequate for prevention of hypoglycaemia, as babies at risk may have a prolonged nadir in blood glucose after birth [8–10] and higher plasma insulin concentrations and lower rates of hepatic glucose production in the first hours after birth than those not at risk [8]. We therefore also investigated both a higher single dose (400 mg/kg) and repeated doses in the first 12 h. Babies were randomised to the resulting eight dosage groups. The primary outcome was neonatal hypoglycaemia in the first 48 h.
The aim of this study was to determine a dose of 40% oral dextrose gel that will prevent neonatal hypoglycaemia in newborn babies at risk.
The trial was approved by the Northern A Health and Disability Ethics Committee of New Zealand (13/NTA/8).
We undertook this randomised, double-blind, placebo-controlled trial at two hospitals providing maternity and neonatal services (Auckland City Hospital and Waitakere Hospital) in Auckland, New Zealand. Eligible babies were infants of mothers with diabetes (any type of diabetes), late preterm (35 or 36 wk gestation), small (birthweight < 10th centile on population or customised birthweight charts or < 2.5 kg) or large (birthweight > 90th centile on population or customised birthweight charts or > 4.5 kg), or those with other risk factors (e.g., maternal medication such as β-blockers). Babies also satisfied all of the following inclusion criteria at the time of randomisation; ≥35 wk gestation, birthweight ≥ 2.2 kg, < 1 h old, no apparent indication for admission to a neonatal intensive care unit (NICU) (this included a special care baby unit), unlikely to require admission to a NICU for any other reasons, and mother intending to breastfeed. Exclusion criteria were major congenital abnormality, previous formula feed or intravenous fluids given, previous diagnosis of hypoglycaemia, admitted to a NICU, or imminent admission to a NICU. Mothers of babies who were likely to become eligible (maternal diabetes, likely late preterm birth, or anticipated high or low birth weight) were identified through lead maternity carers and antenatal clinics and provided with an information sheet before the birth. Written informed consent was obtained before the birth by a member of the research team and confirmed verbally after the birth.
The trial was prospectively registered with the Australian New Zealand Clinical Trials Registry, number ACTRN12613000322730. The study protocol is available online at https://researchspace.auckland.ac.nz/handle/2292/25006.
Eligible babies for whom consent had been obtained were randomised within the first hour after birth. We used computer-generated blocked randomisation with variable block sizes to assign babies to one of eight treatment arms. Allocation was to either 40% dextrose or placebo gel and to one of the following dose regimens: 0.5 ml/kg (200 mg/kg) once, 1 ml/kg (400 mg/kg) once, 0.5 ml/kg for four doses (total 800 mg/kg), 1 ml/kg once followed by 0.5 ml/kg for a further three doses (total 1,000 mg/kg) (Fig 1). The allocation ratios were dextrose:placebo 2:1, with the intention that the placebo arms would be combined for analysis, single:multiple dose 1:1, and low:high dose 1:1. Randomisation was stratified by centre and then by prioritised allocation to primary risk factor—i.e., if more than one risk factor was present, primary risk factor was allocated in the order of maternal diabetes, preterm, small, large, or other. For example, a baby who was both preterm and whose mother had diabetes was allocated the primary risk factor of maternal diabetes. We assigned twins independently. Research staff entered demographic and entry criteria data into an online randomisation website that provided a number corresponding to a numbered trial pack that contained either a single 5 ml prefilled syringe of either 40% dextrose gel or an identical-appearing placebo (2% hydroxymethylcellulose) or four numbered syringes of gel (1 x 5 ml and 3 x 2.5 ml, all containing either dextrose or placebo gel). Clinicians, families, and all study investigators were masked to treatment group allocation throughout the study and remain so for the planned follow-up.
Study gel was massaged into the buccal mucosa, either once at 1 h of age (0.5 ml/kg or 1 ml/kg) or an additional three times (0.5 ml/kg) before feeds in the first 12 h, with gel given no more frequently than every 3 h. Each dose of gel was followed by a breastfeed. Feeding was according to standard hospital guidelines, which included breastfeeding within 1 h after birth and then on demand not less than every 3 to 4 h. Supplemental formula was not given routinely but could be given on parent or clinician decision.
We measured blood glucose concentrations first at 2 h after birth. Subsequent blood glucose measurement was according to the local hospital protocol, with prefeed blood glucose measurements every 2 to 4 h for at least the first 12 h, and until there were three consecutive blood glucose concentrations of ≥2.6 mM. Babies who developed hypoglycaemia were managed by the hospital clinical team according to the standard clinical practice at each site, including supplementation with formula, treatment with 40% dextrose gel, and admission to a NICU for intravenous dextrose if required.
All blood glucose concentrations were measured in whole blood by the glucose oxidase method, either with a portable blood glucose analyser (i-STAT, Abbott Laboratories, Abbott Park, Illinois, United States) or a combined metabolite/blood gas analyser (e.g., ABL 700, Radiometer, Copenhagen, Denmark).
Babies whose parent(s) gave consent had a continuous glucose monitor sensor (iPRO2, Medtronic, MiniMed, Northridge, California, US) inserted subcutaneously into the lateral aspect of the thigh as soon as possible after birth. Interstitial glucose concentrations cannot be viewed on this monitor in real time and therefore were not available to clinicians and had no influence upon clinical management. The sensors remained in situ for at least 48 h. These data (n = 137) will be reported separately.
Parents were contacted at 3 d and 6 wk after birth to determine feeding method and were asked about any health events since discharge and parental perceived treatment allocation on the second occasion.
The primary outcome was hypoglycaemia, defined as any blood glucose concentration < 2.6 mM in the first 48 h after birth. Secondary outcomes were admission to a NICU (defined as admission for > 4 h); admission to a NICU for hypoglycaemia; hyperglycaemia (blood glucose concentration > 10 mM); breastfeeding at discharge from hospital (full or exclusive); received any formula prior to discharge from hospital; formula feeding at 6 wk of age; cost of care until discharge home (to be reported separately), and maternal satisfaction at 6 wk. An independent data monitoring committee undertook interim analyses after 120, 240, and 360 babies had been randomised and recommended that recruitment continue. The safety monitoring committee received reports of serious adverse events (seizures and death) and other adverse events: hyperglycaemia, late hypoglycaemia (blood glucose concentration < 2.6 mM for the first time after 12 h of age), delayed feeding (failure to establish breastfeeding without supplements by the end of day 3), and systemic sepsis [9].
Limitations of the trial intervention were defined as tolerance of gel (small spill [few drops], moderate spill [half of volume administered], and large spill [all of volume administered]) assessed by the clinician at time of administration, length of time to administer dose (time to massage gel into baby’s buccal mucosa), messiness (parental report), hyperglycaemia, late hypoglycaemia, delayed feeding, and acceptability of trial intervention to parent(s) (acceptable, some inconvenience, major inconvenience, or unacceptable).
A total of 416 babies were randomised between August 3, 2013, and November 13, 2014 (Fig 1).
Demographic and baseline characteristics were similar for all randomisation groups (Table 1). The median (range) birthweight was 3,190 (2,200, 5,255) g and gestational age 38 (35, 42) wk; 301/415 babies (73%) were infants of mothers with diabetes, and 199/415 (48%) were born by caesarean delivery. Primary risk factors for hypoglycaemia were similar across all treatment groups. A similar proportion of mothers in each group were uncertain as to the gel type the baby received (163/257 [63%] in those randomised to dextrose gel versus 77/126 [61%] in those randomised to placebo) or thought the baby had received dextrose gel (67/277 [24%] randomised to dextrose versus 34/126 [25%] randomised to placebo), demonstrating effective masking.
The overall incidence of hypoglycaemia was 186/415 (45%, 95% CI 40%–50%), and 32/415 babies (8%, 95% CI 5.5%–10.7%) were admitted to NICU, of whom most (23/415, 6%, 95% CI 3.7%–8.2%) were admitted for hypoglycaemia. For those babies who became hypoglycaemic, the median (range) age at first detection was 2.3 (1.1, 44.5) h. Formula was given to 232/415 (56%, 95% CI 51%–61%) during hospital admission, with the most common indications being medical intervention for hypoglycaemia (93/232, 40%, 95% CI 34%–47%) and maternal choice (73/232, 31%, 95% CI 26%–38%). At discharge from hospital, 279/407 (69%, 95% CI 64%–73%) babies were fully or exclusively breastfeeding.
There was no difference between placebo and dextrose gel groups in timing of gel administration or in volume of gel administered (Table 2). Thirty babies did not receive all doses of gel according to protocol, including 13 who were withdrawn from the trial prior to completing all doses, 7 with missed doses, and 8 who received an incorrect volume. These protocol deviations in gel administration occurred with similar frequency for babies allocated to dextrose gel and placebo (21/277, 8% versus 9/138, 7%). Only 1/209 babies (<1%) allocated to a single dose did not receive all allocated gel, compared to 26/206 babies (13%) allocated to multiple doses (RR = 0.04, 95% CI 0.01–0.28, p = 0.0013).
Twenty-four babies (24/415, 6%) were withdrawn from the trial after randomisation (Fig 1), although consent was given to obtain outcome data from the clinical records for these babies. Withdrawal rates were similar in babies randomised to dextrose gel and placebo (17/277 [6%] versus 7/138 [5%], p = 0.60) but were higher in babies randomised to multiple doses than in those randomised to a single dose of gel (17/206 [8%] versus 7/209 [3%], RR 1.05, 95% CI 1.00–1.11, p = 0.034). The most common reasons for withdrawal were parental concern about blood sampling (despite this not being determined by the trial protocol) and clinician uncertainty about the trial protocol.
When cumulative doses of dextrose gel were plotted against the odds of developing hypoglycaemia, with adjustment for sex, gestational age, and mode of birth (Fig 2), the odds of hypoglycaemia were not significantly lower when all dose regimes of dextrose gel were compared against placebo gel (p = 0.21). However, the 95% CI for the 200 mg/kg dose relative to placebo did not include unity.
In post hoc exploratory analyses, there was no difference in median blood glucose concentration between dose regimes (Fig 3). Amongst babies randomised to multiple doses of dextrose gel who became hypoglycaemic, 56/88 (64%) had done so before the time that they would have completed their allocated four doses of gel.
Babies randomised to any dose of dextrose gel were less likely to develop hypoglycaemia than those randomised to placebo (RR 0.79, 95% CI 0.64–0.98, p = 0.03; number needed to treat = 10, 95% CI 5–115, Table 3). They also developed hypoglycaemia later (dextrose 3.7 [1.1–44.5] h, placebo 2.1 [1.5–43.8] h, p = 0.03). However, the lowest blood glucose concentration in those who did experience hypoglycaemia was similar for babies randomised to dextrose gel or to placebo (2.3 [0.6–2.5] mM versus 2.1 [1.1–2.5] mM, mean difference 0.08 mM, 95% CI −0.02 to 0.18 mM, p = 0.13), as was the number of blood glucose measurements < 2.6 mM in the first 48 h (mean [standard deviation (SD)] 1.8 [1.1] versus 2.0 [1.2], mean difference −0.19, 95% CI −0.54 to 0.16 mM, p = 0.29). One quarter of the babies (109/415, 26%) received supplementary dextrose in addition to study gel, either as open label 40% dextrose gel or intravenous dextrose, with similar rates in dextrose and placebo gel groups (67/277 [24%] versus 42/138 [30%], RR 0.79, 95% CI 0.57–1.10, p = 0.17, Table 2).
There was no difference between dextrose and placebo groups in the rate of admission to a NICU (Table 3), although admission to a NICU for hypoglycaemia tended to be less common in babies randomised to dextrose gel (RR 0.46, 95% CI 0.21–1.01, p = 0.05). Rates of breastfeeding were similar in both groups at discharge (p = 0.92), on day 3 (p = 0.08), and at 6 wk (p = 0.53) (Table 3). Parental satisfaction did not differ for babies who received single doses rather than multiple doses (RR 1.06, 95% CI 1.00–1.13, p = 0.06) or for babies who received dextrose or placebo gel (RR 0.95, 95% CI 0.90–1.01, p = 0.12).
In post hoc subgroup analysis, the effect of dextrose gel on the incidence of hypoglycaemia was similar in babies with different primary risk factors.
Overall, gel was well tolerated, with 918 of 1,030 doses (89%) associated with no spill or a small spill, 33 (3%) with a moderate, and 13 (1%) with a large spill. At least one moderate or large spill was more common in babies receiving multiple doses than after single doses (RR 7.94, 95% CI 2.85–22.09, p < 0.001) but was not different between babies receiving dextrose and placebo gel (RR 1.09, 95% CI 0.55–2.17, p = 0.80).
Most doses took 5 to 10 min to administer. Taking longer than 5 min to administer a dose was more common for multiple doses than for single doses (RR 1.08, 95% CI 1.03–1.14, p = 0.0036) but was similar for dextrose gel and placebo gel administration (RR 1.05, 95% CI 0.99–1.11, p = 0.13).
Similarly, parents reported more messiness with multiple doses than with single doses of gel (RR 7.07, 95% CI 2.14–23.33, p = 0.0013) but no differences between dextrose and placebo gel (RR 1.00, 95% CI 0.44–2.29, p = 0.99). Most parents found the gel acceptable (364/402, 91%), with no differences between multiple and single doses or between dextrose and placebo gel.
No babies met the criteria for hyperglycaemia. There were no differences between treatment groups in the incidence of late hypoglycaemia or delayed feeding (Table 4).
Total limitations scores were similar in all treatment groups. However, more babies in the multiple dose group than in the single dose group experienced at least one limitation (score > 0, RR 1.03, 95% CI 1.00–1.07, p = 0.05), although this was similar in dextrose and placebo gel groups (RR 1.00, 95% CI 0.96–1.03, p = 0.83).
One baby developed seizures, without concurrent hypoglycaemia, that were not considered to be related to the intervention. There were no neonatal or infant deaths. No babies developed hyperglycaemia or systemic sepsis or had a first episode of hypoglycaemia after 48 h. Delayed feeding occurred in 170/405 (42%, 95% CI 37%–47%) babies, and late hypoglycaemia in 14/415 (3.4%, 95% CI 2.0%–5.6%), with similar rates in all treatment groups (Table 3).
Our findings show that the most effective and well-tolerated dose of prophylactic oral dextrose gel to reduce the incidence of neonatal hypoglycaemia in babies born at risk but without indication for a NICU admission is 200 mg/kg (0.5 ml/kg of 40% oral dextrose gel). Further, the intervention was easy to administer, well tolerated, acceptable to parents, and not associated with any adverse outcomes. Neonatal hypoglycaemia is a common problem, occurring in up to 15% of newborn babies and in 50% of those born at risk [1]. Management commonly includes supplementary feeds with formula milk and/or separation of mother and baby for admission to a NICU for more invasive management with intravenous dextrose. The use of formula milk is associated with decreased breastfeeding rates [10], and admission to a NICU separates mother and baby, making breastfeeding establishment more difficult as well as increasing health care costs. Other than feeding early [5], there are no effective interventions for prophylaxis of neonatal hypoglycaemia in babies at risk. This trial is the first to demonstrate that oral dextrose gel reduces the incidence of neonatal hypoglycaemia.
Neonatal hypoglycaemia occurs most frequently in the first 24 h after birth, with lower blood glucose concentrations of ≤2 mM occurring most often within the first 12 h [1]. Babies at risk of neonatal hypoglycaemia commonly receive repeated feeds of supplemental formula milk to manage low blood glucose measurements while maternal lactation is established [6,10]. Therefore, we decided to investigate the effect of a prophylactic regime of multiple doses of dextrose given within the first 12 h following birth. As with the single dose regime, we considered that both a standard and a higher initial dose might be of benefit.
Perhaps surprisingly, there was no evidence of a dose-response effect, with all dose regimes having similar efficacy and resulting in similar median blood glucose concentrations. However, the diagnosis of hypoglycaemia was later in babies randomised to dextrose gel, although the incidence of late hypoglycaemia was unchanged, suggesting that the main effect of dextrose gel may be in reducing the incidence of early hypoglycaemia. This is consistent with our finding that two-thirds of babies randomised to multiple doses who developed hypoglycaemia did so before the time that they would have received all doses. These babies had already met the primary outcome, and therefore, any benefit of subsequent doses in maintaining blood glucose concentrations would not be captured in this analysis. For the same reason, the effective sample size was less than the number randomized in the multiple dose groups, and it is therefore possible that the study was inadequately powered to detect differences between single and multiple dose groups. Within the single dose groups, there was no indication that higher doses might have been effective, with the proportion of hypoglycaemic babies in the 1 ml/kg dose group being closer to that in the placebo group than in the 0.5 ml/kg dose group.
Each dose of dextrose gel was followed by a breast feed. We anticipated that the dextrose gel would be rapidly absorbed into the buccal mucosa but alone would not be adequate to maintain blood glucose concentrations for the length of the period between feeds. Although early colostrum contains few calories, it contains many other factors that are important for early neonatal health, including metabolic regulation during the transition [11,12], and we considered it a priority to encourage early establishment of breastfeeding, with health benefits for both mother and baby [13,14]. It was also possible that the gel might stimulate insulin production. Although there is uncertainty whether increased blood glucose concentration in the early neonatal period does induce an increase in insulin production [15], transient neonatal hyperinsulinism is the likely mechanism underlying most transient neonatal hypoglycaemia [16].
It should be noted that we measured whole-blood glucose concentrations using the i-STAT portable clinical analyser, which utilises the glucose oxidase method and does not adjust the results to plasma glucose concentrations. Screening of babies at risk for neonatal hypoglycaemia is commonly performed using whole blood and bedside analysers, rather than plasma, because of the requirement for immediate results and the risk of glycolysis in specimens sent to the lab [2,6]. Plasma glucose concentrations are approximately 10% to 18% higher than whole-blood concentrations because of the higher water content of plasma [2]. Ten percent of blood tests in this trial were analysed using the blood gas analyser in the NICU, usually after the baby had been diagnosed with hypoglycaemia and admitted to the NICU. Good reliability between the i-STAT and blood gas analysers in measuring blood glucose concentrations in this population has previously been reported [17].
The eligibility criteria for this study were intended to select babies who would not need NICU admission for other reasons and were therefore most likely to benefit if hypoglycaemia could be avoided. Although admission to a NICU for hypoglycaemia appeared to be less common in babies allocated to dextrose gel, overall admission rates were similar in both groups. However, this dose-finding trial was not powered to detect a reduction in admission to NICUs or later neurodevelopmental outcomes, and therefore, a larger trial is needed to determine the effect of prophylactic dextrose gel on these important outcomes.
The potential negative impact of any supplement given during the neonatal period on breastfeeding [10,18,19] necessitated close monitoring of feeding during the trial. In particular, the use of dextrose gel for treatment of hypoglycaemia has previously been reported in one small trial to reduce the volume of formula taken at the subsequent feed [20], although a larger, more recent trial was more reassuring and demonstrated reduced formula feeding rates at 2 wk after treatment dextrose gel [6]. We found no effects of prophylactic dextrose gel on measures of infant feeding (receipt of formula, delayed feeding, or breast feeding at discharge, on day 3, or at 6 wk). However, although all mothers of babies in our trial intended to breast feed, 55% of babies received formula before discharge, and 42% had not established full breast feeding by 72 h. This is perhaps not surprising given that 72% of mothers had diabetes and 48% underwent caesarean delivery; both are risk factors for delayed onset of lactation [18]. There are few comparative data. One study of women birthing in a university hospital in the US, the majority of whom intended to breastfeed and whose babies had no risk factors for hypoglycaemia, reported in-hospital formula supplementation in 47%, with the commonest indication being perceived insufficient milk supply [21]. Furthermore, delayed onset of lactation (≥72 h after birth) has been reported in 35% of healthy women [22].
We used a predefined assessment of limitations to assist with selecting the most appropriate dose, as we anticipated that more than one dose might be effective in preventing neonatal hypoglycaemia. As this prediction of similar efficacy proved correct, we aimed to select the dose that would be most acceptable to clinical staff and parents, with fewest potential adverse effects and best tolerated by the baby. Although the weightings of each component of the limitation score were assigned arbitrarily, they were based on our consensus estimate of clinical importance. This was helpful in clarifying that multiple doses were more likely than single doses to be associated with spilling, slower to administer, and considered messy by parents. However, there were no differences between dextrose and placebo gel groups.
The commonest risk factor for hypoglycaemia in participants in this trial was infant of a mother with diabetes. This was in large part because women pregnant with potentially eligible babies were approached antenatally, and we were able to identify women with diabetes more readily than those in other risk groups. Although this may be considered a potential weakness of this study, the incidence of neonatal hypoglycaemia was similar amongst the risk groups. Furthermore, prespecified subgroup analysis did not show any difference in efficacy of dextrose gel to prevent hypoglycaemia dependent upon the primary risk factor, although our trial was not powered to investigate this difference. Strengths of this trial are the low cost of the intervention and ease of administration of the gel, with potential for positive impact on neonatal health globally.
This trial was designed as a dose-finding trial, with the most effective dose in prevention of hypoglycaemia to be used to inform a subsequent multicentre trial to determine the effect on admission to NICU and on important long-term neurodevelopmental outcomes. Since efficacy of dextrose gel in prevention of hypoglycaemia was similar in all dosage groups, but limitations were more common in babies randomised to multiple doses, we have selected 0.5 ml/kg as the dose to be used in our ongoing trial of dextrose gel prophylaxis (Hypoglycaemia Prevention with Oral Dextrose [hPOD], ACTRN12614001263684) [23]. This also has the advantage of being the same as the dose shown to be effective and safe in treatment of neonatal hypoglycaemia [6], thus minimising any risk of confusion between prophylaxis and treatment in prescription and administration of gel in a clinical setting.
We have shown that in term and late preterm babies at risk of neonatal hypoglycaemia but without indication for NICU admission, the incidence of hypoglycaemia can be reduced by a single prophylactic buccal dose of 0.5 ml/kg 40% dextrose gel at 1 h of age, with an average of ten babies needing treatment to prevent one baby developing hypoglycaemia. It remains to be determined if this will result in other clinically important benefits in the short term and any effects on long-term health.
|
10.1371/journal.pgen.1005902 | Mariner Transposons Contain a Silencer: Possible Role of the Polycomb Repressive Complex 2 | Transposable elements are driving forces for establishing genetic innovations such as transcriptional regulatory networks in eukaryotic genomes. Here, we describe a silencer situated in the last 300 bp of the Mos1 transposase open reading frame (ORF) which functions in vertebrate and arthropod cells. Functional silencers are also found at similar locations within three other animal mariner elements, i.e. IS630-Tc1-mariner (ITm) DD34D elements, Himar1, Hsmar1 and Mcmar1. These silencers are able to impact eukaryotic promoters monitoring strong, moderate or low expression as well as those of mariner elements located upstream of the transposase ORF. We report that the silencing involves at least two transcription factors (TFs) that are conserved within animal species, NFAT-5 and Alx1. These cooperatively act with YY1 to trigger the silencing activity. Four other housekeeping transcription factors (TFs), neuron restrictive silencer factor (NRSF), GAGA factor (GAF) and GTGT factor (GTF), were also found to have binding sites within mariner silencers but their impact in modulating the silencer activity remains to be further specified. Interestingly, an NRSF binding site was found to overlap a 30 bp motif coding a highly conserved PHxxYSPDLAPxD peptide in mariner transposases. We also present experimental evidence that silencing is mainly achieved by co-opting the host Polycomb Repressive Complex 2 pathway. However, we observe that when PRC2 is impaired another host silencing pathway potentially takes over to maintain weak silencer activity. Mariner silencers harbour features of Polycomb Response Elements, which are probably a way for mariner elements to self-repress their transcription and mobility in somatic and germinal cells when the required TFs are expressed. At the evolutionary scale, mariner elements, through their exaptation, might have been a source of silencers playing a role in the chromatin configuration in eukaryotic genomes.
| Transposons are mobile DNA sequences that have long co-evolved with the genome of their hosts. Consequently, they are involved in the generation of mutations, as well as the creation of genes and regulatory networks. Controlling the transposon activity, and consequently its negative effects on both the host soma and germ line, is a challenge for the survival of both the host and the transposon. To silence transposons, hosts often use defence mechanisms involving DNA methylation and RNA interference pathways. Here we show that mariner transposons can self-regulate their activity by using a silencer element located in their DNA sequence. The silencer element interferes with host housekeeping protein transcription factors involved in the polycomb silencing pathways. As the regulation of chromatin configuration by polycomb is an important regulator of animal development, our findings open the possibility that mariner silencers might have been exapted during animal evolution to participate in certain regulation pathways of their hosts. Since some of the TFs involved in mariner silencer activity play a role at different stages of nervous system development and neuron differentiation, it might be possible that mariner transposons can be active during some steps of cell differentiation. Interestingly, mariner transposons (i.e. IS630-Tc1-mariner (ITm) DD34D transposons) have so far only been found in genomes of animals having a nervous system.
| Almost all eukaryotic genomes contain transposable elements (TEs). Some of these, known as DNA transposons, move by a simple ‘cut-and-paste’ mechanism removing DNA from one site and inserting it into a new target site. Others, called retrotransposons, move via an RNA intermediate that is copied into DNA and integrated into the genome. The overall fraction of TEs that make up currently described genomes remains difficult to estimate due to the accumulation of several layers of such elements. These layers originate from TE amplification bursts at different periods during the evolution of the element, followed by ageing of the DNA sequence. Recent improvements in sequence analysis methods have showed that the human genome likely consists of at least 66–69% of repeated or repeat-derived sequences [1], which is much higher than the 45–50% that had been reported when this genome was first sequenced. This suggests that the extent to which genomes have been shaped by TEs has probably been underestimated for many eukaryotic species. Mobility, distribution and exaptation of certain TE sequences have been considered as important sources for expansion and diversification of transcriptional regulatory networks as well as for genetic innovations [2,3]. Today, DNA segments derived from TEs that were exapted or inactivated over time by accumulation of mutations appear as remnants of repeated sequences of various ages. While they are rare, active TEs are still present in the genome of extant species in which de novo insertions can generate genetic variations. In multicellular eukaryotes TE insertions must occur within the germinal lineage or during early development in order to be transmitted to the following generations. This leads to the suggestion that transposition into somatic cells had no value for the TEs or their host. However, in the early 1980’s evidence began to accumulate showing that somatic TE activity (i.e. single excision or excision followed by re-insertion) occurred at high frequency in animal taxa. This was first shown for a DNA transposon, Tc1 in the worm Caenorhabditis elegans [4]. Recently, somatic activity was also observed for mammalian LINE-1 and dipteran R2 retrotransposons [5,6]. Interestingly, all of these somatic transpositions occurred in primordial cells associated with neuron-related lineages during embryonic or metamorphic development.
Activation of TE transcription within some cell lineages requires that the factors silencing their expression be specifically switched off in these lineages. The Neuron-Restrictive Silencer Factor (NRSF) that corresponds to the Charlatan (Chn) protein in arthropods [7] and to the SPR-3/SPR-4 in nematodes [8], represses transcription of many neuronal genes in non-neuronal cell types and in neuronal stem cells prior to their differentiation. NRSF binds to a 21 to 30 bp long element called the Neuron-Restrictive Silencer Element [9] (NRSE). NRSF has never been shown before to interfere with TE transcription, even though NRSEs were found in human retrotransposons such as LINE2 [9,10] and that transcription of Tc1-like DNA transposons was shown to be activated during development of the Xenopus nervous system [11]. We report the existence of a silencer element located in the last 300 bp of the Mos1 transposase (MOS1) ORF that is functional in both vertebrate and arthropod cells. This silencer is able to interfere with the transposon promoter as well as with promoters of genes located downstream of the silencer sequence. We show that the presence and location of this silencer element is conserved in mariner-like elements (MLEs), even though their DNA sequences have significantly diverged. Our data reveal that YY1, NFAT-5, NRSF, Alx1, GAF and GTF proteins have binding sites within these silencer elements. Furthermore, our results are consistent with the hypothesis that these silencers function with the Polycomb Repressive Complexes (PRC). Together, mariner silencers might not only regulate the transcription of active MLEs, but might also modify the expression pattern of genes in which active or remnant MLEs are inserted.
Although it was originally used for another purpose (negative controls of transposition done in absence of a transposase source), a stable expression assay was used to investigate whether Mos1 was able to interfere with the expression of neighbouring genes. This assay consisted in transfecting HeLa cells with plasmids containing a Neomycin Resistance (NeoR) marker gene and one or two Mos1 DNA segments cloned upstream or downstream of the NeoR gene (Fig 1A). After two weeks of selection with G418, resistant colonies were stained and counted. The first evidence that Mos1 could decrease the expression of a marker gene located within its neighbourhood was obtained with the Δ1[NeoR]Δ2 construct which corresponded to a complete Mos1 element containing the NeoR gene inserted in its middle. Colony numbers were at least 20-fold lower with the Δ1[NeoR]Δ2 construct than with those obtained with the [NeoR] reference (Fig 1B). Further constructs were tested in an effort to locate the region responsible for the observed decrease in marker expression within Mos1, a region we refer to hereafter as the silencer element. Results obtained with the Δ3[NeoR]Δ4 and Δ5[NeoR] constructs were not different from those of [NeoR] (Fig 1B).
These observations supported two explanations: i) the silencer element was not in the non-coding terminal regions of Mos1, ii) the optimal activity of the silencer element was position-dependent and had an effect only when located downstream of the marker. The first explanation was supported by observations based on the [NeoR]Δ6 and [NeoR]Δ7 constructs which gave results similar to those obtained with the Δ1[NeoR]Δ2, indicating that the silencer was located within the 3’ half of the MOS1 open reading frame (ORF) corresponding to the Δ7 DNA segment (Fig 1B through 1D). The role of the position and orientation of the Mos1 silencer element was confirmed using four constructs in which the Δ7 DNA segment was cloned upstream or downstream of the NeoR gene, in positive (i.e. with the piece of MOS1 ORF on the same strand as the NeoR ORF) or negative orientations (Fig 1E). Only the [NeoR]Δ7+ construct showed a strong silencer effect. Hence, the Δ7 DNA segment had a silencer effect only when located downstream of the marker gene, in the positive orientation with respect to the NeoR marker. In addition, complementary experiments demonstrated that an intragenic Δ7 DNA segment in frame with a marker gene had a silencer effect on its expression since Δ7-GFP and MOS1-GFP fusions expression is similar and significantly lower (~5-fold) than the GFP control (Fig 1F). These data support that Δ7-MOS1 segment has a silencer effect when it is fused in frame within a gene and that the silencer effect could be operating with the pCMV promoter. These results were confirmed by RT-qPCR using total RNA extracted from transiently transfected cells and GFP specific primers.
To confirm that the Δ7-MOS1 segment contains a real silencer element, we used a transient luciferase expression assay that was previously validated to characterize silencer elements [12] (S1 Fig). Our first results were confirmed with this alternative approach since the only ratio lower than 1 was obtained with the HS2_P_Luc_Δ7+ plasmid (Fig 2A). In addition, they revealed that the expressions of the marker gene in the [NeoR]Δ7+ and HS2_P_Luc_Δ7+ constructs are of the same order of magnitude with respect to controls, [NeoR] (14.3x; Fig 1B and 1E) and HS2_P_Luc (11.7x; Fig 2A), respectively. Therefore, these data confirmed that the Δ7-MOS1 segment contained a silencer element which is more efficient when it is located downstream of the maker gene in a positive orientation. They also confirmed that the results obtained with our stable expression assay did not reflect an ability of the plasmid to be integrated into the genome, but the capacity of the NeoR gene to be expressed post-integration. Given the above observations we decided that rather than continuing with the transient assay we should use our stable expression assay to investigate the impact of the distance separating the marker and the Δ7 DNA segment by cloning a 1.2 or 2.7 kbp spacer between them. We observed that the 1.2 kbp spacer had little or no impact on Δ7 silencing activity while the 2.7 kbp spacer decreased its activity approximately 5-fold (Fig 2B). It is interesting to notice that a Δ7 DNA segment in the negative orientation located a few kbps away from the marker gene had silencing activity comparable to the one on the Δ7 DNA segment in the positive orientation. These results were verified using linearized constructs (S2 Fig) and were not uniform, suggesting that vector configuration is important in such experiments. An orientation effect similar to that previously observed in the absence of a spacer was found with linearized vectors containing a 3 kbp spacer.
The activity of the Mos1 silencer element was tested using [NeoR]Δ7+ and [NeoR]Δ7-, our stable expression assay and two other cellular lineages originating from distantly related species: Speedy cells [13] from Xenopus tropicalis (Amphibia) and Sf21 cells from Spodoptera frugiperda (Insecta). Our results showed that the Δ7 DNA segment had a silencer effect in both cellular systems (Fig 3A and 3B), suggesting that the protein factors with which it interferes are conserved in these two species. Interestingly, the orientation effect of the silencer was recovered in Speedy cells but was absent in Sf21 cells.
The mariner TE family consists of five sub-families designated cecropia, elegans/briggsae, irritans, mauritiana and mellifera/capitata [14]. Based on the phylogeny of its transposase Mos1 belongs to the mauritiana sub-family. The presence of a silencer element was surveyed within the Δ7 DNA segments of three MLEs, Himar1, Mcmar1 and Hsmar1, which respectively belong to the irritans, elegans/briggsae, and cecropia sub-families. Results obtained using the stable expression assay (Fig 3C) showed that the Δ7-MCMAR1 segment had a silencing activity with features similar to those of Δ7-MOS1 (i.e. in terms of intensity and orientation). The Δ7-HIMAR1 and Δ7-HSMAR1 segments also had silencing activity that was not significantly different from those of Δ7-MOS1 and Δ7-MCMAR1, but independent of their orientation. This result is important because it suggests that the presence of a silencer within the Δ7 DNA segment is a characteristic shared by all MLEs. It also suggests that protein factors conserved in most animal species that interfere with the mariner silencer elements might have conserved binding site motifs in their sequences.
Taking into account the sequence of the active promoter in Hsmar1 [15], a variant of transient luciferase expression assay was designed with luciferase expression plasmids containing the Mos1 and the Hsmar1 promoters (Figs 4A and S3). Our results with HeLa cells (Fig 4B and 4C) revealed that both promoters were active. pMos1 was found to be 10-fold less efficient than the early promoter for SV40 (pSV40) under these experimental conditions. pHsmar1 was found to be two-fold more active than the pSV40 contained in the P_Luc control. When their silencer were cloned downstream of the marker gene our results revealed levels of marker expression that were lower than those of the controls (3.3-fold for pMos1 and 2-fold for pHsmar1). This indicated that mariner silencers were able to negatively interfere with their own promoters. Because the closest transcriptional start site (TSS) upstream of the silencer element is that of the transposase ORF this mechanism is probably a way for MLEs to repress their transcriptional activity in their host cells and maintain active copies in a state of latency when host factors required for this repression are available.
In the next four sections we present the results of molecular and cellular biology investigations in which the Mos1 silencer was used as the main model to elucidate the mechanism of its activity. The silencer of Hsmar1, and in a few cases those of Himar1 and Mcmar1, were used as complements to confirm certain results. In the final section of the results pertaining to silencer activity at the scale of a eukaryotic genome, the Hsmar1 and Hsmar2 silencers were used, as they were the only models for which in silico genomic data are available.
NeoR expression was monitored for 24 hours both at the protein and mRNA levels using cellular extracts from cells transiently transfected with our constructs. Western-blot analyses (Fig 5A) revealed that the amount of neomycin phosphotransferase 2 (NeoR protein) was ~5-fold lower in cells transfected with [NeoR]Δ7- than with [NeoR]. However, few or no NeoR protein was detected in cells transfected with [NeoR]Δ7+. This was also supported by RT-qPCR experiments (Fig 5B) which showed that there were respectively 5 and 20-fold fewer NeoR transcripts in cells transfected with [NeoR]Δ7- and [NeoR]Δ7+ than in those transfected with [NeoR]. Taken together these results confirmed that the Mos1 silencer element interferes with the expression of a gene marker located immediately upstream, that it acts at the level of RNA, and that the strength of the effect depends on its orientation since the amount of NeoR transcripts in cells transfected with [NeoR]Δ7- was ~4-fold higher than in those transfected with [NeoR]Δ7+.
Because the silencer element had to be located within or downstream of the marker gene to be effective, we investigated whether it directly interfered with processes occurring after transcription initiation. Transcript quality and RNA interference were examined. Polyadenylation tails of transcripts from cells transfected with [NeoR], [NeoR]Δ7- and [NeoR]Δ7+ were investigated [17,18] according to their concentration in each sample, using GAPDH transcripts as endogenous controls. No difference was found, indicating that polyadenylation was unlikely to be affected. To test if the miRNA pathway was involved, Co115 human cells depleted in a key protein for miRNA processing and DICER function TARBP2 [19] were used in a transient expression assay. Similar silencing activity of [NeoR]Δ7+ was found in both Co115 (Fig 5C) and HeLa cells (Fig 5D), suggesting that there was no link between the silencing activity of Δ7 and the miRNA pathway.
In an attempt to locate a smaller fragment that would keep silencing activity in our expression assays the Δ7-MOS1 segment was fragmented (Fig 6A). The Δ8-MOS1 segment, corresponding to the last 317 bp of the MOS1 ORF was found to have the same silencing activity as the Δ7-MOS1 segment (S4 Fig). The size of the Δ8 segment is of interest since it was close to the upper limits for a usable protein electrophoretic mobility shift assays (EMSA) for investigating the binding of a TF.
Several viruses and transposable elements [20–30] were previously found to contain segments capable of silencing their own transcriptional activity to establish their latency in their eukaryotic hosts. These silencers are bound by the transcription factor Yin Yang 1 (YY1 in vertebrates, Pho in drosophila). In eukaryotes, YY1 and other TFs can bind a chromosomal polycomb response element (PRE) to mobilize the PRC1 and PRC2 and finally induce transcriptional silencing of that chromosomal region.
The presence of YY1 binding sites and TF binding sites involved in PRC2 in drosophila was examined in the Δ7 and Δ8 silencer segments of Mos1 (Fig 6B), Himar1, Mcmar1, Hsmar1, and Hsmar2 (S5 Fig). A set of binding sites for YY1 or Pho, the GAGA factor (GAF), GTGT factor (GTF), and Zeste [31,32] located among the Δ7 segments was found in all natural MLEs, each of which is typically repeated. Together the presence of these sites suggests that PRCs might be able to bind to these silencer elements, at least in dipteran species.
EMSAs were carried out to verify the presence of functional YY1 binding sites in Δ8 mariner segments. Our results showed that a shifted complex was present with the Δ8-MOS1, Δ8-HIMAR1 and Δ8-HSMAR1 probes (Fig 6C). These complexes were super-shifted by anti-YY1 antibodies confirming that they correspond to YY1/Δ8 complexes. The absence of a complex with the Δ8-MCMAR1 probe suggested that the binding site located in Δ8 was not bound under our experimental conditions. Hence, the silencing element in Mcmar1 extended beyond Δ8 and might be located at the 5’ extremity of Δ7, which contains a YY1 binding site (S5 Fig).
Since only one shifted band was observed with the Δ8-MOS1 segment while three binding sites were predicted in its sequence, further EMSA investigations were performed with shorter probes, Δ81 to Δ85 (Figs 6A and 6D and S4). These results were consistent with the sequence binding site prediction analysis, showing that there was one YY1 binding site within Δ81 and Δ83, two in Δ84, and none in Δ82 and Δ85. This last result suggested that the motif located in 3’ was unable to be bound by YY1 under our experimental conditions. However, this was likely an artefact due to its location at the 5’ end of the Δ85 probe. Indeed, when both YY1 sites are located in the middle of the Δ84 probe, two shifted bands were observed (Fig 6D, lane 12), suggesting that both sites could be bound. Taken together, these data supported the conclusion that the silencer activity of the Δ8 segments was possibly mediated by one or several YY1 silencing pathways.
Since the definition of PREs in vertebrate genomes is an issue that has yet to be fully elucidated [31,32], we searched for motifs conserved for both sequence and location using the MEME software suite with the DNA sequences of 34 mariner Δ7 segments (S6 Fig). A single conserved 30 bp motif was found (p-values ranging from 3.36e-21 to 6.11e-14) that spanned the region coding one of the two signature motifs of mariner transposases, the PHxxYSPDLAPxD peptide [33], and located in the region as a putative non-cardinal binding site for NRSF [34,35] and charlatan [36]. In mammalian genomes, approximately 80% of the 2 000 characterized NRSEs (called RE1) correspond to a 21 bp motif consisting of two conserved motifs of 9 and 10 bp separated by a 2 bp linker. Approximately 12% consist of multiple rearrangements of this motif [34–36]. The remaining 8% are sites with no conserved motifs. The putative NRSE in the mariner silencer element described above belongs to the second category.
EMSAs were carried out to assay these predicted NRSEs. HeLa nuclear extracts containing charlatan, human or fugu NRSF tagged with FLAG or Myc were prepared as described in previous studies [36,37]. The activity of the nuclear extracts was validated using an NRSE probe (RE1) in EMSA (S7 Fig) carried out with appropriate competitor and/or antibodies [36–38]. The binding of NRSF to Δ8-MOS1 was then further investigated with EMSA using shorter versions of Δ8, Δ81 to Δ85 segments as probes and HeLa nuclear extracts containing charlatan, human or fugu NRSF tagged with FLAG or Myc. No shifted bands were obtained with Δ83, Δ84 and Δ85 probes. By contrast, shifted complexes sensitive to the competition by a specific competitor (unlabelled RE1 fragment) were obtained with the Δ82 probe for the three NRSF proteins (Fig 7A). Since this probe contained the motif encoding the PHxxYSPDLAPxD peptide, we concluded that it was an NRSE.
To verify whether these NRSE intervened in the silencer activity under our experimental conditions, transient luciferase expression assays were performed using constructs with the Δ8, Δ8-ΔNRSF, or Δ8-mutNRSF of Mos1 or Hsmar1 (S8A and S8B Fig) cloned in positive orientation downstream of the marker cassette of HS2_P_Luc plasmids (Fig 7B and 7C). Δ8-MOS1-ΔNRSF and Δ8-HSMAR1-ΔNRSF were specified by the deletion of the NRSE motif and Δ8-MOS1-mutNRSF and Δ8-HSMAR1-mutNRSF by the mutagenesis of the NRSE by randomly shuffling its sequence. Results revealed that the DNA motif encoding the PHxxYSPDLAPxD peptide, i.e. the NRSE, was not essential for the silencer activity of the Δ8 silencers of Mos1 and Hsmar1.
Two shifted complexes were also obtained with the three NRSF proteins and the short Δ81 probe in which there was one YY1 binding site and no overlap with the NRSE encoding the PHxxYSPDLAPxD peptide (Fig 7D lanes 2, 5 and 8; S9 Fig, lanes 2 and 5). These two complexes were sensitive to competition by a specific competitor of the YY1 binding (Fig 7D, lanes 3 and 6; S9 Fig, lane 3), indicating that they involved YY1. Interestingly, we observed that the bigger complex (indicated by a red star in Figs 7D and S9) disappeared when antibodies directed against the tag of the human or fugu NRSF were added (Fig 7D, lane 9; S9 Fig, lane 6). Together these results indicated that there was a second NRSF binding site in the Δ8 silencer that required the cooperative binding of YY1 to be efficient. In spite of our efforts we failed to locate an NRSE or a charlatan binding element in this region. Therefore, it remains possible that NRSF only interacts with YY1 when it is bound to Δ8. In order to verify whether this second site of NRSF binding was required for the silencer activity two Δ8 variants for Mos1 (Fig 8A) and Hsmar1 (S8B Fig) were generated by PCR, cloned downstream of the marker gene into HS2_P_Luc plasmid constructs in positive orientation, and tested in transient luciferase expression assays in HeLa cells. Results obtained with constructs HS2_P_Luc_Δ8-MOS1-[47–311]-ΔNRSF (Fig 8B) and HS2_P_Luc_ Δ8-HSMAR1-[61–311]-ΔNRSF (S8C Fig) revealed that the silencer activity was conserved in spite of the fact that both regions bound by NRSF proteins in Mos1 were deleted in Δ8 segments.
Finally, these data supported that there were either one or two sites where NRSF was able to interfere with the Δ8 segment. However, the binding of NRSF to the mariner silencers of Mos1 and Hsmar1 was not essential to the silencer activity under our experimental conditions. Therefore, we continued our efforts to find out the regions essential for the silencer activity of the mariner Δ8 segment.
In addition to the variant Δ8-MOS1-[47–311]-ΔNRSF, four other variants were made (Fig 8A). The first, Δ8-MOS1-[1–152], contained the 5’ half of Δ8-MOS1 (i.e. one YY1 binding site plus the two NRSF binding sites). The second, Δ8-MOS1-[86–311], contained the 3’ half plus the NRSF binding site overlapping the DNA motif encoding the PHxxYSPDLAPxD peptide (i.e. two YY1 binding sites plus one NRSF binding site). The third, Δ8-MOS1-[116–311] was similar to the second with the exception that its NRSF binding site was removed. The fourth, Δ8-MOS1-[66–311]-ΔNRSF, was similar to Δ8-MOS1-[47–311]-ΔNRSF but its 19 residues on the 5’ end were deleted.
Transient luciferase expression assays in HeLa cells revealed that all of these Δ8 variants had kept their silencer activity but with variable efficiency (Fig 8B). For Δ8-MOS1-[1–152], Δ8-MOS1-[86–311] and Δ8-MOS1-[116–311], the luciferase expression is higher than the P_Luc control but, significantly, it was 2.5 to 3.5-fold lower than that of HS2_P_Luc (e.g. in Figs 2A and 5B). This indicated that the YY1 binding site motif within the 5’ half of Δ8 and other TF binding sites within the positions 116 to 311 were enough to trigger weak silencer activity in HeLa cells. Interestingly, it also indicated that Δ8-MOS1-[47–311]-ΔNRSF and Δ8-MOS1-[66–311]-ΔNRSF had better silencer activity than Δ8 in HeLa cells. Taken together, these results suggested that several combinations of TFs could bind to Δ8 and could cooperatively act with YY1 to trigger the silencer activity.
To verify whether this property could be recovered in another mariner silencer, Δ8 variants were also made from Δ8-HSMAR1 (S8B Fig). Their analysis under similar experimental conditions first revealed that the 3’ half (positions 130 to 310) was enough to trigger weak silencer activity in HeLa cells, but was more efficient when the DNA motif bound by NRSF and overlapping the DNA motif encoding the PHxxYSPDLAPxD peptide was present (positions 86 to 310). This indicated that this NRSF binding site favoured the silencer activity in Δ8-HSMAR1. In addition, the two variants Δ8-HSMAR1-[61–310]-ΔNRSF and Δ8-HSMAR1-[81–310]-ΔNRSF that have sequence properties similar to those of Δ8-MOS1-[47–311]-ΔNRSF and Δ8-MOS1-[66–311]-ΔNRSF have kept a strong silencer activity, but this was significantly less strong than that of Δ8.
A search for TF binding sites motifs within the regions 47 to 96 in Δ8-MOS1 and 61 to 104 in Δ8-HSMAR1 was achieved using the MatInspector facilities of the GENOMATIX software suite (Munich, Germany). Our results revealed that there were two NFAT-5 and one Alx1 binding sites in the 50 bp MOS1 segment (Fig 8A), and one NFAT-5 and one Alx1 binding sites in the 44 bp HSMAR1 segment (S8B Fig). Under the hypothesis that the same TFs acted on this region, results obtained with Δ8-MOS1-[47–311]-ΔNRSF and Δ8-MOS1-[66–311]-ΔNRSF on the one hand, and Δ8-HSMAR1-[61–310]-ΔNRSF and Δ8-HSMAR1-[81–310]-ΔNRSF on the other hand, suggested that both these TFs might cooperatively intervene in the silencer activity.
In order to further investigate this feature the HeLa, Co115, H4 and DT40 cells used in our work were phenotyped in order to determine their expression for NRSF, YY1, NFAT-5, Alx1 and TARBP2. We found that HeLa cells were NRSF +, YY1 +, NFAT-5 +, Alx1 + and TARBP2 +. Others cells presented differences since Co115 cells were TARBP2 -, DT40 were NFAT-5 -, and H4 were NRSF—and NFAT-5 -. Taking into account these phenotypes, the effect of the Δ7-MOS1 segments in transient luciferase expression assay was analyzed (Fig 5C and 5D and Fig 8C and 8D). Under these experimental conditions the absence of NFAT-5 significantly weakened the silencer effect of Δ7-MOS1 in H4 and DT40, but did not suppress it entirely.
In conclusion, our results suggested that TFs NFAT-5, Alx1 and NRSF, might intervene alone or cooperatively with YY1 to bind to the silencer of Mos1 and Hsmar1 and elicit the silencer activity. In addition, the weak silencer activity of the region located downstream of the DNA motif encoding the PHxxYSPDLAPxD peptide of Δ8-MOS1 and Δ8-HSMAR1 might be related to the cooperative binding of YY1 and GAF and/or GTF TFs.
The sequence features of the Δ8 mariner silencers described above might match those of the Polycomb Responsive Elements/Trithorax Responsive Elements (PRE/TRE) [31,39–42] that respectively silence or activate gene transcription by modifying chromatin histone marks. In order to further investigate whether the silencing depended on the polycomb pathway we used a specific inhibitor of PRC2, the 3-deazaneplanocin A (DZNep), in expression assays using plasmid constructs containing different variants of the mariner silencers of Mos1 (Fig 9) and Hsmar1 (S10 Fig). DZNep is an analogue of 3-deazadenosine that inhibits the activity of S-adenosylhomocysteine hydrolase, leading to the indirect inhibition of various S-adenosylmethionine-dependent methylation reactions, such as those catalysed by EZH2 in animal cells, including HeLa cells [43–45]. DZNep efficiently inhibits EZH2 after 8 h of treatment and can induce strong apoptotic cell death reaction in cancer cells beyond 48 h [43–47]. Cells were treated overnight with 5 μM DZNep prior DNA transfection and treatment was maintained until Firefly and Renilla luciferase activity measurements.
Before experimenting with DZNep on mariner silencers, the impact of this chemical was verified on the Firefly luciferase expression of the P_Luc and HS2_P_Luc constructs (Fig 9A). Results revealed that DZNep had no effect on P_Luc. However, this chemical decreased the capacity of the HS2 enhancer to boost the Firefly luciferase expression (~3.5-folds) even if the difference between P_Luc and HS2_P_Luc constructs remained significant. Because our silencer DNA segments were cloned into an HS2_P_Luc plasmid backbone, expression results obtained with HS2_P_Luc in the presence or absence of 5 μM DZNep were used as references to calculate the expression rate obtained with the mariner silencer constructs in the presence or absence of 5 μM DZNep (Fig 9B and S10 Fig).
Results showed that DZNep significantly increased (p<0.05) the Firefly luciferase expression from HS2_P_Luc_Δ8-MOS1, HS2_P_Luc_Δ8-MOS1-[47–311]-ΔNRSF, HS2_P_Luc_Δ8-HSMAR1, HS2_P_Luc_Δ8-HSMAR1-ΔNRSF and HS2_P_Luc_Δ8-HSMAR1-[61–310]-ΔNRSF. This supported the hypothesis that the Mos1 and Hsmar1 might contain functioning silencers depending on the PRC2 pathway, since the silencer activity of these constructs is decreased. Interestingly, five constructs responding to a DZNep treatment shared the presence of YY1, NFAT-5, Alx1, GAF and GTF binding sites in their DNA sequences.
By contrast, the Firefly luciferase expression from constructs containing the 5’ half of the Mos1 silencer (HS2_P_Luc_Δ8-MOS1-[1–152]) or the 3’ half of the Mos1 or Hsmar1 silencers (HS2_P_Luc_Δ8-MOS1-[116–311] and HS2_P_Luc_Δ8-HSMAR1-[115–310]) were not affected by the DZNep treatment. This suggested that the weak silencer effect resulting from the presence of these DNA segments might result from another silencing mechanism and is detected only when PRC2 is disrupted. Such a duality between silencing pathways was previously described for ITm TEs contained in the genome of murine ES cells. Indeed, these TEs can switch from heterochromatinization mediated by the HP1 (Heterochromatic Protein 1) dependent pathway to a PRC2-dependent silencing when the Histone-lysine N-methyltransferase Su(var)39/HP1 is disrupted [29]. Here, the duality between silencing pathways might also help explain why weak residual silencer effects were observed in some cases, such as in the H4 and DT40 cells (Fig 5C and 5D).
Since there is no available animal model with active MLEs for which high throughput chromatin data are available in public databases, we have investigated the chromatin status of two human MLEs, Hsmar1 and Hsmar2, that appeared in the human genome approximately 50 and at least 80 million years ago respectively [48,49]. Currently these elements have lost their ability to transpose due to the accumulation of nucleotide mutations in the ORF coding for their transposase. The advantage of the human model is that it has the richest set of ChIP-Seq data for TFs and histone modifications. Because the recruitment of TFs bound to DNA at the moment of the establishment of histone modifications is not subsequently required for their maintenance and transmission over cell divisions [50–53], we have focussed our investigations on histone modifications. This was carried out in order to highlight potential associations between the presence or the absence of a complete mariner silencer within each human MLE, their genomic location, and two important silencing pathways: polycomb/trithorax and Su(var)39/HP1. These two pathways lead to specific signatures of histone modifications: (i) H3K27me3 when the genomic loci is inactivated by PRC, (ii) H3K27me3/H3K4me3 when PRC and Trithorax complexes interfere together at level of inactive poised regions, (iii) H3K27ac/H3K4me3 when the genomic loci is activated by Trithorax complexes, and (iv) H3K9me3 and H4K20me1 when it is silenced and heterochromatinized by the Su(var)39/HP1 pathway [39,41,54,55]. Since it was previously shown that human TEs carry more histone modifications when they are located within or near genes [56], we have distinguished two categories of MLEs: those located in genes coding for proteins and those in inter genic regions. As a first step in our analysis, we inventoried the sequence features of Hsmar1 and Hsmar2 in the human genome using the hg19 RepeatMasker annotation (S11 Fig, S1 Table). Among the 592 and 1240 loci containing respectively an Hsmar1 or an Hsmar2 segment, 361 and 595 contained a nearly full-length copy and 315 and 644 contained Hsmar1 and Hsmar2 Δ8 silencers. Their chromatin status (Polycomb (P), Trithorax (T), Su(var)39/HP1 (H) or a mix of these statuses) was then inventoried in 14 human cell lines using CHIP-seq peaks (S2 Table; S12 Fig). In a second step, an analysis of the chromatin status was carried out at the scale of complete populations of Hsmar1 and Hsmar2 using a silencer definition in which the sequence of Δ8 segments was complete, absent, or damaged (Sil+, Sil- and U) and their genic or inter genic location in the human genome (S1 Table). Results indicated that the chromatin status was only statistically defined for 25 to 71% of loci, depending on the cell type and the features of the mariner element (S2 Table). Statistical analyses were carried out to test putative associations between the chromatin status, the presence of a silencer, and their genome location (S1 Table). A Wilcoxon test verified the associations between the percentage of Sil+ and Sil- and the polycomb status in cell lines. A Student t-test was used to search for associations between the quantity of polycomb status in Sil+ and Sil- loci. Only one robust association was found with both tests for Hsmar2. It revealed that Hsmar2 Sil+ has significantly more often a polycomb status than Hsmar2 Sil- in genic regions (p value = 0.00428 with the Wilcoxon signed-rank test and 0.02084 with the t-test, see methods). Features of Hsmar1 and Hsmar2 elements were therefore further investigated in order to i) verify whether genomic Hsmar1 silencer were still active and ii) verify whether the propensity of at least a part of Hsmar2 Sil+ to have a polycomb status was due to their activity.
We verified that at least a part of the Hsmar1 elements still contained an active silencer because remnants of human MLEs had accumulated significant amounts of mutations due to their age (S11 Fig). Eight Hsmar1 Δ7 segments were amplified by PCR from human gDNA, sub-cloned, sequenced, and located in hg19 (Fig 10A and 10B). These Hsmar1 Δ7 segments were then assayed with our stable expression and transient expression assays to verify their silencer activity. All of them were found to be strong silencers (luciferase/renilla ratio > 0.5; p<0.05; Fig 10C). Their putative co-localizations with CHIP-seq peaks on their Δ8 moiety were investigated and our results showed that the chromatin status was statistically defined for 59% of cases (S3 Table). They suggested that 50–56% of the 8 loci had a Su(var)39/HP1 status, whatever the cell type and the loci, the other 44–50% having polycomb status. In agreement with the literature [42], this suggested that the impact of these 8 strong silencers on the local chromatin status in somatic cells mainly depended on their genomic environment and the origin of the cells. If Hsmar1 silencers play a role in their host genome, our hypothesis is that they would intervene in chromatin organization during development or cell differentiation but not in adult somatic cells.
Because lacking data about the chromatin status (see loci with an undetermined chromatin status (ucs) in S1 Table) of human mariner silencers prevented the calculation of heat maps, only Sil+, Sil- and U located in intragenic regions and being annotated in at least seven cell lines were selected (187 loci, 95 Sil+, 67 Sil- and 25 U) to generate a heat map of the chromatin status Hsmar2 silencers (S13 Fig). Both cladograms on the top and the left of the heat map indicated that there was a suitable segregation of loci which were preferentially associated with a polycomb (green box) or a Su(var)39/HP1 status (yellow box), excepted for H1-hESC. This observation about hESC was in agreement with previous works indicating that hESC had a global chromatin status that is less marked than in somatic adult cells [57]. This heat map also allowed locating loci with a bivalent status (P/T loci in the blue box and P/H loci in the orange boxes). In agreement with our previous results, we observed that the density of Hsmar2 Sil+ loci associated with a polycomb status (91.5% of intragenic Sil+) was significantly above that of Sil- (73% of intragenic Sil- and U). Reciprocally, the density of Sil- associated to a Su(var)39/HP1 status (27% of intragenic Sil- and U) was significantly above that of Sil+ (8.5% of intragenic Sil+). Results with intragenic Sil- and U therefore suggested that only 20% of the Hsmar2 Sil+ would have a chromatin status depending on the activity of their silencer.
To verify whether the propensity of intragenic Hsmar2 Sil+ to be polycomb was due to their activity, we verified whether certain YY1 and NFAT-5 binding sites were significantly associated to the polycomb phenotype, taking into account that at least 1 YY1 and 1 NFAT-5 binding sites are required in an active mariner PRE. Because no result was statistically significant with the YY1 sites of Δ8 regions, sequences were extended in 5’ in order to match with a Δ7 segment. The YY1 and NFAT-5 binding sites were located in all Hsmar2 loci with a segment Δ7. In agreement with the Hsmar2 consensus sequence (S5 Fig), we found four YY1 binding sites at positions 11, 382, 431 and 475 (Fig 11A) of Δ7 segment and three NFAT-5 at positions 202, 293 and 294 (Fig 11B), all well conserved in numerous elements. For each binding site, a Wilcoxon test was used to verify the association between its presence and the propensity to have polycomb status in various cell lines. These tests revealed that the YY1 binding site at position 11 and the two NFAT-5 binding sites at positions 202 and 352 were significantly associated to loci with a polycomb status (p value = 0.014, 0.019 and 0.023 with the Wilcoxon signed-rank test, respectively). In agreement, the association of the YY1 site and one of the two NFAT5 sites in silencers was found to be significantly associated to loci with a polycomb status (p value = 0.017 with the Wilcoxon signed-rank).
Together, these results indicated that numerous intragenic Hsmar2 elements displaying the Δ7 region would contain a silencer still active in somatic cells. These results also confirmed that the size of a minimal mariner silencer was variable and depended on the MLE “species”. It corresponded to the Δ7 region in Hsmar2 and Mcmar1 (S5 Fig) and only to the Δ8 region in Himar1, Hsmar1 and Mos1.
Our work demonstrates that among all the MLEs we analysed all contain a silencer element within the last 300 to 500 bp of the transposase ORF. Under our experimental conditions the assayed silencers were found to have an optimal gene silencing effect when they were located from just downstream of the marker gene TSS to a few kpbs downstream of the gene transcription arrest site. We found that mariner silencers were able to silence strong (pCMV and pIE1), moderate (pSV40 and pHsmar1) and weak (pMos1) promoters, and were restricted by the host origin suggesting that they likely function with TFs that are conserved among animal species. Finally, our results support that mariner silencers largely function by promoting the PRC2 pathway, but they might also be able to trigger an alternate silencing pathway when PRC2 cannot be activated.
In agreement with our hypothesis that mariner silencers function with TFs conserved among animal species we found that their activity may depend on the binding of at least five TF candidates and YY1. The binding of NFAT-5 alongside with YY1 (NFAT in D. melanogaster) to the Mos1 and Hsmar1 silencers is likely the key for the activity of mariner silencers. Alx1 (Php13-Hazy in D. melanogaster) and NRSF are also be able to promote the silencer activity but with lower efficiency. In addition, expression data obtained with HS2_P_Luc_ Δ8-MOS1-[116–311] (Fig 8A and 8B) indicate that Δ8-MOS1-[116–311] keeps a weak silencer activity in spite of the absence of the NFAT-5, Alx1, and NRSF sites, and the YY1 site located near the 5’ end of the Δ8-MOS1 segment. This supports the conclusion that other TFs, such as GAF and-or GTF, might intervene in the silencer activity by binding to the 3’ half of the Δ8 segments (Figs 6B and S6). Since none of the Δ8-MOS1 and Δ8-HSMAR1 variants lost their silencer activity completely, it suggests that NFAT-5, Alx1, NRSF, GAF and GTF might function alone or more likely cooperatively with YY1 to trigger the silencing activity, depending on the cellular context. It should be noted that the variations in silencing efficiency of the Δ8-MOS1 and Δ8-HSMAR1 variants must be carefully considered. Indeed, they might also be due to the relative concentration of each TF in the various cell lines used in our assays rather than to the DNA affinity of each TF for the silencers.
Together, the profiles of TF binding sites in the mariner silencers looks like the numerous PRE/TRE that have previously been described in D. melanogaster and the few well-characterized PRE/TRE in mammal genomes [31,32]. Because the closest TSS upstream of these MLE silencers is that of their transposase ORF, these PRE/TRE were probably originally dedicated to the repression of the MLE transposon activity in cells expressing one or several of the identified TF candidates. Because MLEs have co-evolved with their animal hosts it is no surprise to observe that they have co-evolved to use conserved TFs and host housekeeping pathways to control their activity.
To our knowledge no functional link between NFAT-5, Alx1 and NRSF in adult insects or vertebrates have been published. Indeed, NFAT-5 is primarily implicated in the response to osmotic stress [58–60], Alx1 in osteogenesis during vertebrate development [61], and NRSF as a negative regulator of neuronal fate by silencing neuronal-specific genes in non-neuronal cells [62,63]. Furthermore; NFAT-5 and Alx1 appear to share with NRSF the property of participating at different levels in the development and differentiation of the nervous system [64–67]. NRSF was also reported to be involved in non-neuronal pathways of development or cell differentiation [67–69]. Even if the role of NRSF in the functioning of mariner silencers is not yet fully elucidated it suggests that the activity of these silencers in somatic cells might be dependent on the particular development pathway being used and the cellular environment. The fact that NRSF was involved in a polycomb dependent silencer is of interest. Indeed, this TF was found to have context dependent functions for the PRC1 and PRC2 recruitments [70–72] and is able to act as a recruiter for both complexes or as a limiting factor for the PRC2 recruitment [71]. NRSF is therefore an excellent candidate to positively or negatively regulate the commitment of mariner silencers in the polycomb pathway. Confirmation of a functional interplay between NRSF and MLEs would match well with the host range of these TEs. To our knowledge, MLEs are restricted to animal genomes having a nervous system (i.e. present in the genomes of cnidarians through arthropods and chordates). Nevertheless, it should be noted that in the original manuscript describing the discovery of mariner in D. mauritiana [73] mariner activity was not restricted to a particular cell type. Indeed, although the excision activity of Mos1 from the white peach locus was found to occur in neuron-like primordial cells of eye facets, it also occurs in the primordial cells of the larval Malpighian tubules and adult male testis sheaths.
The silencing machineries used by mariner silencers can also explain why neo-integrated Mos1 transposons are so stable and inefficient for remobilisation in transgenic insects [74]. Indeed, when silencing pathways promote and propagate H3K27 trimethylation in the neighbouring regions of their primary binding site [75], the MLE silencer element could extend the silent state of chromatin beyond the transposon, making it inaccessible for the transposase. Self-regulation using host silencing pathways is therefore potentially a mean to control MLE activity at two levels: transposase expression and transposon mobilization.
In somatic cells TEs can either move or rearrange themselves within the genome. Therefore, they need to be finely tuned to avoid deleterious side effects due to their activity. Until now it was the TE host that was most often considered the main actors of this control or defence against TEs, using epigenetic mechanisms including RNA interference (RNAi), DNA methylation and histone modifications to silence TE transcriptional activity. In spite of their widespread presence in animal genomes, master loci coding for small interfering RNA and other host mechanisms have not, so far, been demonstrated to be an important mechanism for repression of MLE transcription in animal genomes [76,77]. It is therefore possible that other mechanisms exist that control MLE transcription. Our results support that, just as certain viruses and endogenous retroviruses [20–30], MLEs control their activity using a self-regulation mechanism that uses the host polycomb machinery and certain host TFs. This self-regulation would not be the only mechanism that is controlled by MLEs. Indeed, cells that temporarily do not express TFs that elicit mariner silencers also show evidence of self-repression. Two other non-exclusive mechanisms were proposed to also mediate MLE self-repression. Beyond a certain threshold of transposase concentration, the first mechanism would lead to a partial or complete transposase aggregation outside the nucleoplasm, the compartment in which MLE transposition occurs. This sequestration would likely depend on certain host proteins [78]. The second mechanism would rely on communication between transposase subunits, their concentration, and the number of transposons that can be mobilized in the environment [79]. Whatever the features of their hosts and the role of these mechanisms, it is striking that MLEs might use certain host housekeeping pathways as the main modulator of their expression. This also applies to MITE derivatives that lack a silencer (e.g. MADE1 for Hsmar1 [80]), but their mobility is controlled via availability in the nuclear environment of transposases encoded by related functional elements.
Overall, data accumulated on the self-management of some herpesviruses and retrovirus latency by using host silencing machineries support the suggestion that some endogenous retroviruses and MLEs are themselves the main actors of their “latency” regulation in the germ line and the soma of their hosts. This view is a breakthrough compared to the widely accepted idea that the host restrain the activities of all TEs in its genome. It also suggests that some TEs might be able to master their own invasion dynamic within their host genome, and that this would vary depending on their ability to use the host silencing machineries.
This change in the conception of TE activity does not modify our understanding of their involvement in the host genome evolution. It is tempting to propose that insertions of MLEs might have had beneficial effects for their host's evolution by spurring the complexity of silencing regulatory networks [6,81]. The presence of human MLEs within genic regions supports this hypothesis. However, their distribution might also reflect their preference for inserting into gene loci. This could, for example, be because they would be more accessible to the MLE insertion complex. Silencers are not only located in gene promoters, several of them are scattered downstream of the TSS and the stop codon [82]. As supported by our data, such locations would not hamper the ability of each of these elements from participating in fine-tuning gene expression based on developmental stage, tissue, and cell type. Further investigations will be necessary to develop efficient experimental approaches to determine whether MLE silencers i) use one or several host silencing pathways to be effective, ii) have a silencer activity that is fully ubiquitous in animal bodies or have an activity that can cease at some steps of the life cycle, and iii) were exapted several times in order to intervene in the silencing regulation networks during evolution of animal taxa [83,84].
Although we indicated in the result section that the data of our in silico investigations must be viewed only as a prospective study, they suggest that part of the 109 Hsmar2 silencers located in genic regions have kept their ability to induce the PRC2 silencing pathway. Taking into account the lack of transposition activity of Hsmar2 in the human genome and their sequence degeneracy, the conservation of this ability to silence chromatin might have been exapted during evolution by the host genome. Therefore, it might correspond to a putative network of Hsmar2 PRE-like interspersed in certain genic regions of the human genome.
Concerning Hsmar1, we were disappointed when we did not obtain a correlation similar to that obtained with Hsmar2 silencers. Indeed, our experimental data in HeLa cells supported that at least a part of Hsmar1 silencers efficiently silenced gene expression. However, our in silico approaches failed to reveal an impact on local histones. This could be explained by two hypotheses. The first is that Hsmar1 elements have so far not been exapted for this functionality in the human genome. Therefore, the status of their current chromatin was gradually dictated by their genomic environment throughout their evolutionary sequence inactivation. The second implies that only a small population would currently be exapted in the human genome, which hampers localizing them with our analytical approaches. Previous reports support this second hypothesis since a small part of TEs (~5%) located near genes undergo purifying selection in mammal genomes, and might have regulatory functions at the levels of histone modification or gene expression [56,85–87]. Novel approaches will also be necessary to investigate the possible role of Hsmar1 and Hsmar2 silencers and whether they were also exapted during human evolution.
As noted above, MLEs have co-evolved with their animal hosts and it is therefore not a surprise to observe that they use certain housekeeping proteins to control their activity. Even if our results do not elucidate the involvement of NRSF in the functioning of the mariner silencers and its possible links with Alx1 and NFAT-5, we found information in various databases and in the literature indicating that part of the genes containing a mariner silencer might be related to the functioning of neuron and the central nervous system. Unfortunately, these preliminary data were not statistically confirmed using facilities of the GREAT platform [88].
Future investigations will require the development of specific approaches to further scrutinize and confirm the determinants of the mariner silencers. Another important issue will be elucidating what the development, differentiation, or physiological pathways are, how they might intervene, and to confirm that they were exapted during evolution of the human genome, and/or in any other animal genomes in which they are widespread.
Six cells lineages were used. Dr G. Sui (Harvard Medical School, ME USA) provided DT40 and DT40yy1- cells, Dr M. Esteller (CNIO, Madrid, Spain.) provided Co115 cells and Dr HY. Hwang (Standford University, USA) provided Speedy (known as 91.1.F1) cells. Sf21 cells were acquires from Sigma-Aldrich, HeLa-S3 and H4 cells from the ATCC.HeLa cells derived from human cervical cancer cells and H4 cells from malignant human glioma were cultured in DMEM (Gibco) supplemented with 10% fetal bovine serum (Gibco). DT40 cells from chicken B lymphoma were cultured in RPMI-Glutamine (Gibco) supplemented with 10% fetal bovine serum and 1% chicken serum (Gibco). The lineage of malignant human colorectal cells Co115 was cultured in RPMI-Glutamine supplemented with 10% fetal bovine serum. The Xenopus tropicalis speedy cell line [14] is a secondary lineage derived from a primary lineage established from a X. tropicalis limb. Cells were cultured in 67% (v/v) L15 medium adjusted to amphibian osmolarity by dilution with sterile water, supplemented with 10% heat inactivated fetal bovine serum (Sigma) and a cocktail of penicillin G (50U/mL) and streptomycin (50μg/mL) (Invitrogen). Sf21 cells from Spodoptera frugiperda ovary were cultured in Grace’s insect medium with L-glutamine (Gibco) supplemented with 10% fetal bovine serum.
pBlueScript SK+ plasmids were used as a vector backbone to make constructs for the stable expression assays. A [NeoR] marker cassette corresponding to a neomycin resistance gene coding a neomycin phosphotransferase 2 was cloned between the EcoRI and BamHI sites of pBS SK+. This gene was flanked by an early SV40 promoter (a moderate promoter) and an SV40 terminator except for the plasmids used in Sf21 cells, where the SV40 promoter was replaced by the immediate early protein 1 promoter (IE1; baculovirus AcMNPV). Each assayed DNA segment was cloned upstream (EcoRI site) or downstream (NotI site) of the marker in positive (+) or negative (-) orientation. DNA spacers of 1.2 kbp or 2.7 kbp were cloned between the 3’ end of the marker and the 5’end of the assayed DNA segment at the XbaI site as described [89].
Cells were co-transfected with approximately 150 ng of a two plasmids mix using jetPEITM as described by the manufacturer (Polyplus Transfection). Two third of the mix (100 ng) corresponded to the pGL3 plasmid (Promega), used to check for effective transfection. One third (50 ng) consisted of the assayed DNA plasmid. The amount of plasmid was fitted to its size with respect to that of the smallest plasmid used as a control in each experiment, [NeoR]. Two days after transfection, 1/3 of the transfected cells were evaluated for luciferase activity with the Luciferase Assay System Kit (Promega). The remaining 2/3 of the cells were transferred in 100 mm Petri dishes followed by G418 sulfate selection (800 μg/mL, PAA France) for 15 days. Cells were then fixed and stained with 70% EtOH-0.5% methylene blue for 3 h. Only colonies with a diameter > 0.25 mm were counted.
Plasmid constructs are presented in S2 Fig. The fragments pMos1, pHsmar1, Δ8-MOS1-ΔNRSF, Δ8-MOS1-mutNRSF, Δ8-HSMAR1-ΔNRSF, and Δ8-HSMAR1-mutNRSF were synthesized by ATG:Biosynthetics. To use the plasmids containing promoter pMos1 or pHsmar1 in transient luciferase expression assay, the NcoI-BamHI DNA fragment containing the luciferase ORF and an SV40 late polyadenylation signal was purified from the P_Luc plasmid, then cloned into each of both plasmids between NcoI and BamHI sites. In pMos- and pHsmar1-Luc plasmids, the BamHI site at the 3’ end of the luciferase cassette was used to clone the DNA fragment to assay. For the transient luciferase expression assay in HeLa and H4 cells, 6 x 104 cells were seeded onto a 24-well plate one day prior to transfection. Transfection was performed using jetPEI, according to the manufacturer’s instructions, using 400 ng of test DNA and 50 ng of pRL-Tk Renilla. For DT40 cells, 5 x 105 cells were seeded onto a 24-well plate one day prior to transfection. jetPEI was also used to transfect about 400 ng of test DNA and 50 ng of pRL-Tk Renilla. For Co115, 4 x 105 cells were seeded onto a 24-well plate one day prior transfection. For each test plasmid, its amount (400 ng) was fitted to its size with respect to that of the smallest plasmid used as a control in each experiment, P_Luc. Transfection was performed with ICAFectin441 DNA transfection reagent, according to the manufacturer’s instructions (In Cell Art), using 400 ng of test DNA and 400 ng of pRL-Tk Renilla. Luciferase expression was measured in a 96-well plate format with detection of fluorescence using the Dual-Glo Luciferase Assay System (Promega). Measurements were recorded on a Berthold plate-reader luminometer. Similar assays were used to investigate whether PRC2 was involved in the silencing effect observed with our constructs. However, a PRC2 inhibition was achieved by adding DZNEp (Sigma-Aldrich, USA) in the cell culture medium from the seeding until measuring Firefly and Renilla luciferase activities.
The expression profile of NRSF, YY1, TARBP2, Alx1, and NFAT-5 in HeLa, Co115 and H4 cells was determined by Western-blot analysis using commercial antibodies for NRSF (ab75785; Abcam), YY1 (ab12132; Abcam), TARBP2 (ab42018; Abcam), Alx1 (ABIN785202; antibodies-online GmbH) and NFAT-5 (ABIN183505; antibodies-online GmbH). For DT40 and HeLa cells, expression profile was determined by RNA-seq analysis using data available in databases, GEO datasets SRX286375 and SRX083286, respectively. During the analysis, we observed only one discrepancy between the RNA-seq data and the Western-blot analyses. Indeed, our HeLa cells were found to express NFAT-5 whereas the RNA-seq analyses done on another HeLa cell batch led to the opposite conclusion.
8 x 104 cells were seeded onto a 24-well plate one-day prior transfection and then transfected with jetPEI, according to the manufacturer’s instructions using 0.5 μg of plasmid DNA. Cells recovered from the culture 24 h post-transfection were washed three times with 1X PBS. The cell pellet was finally suspended in 400 μL 1X PBS-2% paraformaldehyde (w/v), and stored at 4°C. The analyses were performed using a flow cytometer FACSORT and the Cell Quest program (Beckton Dickinson). A total of 20 000 cells were acquired for each sample. Dead cells and debris were excluded from the analysis based on forward angle and side scatter light gating. Analysis gates were determined from the green fluorescence intensity using transfection controls done with or without plasmids expressing GFP.
Annotations from RepeatMasker were used to select positions of Hsmar1 and Hsmar2 in the hg19 human genome version. Those containing a Δ8 segment (Sil+), a damaged Δ8 segment (U), or no Δ8 segment (Sil-) and their location in a genic or an intergenic region were inventoried using home made perl scripts (S1 Table). Here, genic regions corresponded to those for which maximal efficiency of the mariner silencer was observed from our experimental data (i.e. from the TSS of each gene to 5 kbp downstream of its 3’end). Intergenic regions corresponded to any region of the genome that was not genic. ChIP-seq peaks files (EzH2, H3K27me3, H3K27ac, H3K4me3, and H3K9me3; no data about H4K20me3 were available for all cell lines) were located within UCSC resources for ENCODE data and downloaded at https://genome.ucsc.edu/ENCODE/dataMatrix/encodeDataMatrixHuman.html [86]. Intersections between the location of ChIP-seq peaks and Sil+, Sil- or U elements were performed using bioconductor. The chromatin status of each silencer was then inventoried and classified in four main categories: (i) undetermined chromatin (ucs) when no peak co-localized, (ii) Su(var)39/HP1 (H) when H3K9me3 peaks co-localized, trithorax (T) when H3K27ac and-or H3K4me3 peaks co-localized, and polycomb when EZH2 and-or H3K27me3 peaks co-localized. Mixed statuses (T-H, P/H, P/T, P/T/H) were proposed when appropriate (S1 and S2 Tables). 96.7% (1792/1855) of the Sil+, Sil- and U elements had at least one annotation about their chromatin status. Consequently, we considered that “ucs” annotations were not due to weak mapping ability of Hsmar1 and Hsmar2 elements but rather to variation of the quality of the CHIP-seq signal probably because sequencing depths were not deep enough [90,91] and-or variation of “sequencing ability” from one locus to another [92–94]. Therefore, statistical analyses were performed using datasets in which polycomb, trithorax and Su(var)39/HP1 frequencies were calculated and differences between MLEs Sil+/Sil- was tested using a Wilcoxon signed-rank test without “ucs” data (S2 Table).
Using HMMER, a HMM model for YY1 was calculated from the YY1 binding sites found in Mos1 and the consensus sequences of Hsmar1, Hsmar2, Himar1 and Mcmar2. YY1 sites were then detected in genomic Hsmar2 sequences using HMMER and score for positive hits were recorded. Textual searches were done to identify putative NFAT-5 binding site using the motif AAGGG/CCCTT.
The graphics calculated from data analysed with non-parametric statistics followed recommendations of the guidelines for journals of the American Society of Microbiology [95]. All values represented in graphics corresponded to the median value obtained from three experiments done in triplicate (9 data points). Bars corresponded to the values of quartiles 1 and 3. In the text, figures and supplementary data, all the results indicated as being different were previously verified to be significant with a p-value < 0.05, using a Kruskal-Wallis test. Wilcoxon signed-rank tests, Student t-tests and hierarchical clustering were done using R facilities and libraries [96].
|
10.1371/journal.pgen.1005467 | Identification of Driving ALK Fusion Genes and Genomic Landscape of Medullary Thyroid Cancer | The genetic landscape of medullary thyroid cancer (MTC) is not yet fully understood, although some oncogenic mutations have been identified. To explore genetic profiles of MTCs, formalin-fixed, paraffin-embedded tumor tissues from MTC patients were assayed on the Ion AmpliSeq Cancer Panel v2. Eighty-four sporadic MTC samples and 36 paired normal thyroid tissues were successfully sequenced. We discovered 101 hotspot mutations in 18 genes in the 84 MTC tissue samples. The most common mutation was in the ret proto-oncogene, which occurred in 47 cases followed by mutations in genes encoding Harvey rat sarcoma viral oncogene homolog (N = 14), serine/threonine kinase 11 (N = 11), v-kit Hardy-Zuckerman 4 feline sarcoma viral oncogene homolog (N = 6), mutL homolog 1 (N = 4), Kiesten rat sarcoma viral oncogene homolog (N = 3) and MET proto-oncogene (N = 3). We also evaluated anaplastic lymphoma kinase (ALK) rearrangement by immunohistochemistry and break-apart fluorescence in situ hybridization (FISH). Two of 98 screened cases were positive for ALK FISH. To identify the genomic breakpoint and 5’ fusion partner of ALK, customized targeted cancer panel sequencing was performed using DNA from tumor samples of the two patients. Glutamine:fructose-6-phosphate transaminase 1 (GFPT1)-ALK and echinoderm microtubule-associated protein-like 4 (EML4)-ALK fusions were identified. Additional PCR analysis, followed by Sanger sequencing, confirmed the GFPT1-ALK fusion, indicating that the fusion is a result of intra-chromosomal translocation or deletion. Notably, a metastatic MTC case harboring the EML4-ALK fusion showed a dramatic response to an ALK inhibitor, crizotinib. In conclusion, we found several genetic mutations in MTC and are the first to identify ALK fusions in MTC. Our results suggest that the EML4-ALK fusion in MTC may be a potential driver mutation and a valid target of ALK inhibitors. Furthermore, the GFPT1-ALK fusion may be a potential candidate for molecular target therapy.
| Little is known about the molecular biology of medullary thyroid cancer (MTC), which is a rare disease. Genomics are increasingly being used to improve our knowledge about disease biology and to identify therapeutic targets in many cancers. Here, we report the largest genomic results of MTC to date. MTC tissue frequently included several mutations. For the first time, anaplastic lymphoma kinase (ALK) rearrangements were detected in MTC: one case with a glutamine:fructose-6-phosphate transaminase 1 (GFPT1)-ALK fusion, and another case with an echinoderm microtubule-associated protein-like 4 (EML4)-ALK fusion. The fusion mechanism of the novel GFPT1-ALK fusion was successfully investigated using molecular biology techniques. In addition, an inhibitor of ALK (crizotinib) dramatically decreased the number of metastatic MTC lesions harboring the EML4-ALK fusion, thus verifying the fusion as a promising target in MTC. Our findings suggest that using rapidly improving sequencing techniques and accumulated genomic data to comprehensively perform genetic analyses on rare tumors, such as MTC, will help to improve the poor prognosis of orphan diseases.
| Many cancer gene profiling studies have recently been published, describing genetic trends that are not limited to specific cancers. Next-generation sequencing (NGS) is an important tool for detecting genetic alterations in many kinds of cancers, as it allows for millions of nucleic acid sequences to be simultaneously sequenced within a short period of time and is more cost-effective than older methods. Thus, many researchers and physicians anticipate that NGS will bring the concept of personalized cancer therapy to fruition.
Medullary thyroid cancer (MTC) is a rare malignancy that accounts for up to 3–5% of thyroid cancers. It is derived from calcitonin-secreting para-follicular C cells and can arise in a familial (25%) or sporadic (75%) pattern. Genetic and epigenetic alterations play important roles in the progression and prognosis of MTC [1–3]. Genes encoding the ret proto-oncogene (RET) and Ras (RAS) are commonly mutated in MTC [4, 5]. The RET mutation is believed to be a causative event in both familial and sporadic MTC [6, 7]. In the Mitogen-activated protein kinase (MAPK) pathway, the RAS mutation is another genetic rearrangement that is prevalent in sporadic MTC and other types of thyroid cancer [2] but the prevalence and significance of other genetic mutations including BRAF in MTC remain unclear.
MTC has a different response to treatment than that of well-differentiated thyroid cancers. Because radioactive iodine does not accumulate in MTC, few therapeutic options are available for advanced MTC. Inhibitors of RET, such as cabozantinib and vandetanib, have recently been shown to be effective in advanced MTC [8, 9]. However, whether the RET mutation is a predictive factor for the success of these drugs is unclear [9].
Recently, the rearrangement of anaplastic lymphoma kinase (ALK) was detected in a small but significant proportion of patients with non-small cell lung cancer (NSCLC) [10]. Several ALK inhibitors, including crizotinib, have achieved dramatic responses in cases of NSCLC harboring ALK rearrangements [11–13]. Although ALK rearrangement has also been episodically observed in a small set of other cancer types, little is known about ALK rearrangements in MTC [14, 15].
In this study, we used targeted NGS and various methods to examine the genetic profiles of MTC and detect ALK rearrangements.
Eighty-four samples (11 hereditary, 41 sporadic and 32 unknown) from patients with MTC (mean age of 48.5 years) and 36 paired normal thyroid tissue samples were successfully sequenced. The normal thyroid tissue samples in the MTC patients were used as matched control samples. Of the cases, 32 were male and 52 were female. Detailed demographic, clinic-pathological and genetic characteristics are listed in Table 1 and S2 Table. Hereditary MTCs were defined as cases having either positive germ-line RET mutations in blood tests or possession of a strong family history with MTC in at least four family members [16]. The unknown group was composed of MTC cases with no blood RET test and no family history of MTC/MEN. The mean value of variant coverage was 593 reads, and the variant coverage ranged from 19 to 1,482 reads. Overall, 101 mutations were observed in the MTC samples. Most mutations (N = 96, 95.0%) were single-nucleotide variants; 5 were deletions. The most common mutation occurred in RET, which was observed in 47 cases, followed by mutations in genes encoding Harvey rat sarcoma viral oncogene homolog (N = 14), serine/threonine kinase 11 (N = 11), v-kit Hardy-Zuckerman 4 feline sarcoma viral oncogene homolog (N = 6), mutL homolog 1 (N = 4), Kiesten rat sarcoma viral oncogene homolog (N = 3), MET proto-oncogene (N = 3), ATM serine/threonine kinase (N = 2), kinase insert domain receptor (N = 2), adenomatous polyposis coli (APC; N = 2), B-raf proto-oncogene (N = 1), cadherin 1 (N = 1), epidermal growth factor receptor (N = 1), cyclin-dependent kinase inhibitor 2A (CDKN2A, N = 1), Janus kinase 3 (N = 1), protein tyrosine phosphatase, non-receptor type 1 (N = 1), SMAD family member 4 (N = 1) and von Hippel-Lundau tumor suppressor (N = 1). We did not detect any dominant gene mutations in 20 MTC samples, which all exhibited wild-type RET, HRAS and KRAS. These are listed in S1 Table and shown in Fig 1.
The commonly observed RET mutations occurred in exons 10, 11, 15, and 16. Previous studies have shown that M918T is the most common RET mutation in MTC [2, 10]. Similarly, M918T (N = 19) was the most common RET mutation in our samples, followed by C634Y (N = 7), C634W (N = 4), C634G (N = 4), C630R (N = 4), D631Y (N = 2), and others (N = 7). All HRAS mutations occurred in exon 3. The mutant amino acid sequence in each of the HRAS mutant cases was Q61K (N = 13). KRAS mutations were observed in three cases (Q61R, 2 and G48R, 1), and BRAF mutation was found in only one case. The dominant amino acid sequence in STK11 was F354L (N = 7). Other mutated genes are shown in S1 Table.
We compared the genetic landscapes between 36 MTC tissues and their matched normal thyroid tissues: this group was composed of 16 sporadic, 5 hereditary and 15 cases with unknown information about heredity (Fig 2). In the hereditary MTC cases, RET mutations were observed in MTC and their matched normal thyroid tissues: these RET mutation types included C634Y, D631Y, and C634W, which are well known to be associated with the MEN2A [17, 18]. One case, which had been classified as an unknown subgroup based on blood test or family history, was found to have C634W mutation in both MTC and normal tissue, leading us to suspect that this case might be hereditary MTC. In 16 sporadic MTC group, several RET mutation types (M918T, C630R, C618S and deletion) were detected in MTC tissues, but not in the matched normal thyroid tissues. The M918T RET mutations and Q61K HRAS mutations were observed only in the MTCs of the sporadic or unknown subgroups, suggesting that these mutations are pathognomonic somatic mutation in MTC. In addition, two KRAS (G48R, Q61R) and one MET (A986T) mutations were also observed only in MTC tissues. However, STK11 (F354L), MLH1, KIT, and KDR mutations were observed in both MTC and normal thyroid tissues, which leads their pathognomonic natures unresolved in MTC.
In parallel with targeted sequencing using AmpliSeq, we screened for ALK rearrangements. Ninety-eight cases were screened using immunohistochemistry (IHC), and 83 of these cases were also evaluated using AmpliSeq. Nine ALK-positive cases were found with IHC scores of 1+ (N = 7), 2+ (N = 1), and 3+ (N = 1). We also performed ALK fluorescence in situ hybridization (FISH) testing on ALK-positive samples that were identified via IHC. The two samples with 2+ and 3+ IHC scores exhibited ALK break-apart rearrangements (Fig 3).
For the two cases harboring ALK break-apart rearrangements, targeted cancer panel sequencing (HiSeq 2500, Illumina, USA) was performed to detect the breakpoints and 5’ fusion partner genes of ALK. This process revealed two distinct ALK fusions. For the first case, a novel fusion gene was detected: 5’ glutamine:fructose-6-phosphate transaminase 1 (GFPT1; located in 2p13) was fused to 3’ ALK (located in 2p23) with preservation of the ALK kinase domain (Fig 4A and 4B). The breakpoints in GFPT1 and ALK were in intron 18 and exon 20, respectively. Based on the gene direction and location, the structural variation was presumed to be intra-chromosomal translocation or deletion. To confirm the fusion, we amplified the genomic fusion point between GFPT1 and ALK using genomic DNA of the MTC tissue. PCR analysis and Sanger sequencing revealed the same results as that of the customized targeted cancer panel (Figs 4B and S1). For the second case, the echinoderm microtubule-associated protein-like 4 (EML4)-ALK fusion was detected. The breakpoints were located in intron 13 of EML4 and intron 19 of ALK, which indicates that this fusion is the most common variant (E13; A20) in NSCLC [19, 20]. This case exhibited metastatic lesions after thyroidectomy and was enrolled in a Phase I crizotinib trial (NCT01121588). After crizotinib therapy, the tumor lesions in the lung, liver, and bone shrank remarkably, and plasma calcitonin levels decreased. The final results will be disclosed with the full clinical study.
We identified two types of ALK fusion genes in MTC by sequencing via IHC, FISH, and NGS analyses. Of the two fusion types, the EML4-ALK fusion was the same as the most commonly detected variant in NSCLC, [19] where the EML4-ALK fusion is a strong predictive factor for the efficacy of ALK inhibitors [13, 21, 22]. In the current study, the patient with metastatic MTC harboring the EML4-ALK fusion showed a dramatic response to crizotinib. We are the first to report an MTC case with a targetable EML4-ALK fusion gene. Previously, Kelly et al. used the Illumina HiSeq sequencing system to identify one papillary thyroid cancer case with an EML4-ALK fusion [15]. However, they also tested 22 medullary carcinoma cases and did not find any cases with the EML4-ALK fusion, as evaluated by reverse transcription-PCR. Their failure to detect the ALK rearrangement in MTC is understandable, given that our prevalence rate of ALK fusions in the current study was only 2% (2 out of 98 cases). This suggests that more efficient strategies are needed to detect the ALK rearrangement. Results from the current study suggest that IHC-based screening, along with FISH-based confirmation and targeted NGS, may be a cost-effective and reliable method to detect ALK rearrangements.
Most importantly, we detected a novel GFPT1-ALK fusion that has not been reported in any type of cancer. GFPT1 is a key enzyme in the biosynthesis of N-acetylglucosamine and is required for critical events in neuromuscular transmission [23]. Until now, several fusion partners of ALK have been reported in various cancers [24–28]. Among them, huntingdon-interacting protein (HIP1)-ALK and RAN-binding protein 2 (RANBP2)-ALK, which have been reported to exist in NSCLC and inflammatory myofibroblastic tumors, respectively, show clinical responses to crizotinib [25, 26]. In the current study, the MTC case harboring the GFPT1-ALK fusion showed strong ALK protein expression and did not exhibit co-existing genetic mutations; both of these factors may support an important role for this fusion gene in the pathogenesis of this MTC case. However, we were unable to validate whether GFPT1-ALK was a driving oncogene or a therapeutically targetable gene. Whether GFPT1-ALK is also a predictor for ALK inhibitors is unclear.
Currently, vandetanib and cabozantinib are approved for the treatment of MTC by the U.S. Food and Drug Administration. However, the prognosis of patients with metastatic MTC is still poor, due to the inherent resistance to radioiodine therapy and aggressive nature of this disease. Furthermore, the rarity of MTC makes it hard to perform prospective studies to find new agents. Therefore, the comprehensive genetic analysis of MTC can help to identify effective ways to improve its prognosis. Despite the low frequency of ALK rearrangements in MTC, our techniques can be used to detect target genes in other rare diseases.
In addition, our sequencing analysis of MTC is the largest to date. Previously, Agrawal et al. published the largest genomic analysis of MTC [5], where they performed whole-exome sequencing of 17 sporadic MTCs and 40 additional MTCs (hereditary or sporadic) for validation. RET was the dominant mutation (43/57) in that study. We used a larger sample size and accurate verification by comparing 36 pairs of MTC with matched normal thyroid tissues that were acquired from the same person.
In the comparison analyses, all five hereditary cases were observed to have germ-line RET mutations in both MTC and control tissues. However, M918T RET (N = 10), Q61K HRAS (N = 7), KRAS (N = 2), and MET (N = 1) mutations were harbored dominantly in MTCs. Simbolo et al. identified RET, HRAS, KRAS and STK11 mutations as significant somatic mutations in MTCs, whereas TP53, KDR, KIT, MET, PIK3CA and ATM mutations were classified as nonpathogenic germ-line variants [29]. Our current data are compatible with that report. Interestingly, however, the F354L STK11 mutation, regarded as significant somatic mutation by Simbolo et al., was observed in both MTCs and control tissues of our seven cases. Therefore, we presume that the F354L STK11 mutation is a germ-line mutation in MTC.
In conclusion, we report that the EML4-ALK fusion, which was found for the first time in MTC, could be an effective molecular target of crizotinib. Furthermore, our results also suggest that the novel GFPT1-ALK fusion can be a potential candidate for molecular target therapy. This study included the largest set of molecular profile data in MTC to date, which was achieved by using high-depth NGS panel sequencing, and also presented the genetic landscape of MTC. Further translational research is needed to determine the oncogenic roles of these mutations in MTC.
Written informed consent was obtained from all participants, and this study was approved by the Institutional Review Board of Samsung Medical Center. (SMC 2013-02-010).
We collected data on patients who were histologically diagnosed with MTC without the coexistence of tumors on the parathyroid and adrenal gland. All patients received surgical treatment at Samsung Medical Center between June 2000 and January 2013. Among 101 MTC specimens, 17 were excluded based on quality control (N = 5), preparation failure (N = 11), and sequencing failure (N = 1). The remaining 84 MTC samples were sequenced using an Ion Torrent Personal Genome Machine (IT-PGM, Life Technologies, Grand Island, NY, USA), which takes real-time measurements of hydrogen ions that are produced during DNA replication and allows for rapid sequencing. Eight normal thyroid tissues were obtained by thyroidectomy and sequenced. Mutation profiles between MTC and normal thyroid tissues from eight individuals were compared.
We constructed libraries using the Ion AmpliSeq Panels, Ion AmpliSeq Library Kit, and Ion Xpress Barcodes, as well as 10 ng of DNA sample per pool (Life Technologies). The amplicons were ligated to Ion Adapters and purified. For barcoded library preparations, barcoded adapters from the Ion Xpress Barcode Adapters 1–96 Kit were substituted for the non-barcoded adapter mix in the Ion AmpliSeq Library Kit. Next, the multiplexed barcoded libraries were enriched by clonal amplification using emulsion polymerase chain reaction (PCR) on Ion Sphere Particles (Ion PGM Template 200 Kit) and loaded on an Ion 316 Chip. Massively parallel sequencing was carried out on an Ion PGM using the Ion PGM Sequencing 200 Kit v2. The Ion AmpliSeq Cancer Hotspot Panel v2 covered hotspot regions of 50 oncogenes and tumor suppressor genes.
The primary filtering process was performed with the Torrent Suite v4.0.0 and Ion Torrent Variant Caller v4.0 software and included signal processing, base calling, assigning quality scores, adapter trimming, PCR duplicate removal, read alignment (to human genome reference 19), mapping quality control, coverage analyzing, and variant calling [30]. To detect variants, a minimum coverage of 100 reads was achieved with a cutoff value of at least 5% in the variant calling rate (frequency). Variant calls were further analyzed by using ANNOVAR variant filtering and COSMIC database (dbSNP build 137) annotating, and these analyses were based on changes in the amino acid sequence.
The ALK IHC assay used a mouse monoclonal ALK antibody (5A4, Novocastra, Newcastle, United Kingdom) and the antibody for ALK was diluted to 1:30, treated, and incubated at 42°C for 2 hours. ALK IHC scores were assigned as follows: 0, no staining; 1+, faint or weak staining intensity with more than 5% tumor cells or any staining intensity with ≤5% tumor cells; 2+, moderate cytoplasmic reactivity with more than 5% tumor cells; and 3+, granular cytoplasmic reactivity of strong intensity in more than 5% of tumor cells [31]. Cases that showed ALK-positive staining with a score of 1+ or greater were analyzed by FISH with the Vysis ALK Break-Apart FISH Probe Kit (Abbott Laboratories, Abbott Park, IL). Samples were considered positive for ALK FISH if more than 15% of cells were positive or an isolated red signal (IRS) in tumor cells.
Genomic DNA extraction was performed using the QIAamp DNA mini kit (Qiagen, Valencia, CA, USA), according to the manufacturer’s instructions. The Nanodrop 8000 UV-Vis spectrometer (Thermo Scientific Inc., DE, USA), Qubit 2.0 Fluorometer (Life Technologies), and 2200 TapeStation Instrument (Agilent Technologies, Santa Clara, CA, USA) were used to check the concentration, purity, and degradation of extracted genomic DNA. For the next step, samples that passed our quality control thresholds were used.
Genomic DNA (250 ng) from the tissues was sheared by the Covaris S220 (Covaris, Woburn, MA, USA) and used for the construction of the library using customized RNA baits and the SureSelect XT reagent kit, HSQ (Agilent Technologies), according to the manufacturer’s protocol. The customized RNA baits covered whole exons and flanking intronic sequences of the 83 genes. After enriched exome libraries were multiplexed, the libraries were sequenced on the HiSeq 2500 sequencing platform (Illumina, USA), as described previously [32]. Briefly, a paired-end DNA sequencing library was prepared through the following processes: genomic DNA shearing, end-repair, A-tailing, paired-end adaptor ligation, and amplification. After the library was hybridized with bait sequences for 16 hours, the captured library was purified and amplified with an index barcode tag. Then, the quality and quantity of the captured library were measured. Sequencing of the exome library was carried out using the 100-bp, paired-end mode of the TruSeq Rapid PE Cluster kit and TruSeq Rapid SBS kit (Illumina, San Diego, CA, USA).
The newly identified glutamine:fructose-6-phosphate transaminase 1 (GFPT1)-ALK fusion gene was detected by targeted cancer panel sequencing, and its respective genomic rearrangement was confirmed by genomic PCR analysis, followed by Sanger sequencing. Genomic DNA was isolated from formalin-fixed, paraffin-embedded (FFPE) tumor samples using a ReliaPrep FFPE genomic DNA extraction kit (Promega, Madison, WI, USA). The PCR products were indicative of fusion points within intron 18 of GFPT1 and exon 20 of ALK, based on target sequencing results. PCR analysis of genomic DNA for GFPT1-ALK was performed with a pair of primers flanking the putative fusion point: GFPT1 F (5’-TCTGTGTGAACTGGCACCTT-3’) and ALK R (5’-ATTCAGCCCCTACACTGCAC-3’). PCR products were then separated on a 2% E-Gel SizeSelect agarose gel (Invitrogen, Carlsbad, CA, USA). For genomic PCR controls, we used DNA from the same FFPE tumor samples with glyceraldehyde-3-phosphate dehydrogenase PCR primers. In reactions that produced a PCR product of the expected size, the amplicons underwent gel purification and sequencing using a 3130 XL ABI Prism sequencer (Applied Biosystems, Foster City, CA, USA) with Bigdye Terminator v3.1 Cycle sequencing kits, according to the manufacturer’s instructions.
|
10.1371/journal.pgen.1002598 | Broad-Specificity mRNA–rRNA Complementarity in Efficient Protein Translation | Studies of synthetic, well-defined biomolecular systems can elucidate inherent capabilities that may be difficult to uncover in a native biological context. Here, we used a minimal, reconstituted translation system from Escherichia coli to identify efficient ribosome binding sites (RBSs) in an unbiased, high-throughput manner. We applied ribosome display, a powerful in vitro selection method, to enrich only those mRNA sequences which could direct rapid protein translation. In addition to canonical Shine-Dalgarno (SD) motifs, we unexpectedly recovered highly efficient cytosine-rich (C-rich) sequences that exhibit unmistakable complementarity to the 16S rRNA of the small subunit of the ribosome, indicating that broad-specificity base-pairing may be an inherent, general mechanism for efficient translation. Furthermore, given the conservation of ribosomal structure and function across species, the broader relevance of C-rich RBS sequences identified through our in vitro evolution approach is supported by multiple, diverse examples in nature, including C-rich RBSs in several bacteriophage and plants, a poly-C consensus before the start codon in a lower eukaryote, and Kozak-like sequences in vertebrates.
| In order to maintain an appropriate balance of proteins in the cell, the protein factories (ribosomes) translate different messages (mRNAs) into protein at different rates. Many human diseases, including cancer and certain hereditary diseases, are caused by making too much or too little protein. Additionally, infections caused by bacteria and viruses are enabled by the ability of these organisms to produce protein very quickly while situated in their host. For these reasons, it is important to understand the ways in which ribosomes may recognize mRNAs and initiate translation into protein. We developed an experimental system that allowed us to uncover the inherent mRNA–binding ability of the ribosomes in a common bacterium, Escherichia coli. We found evidence that, when removed from the native cellular environment, these ribosomes are able to make protein very efficiently using previously unidentified ribosome binding sites on the mRNA that closely resemble known ribosome binding sites in diverse organisms, including plants and humans. Our results suggest a general, ubiquitous mechanism of mRNA–ribosome association during translation initiation.
| The ribosome is widely recognized as a broad-specificity ribozyme that is able to translate mRNA at different rates to maintain appropriate relative protein levels and thereby fulfill the dynamic needs of the cell [1]–[3]. Problems with increased or decreased translation of certain messages are known to lead to cancer and various other hereditary diseases in humans [4]. One of the major determinants of translational efficiency is the 5′ untranslated region (5′ UTR), which may contain a canonical RBS such as the Shine-Dalgarno (SD) sequence [5] in prokaryotes or the Kozak sequence [6] in vertebrates. Recently, it has been noted that, while the SD consensus sequence (5′-GGAGGU-3′) is generally an important cue for ribosome binding in prokaryotes, there are actually more non-SD-led genes than SD-led genes in some microbial genomes [7]. Additionally, the Kozak sequence is a relatively weak consensus, as only a very small fraction of vertebrate genes (∼0.2%) have the exact GCCGCC(A/G)CCAUGG sequence [8]. These observations do not immediately suggest a universal answer to the following fundamental question: what 5′ UTR sequences inherently enable a ribosome to bind mRNA, initiate translation, and proceed to elongation as quickly as possible?
Although efficient RBSs have been previously identified by library approaches both in vivo [9], [10] and in cell extracts in vitro [11], [12], the mechanisms of efficient translation are confounded by the multitude of uncharacterized biomolecular interactions in these environments. Furthermore, both the library size and the sequencing throughput in earlier studies have been limited, hindering identification of statistically significant motifs. To more directly answer the question posed above, we performed selections on a large RBS library (∼3.7×1013 mRNA molecules; ∼6.9×1010 unique sequences) in a minimal, well-defined, E. coli-based translation system [13]–[15] using ribosome display [16]. By using a minimal translation system, we removed unnecessary confounding variables and took a “bottom-up” approach to address the question of what sequences inherently promote the fastest translation.
One of the major goals of synthetic biology is to reveal new fundamental biological insights through the use of well-defined systems. The present study complements previous advances in the field that utilized or focused on differential RBS function, including work on riboregulators [17]–[19] and the RBS Calculator [20], as well as early work on synthetic gene networks that used RBSs of various strengths to adjust the gene expression dynamics of synthetic constructs [21]. Here, we were able to attribute the selected RBSs directly to the contents of the translation system because of its fully defined nature; additionally, we were able to consider general aspects of RBSs, which are not necessarily E. coli-specific, as the basic translational machinery is highly conserved across species.
High-throughput sequencing of the library after stringent selection for translational efficiency surprisingly revealed mostly non-SD motifs. These library members, some of which were nearly as efficient as the SD-containing 5′ UTR sequence derived from enterobacteriophage T7 gene 10, were generally highly C-rich. While it is well appreciated that SD sequences help to form the preinitiation complex by binding to the anti-SD sequence in the unpaired 3′ end of the 16S rRNA in the 30S ribosomal subunit, we further hypothesized that our efficient non-SD RBSs also achieve fast translation by optimizing binding to the 16S rRNA. (“Fast translation” in our study should be considered rapid in the context of the minimal system; the potential speed of translation may be much higher in vivo.)
Based on statistical analyses and competition studies, we conclude that base-pairing between the short, C-rich motifs of the non-SD RBSs and the G-rich rRNA of the small ribosomal subunit allows for fast translation, most likely through fast repositioning of the mRNA on the small ribosomal subunit to form a productive preinitiation complex that is then able to join the large ribosomal subunit and proceed quickly to elongation. We have demonstrated that pure poly-cytosine (poly-C) is a poor RBS, and we have used rational mutagenesis to show that the specific positioning of non-C nucleotides in a C-rich context is a critical determinant of translational efficiency. We also show that the activity of C-rich RBSs, but not SD RBSs, can be strongly decreased in vitro by the addition of random oligonucleotide competitor sequences, which can explain their differential activities in vivo. Furthermore, we report similarities between the most common motifs in our selected RBSs and those in human RBSs, suggesting that structurally and functionally conserved ribosomes from diverse organisms are inherently capable of utilizing C-rich sequences directly upstream of AUG start sites. The broader relevance of C-rich RBSs is further supported by several other examples in nature, including C-rich RBSs in non-E. coli bacteriophage, C-rich RBSs that base-pair to a G-rich rRNA element in plants [22], [23], and a poly-C consensus before the start codon in a lower eukaryote [24]. The overall goal of this study was to determine inherent requirements for fast translation, and our experimental and computational results together provide evidence of a general, broad-specificity mechanism for efficient protein synthesis.
To investigate what upstream sequences promote fast translation, we chose a minimal, reconstituted, E. coli-based in vitro translation system: PURExpress (New England Biolabs) developed from PURE technology [13], [25], [26]. Ribosome display has previously been used to evolve peptides and proteins with desirable properties, including enhanced affinity and stability [16], [27]–[29]. Briefly, the standard method involves multiple cycles of generating a DNA library, in vitro transcription, in vitro translation, selection through binding, and recovery. The mRNA contains, at minimum, an RBS followed by a region encoding the gene of interest and an unstructured protein spacer with no stop codon, so that the ribosome stalls at the end of the mRNA, forming an mRNA-ribosome-polypeptide complex (hereafter called a ribosomal complex). In our adaptation (Figure 1A), we used a randomized 5′ UTR (Figure 1B) and progressively shortened the translation time in each round to impart an increasing selection pressure.
The 5′ UTR from the ribosome display vector pRDV [30] was considered the wild-type (WT) sequence. It includes a 5′ stem-loop to prevent degradation and a translational enhancer and SD RBS derived from enterobacteriophage T7. In the library version, the 18 nucleotides just prior to the start codon (5′-TAAGAAGGAGATATATCC-3′ in WT; SD sequence underlined) were fully randomized, creating a theoretical diversity of 418 = 6.9×1010 different sequences, which can be nearly exhaustively sampled in our in vitro system. The SD sequence, when present, generally has a context-dependent optimal position within this region [31]. Additionally, previous studies investigating the position of mRNA on the 30S ribosomal subunit have suggested that approximately 15 bases prior to the start codon are protected by the ribosome during initiation [32], making this a region of particular interest. The invariant coding region was chosen to be a fusion protein containing (from N- to C-terminus) an initiating Met, Ala, FLAG-tag, Gly-Ser (BamHI site), off7 [30], Lys-Leu (HindIII site), and a modified version of the pRDV tolA spacer that contains out-of-frame stop codons. Off7 is a designed ankyrin repeat protein (DARPin) that was evolved to bind maltose-binding protein of E. coli with nanomolar affinity (∼4.4 nM) [30]. We chose this model protein because it translates and folds well in vitro. Additionally, its high affinity for maltose-binding protein enables easy affinity purification of only those ribosomal complexes with fully translated protein.
We performed three rounds of selection (30 min, 5 min, and 3 min translation at 37°C; the “30-5-3 selection”) and, despite increasingly stringent translation times, the number of recovered mRNA molecules climbed from ∼4.4×109 in the first round to ∼1.5×1010 in the second round to ∼2.2×1010 in the third round. Quantitative reverse transcription-PCR (qRT-PCR) data and accompanying experimental details are presented in Figure S1. mRNA recovery from the third round was comparable to that produced from the WT pRDV RBS, which is highly efficient both in vitro and in vivo. This third round pool was subjected to in-depth analysis.
We sequenced the enriched pools from each round in the 30-5-3 selection using the Roche 454 platform. Approximately 7,000 raw sequences were obtained from each round: 7,268 from round 1; 6,825 from round 2; and 7,525 from round 3. Sequences were excluded from analysis if they did not have 18 bases in the randomized region, if they included an in-frame AUG within the randomized region that could serve as an alternate start site, or if there were errors in the 10 bases on either side of the randomized region. Approximately 5,000 sequences were analyzed from each round: 5,202 (4,933 unique) from round 1; 4,880 (4,586 unique) from round 2; and 4,863 (4,551 unique) from round 3. SD sequences were broadly defined as any sequence containing one of the following four-base motifs which could base-pair to the 3′ tail of the 16S: AAGG, AGGA, GGAG, GAGG, and AGGU. The overall incidence of SD motifs in each round is shown in Figure 1C. The positional and overall frequencies of each individual SD motif at the end of each round are presented in Figure S2. In our data, the first G of GGAG is enriched most prevalently around position −12, while the same nucleotide is favored around position −10 in E. coli [31]. Certainly, mRNA context may affect the optimal position of SD motifs, as may different in vitro or in vivo conditions. Position-dependent enrichment of SD motifs validated our selection method.
Remarkably, of the sequences analyzed from the third round, 3,696 (76%) were considered non-SD candidates (3,491 unique). While we expected that perhaps a significant portion of these non-SD candidates could still be acting by slightly mismatched SD-anti-SD interactions, this did not appear to be the case. In fact, we observed that these sequences were highly C-rich. Of the non-SD candidates, 2,244 (61%) contained nine or more cytosines out of 18. This cytosine richness did not appear to be position-dependent. Base frequency versus position and cytosine content histograms are shown in Figure 2A and 2B, respectively, for non-SD, SD, and combined populations from the third round of selection.
We hypothesized that these C-rich sequences might be operating by base-pairing with the 16S rRNA in the 30S ribosomal subunit, which is generally G-rich. Indeed, this idea has been suggested in both prokaryotic [33] and eukaryotic [34] systems, although consensus on the issue is lacking [35], [36]. We looked at four-, five-, six-, seven-, and eight-base potential complementarities. Overlapping windows of these lengths from the 18-base randomized region of third-round products were compared to all identically-sized windows of E. coli 16S rRNA. We considered all 4,863 18-base regions in this analysis, including both SD and non-SD sequences. The frequency of motifs in our data set that were Watson-Crick (A/U or C/G) reverse complements of each window on the 16S rRNA was determined. We assigned a p-value to each window on the 16S rRNA based on the probability distribution obtained from analyzing ∼100,000 randomly generated libraries equal in size to the dataset (probability of each base = 0.25). The 30S ribosomal subunit of E. coli (PDB 3DF1; [37]) is shown in Figure 3 with potential mRNA-rRNA base-pairing sites shown in red. To be highly stringent, only significant (p<0.01; Bonferroni-corrected) seven-base windows that shared six bases with at least one neighboring significant window were highlighted. Potential mRNA-rRNA base-pairing sites primarily fell on the body of the 30S subunit on the face that becomes buried after assembly with the 50S (Figure 3, first panel). The mRNA tunnel lies between the body and head on this face. Full results from the 16S rRNA comparison are presented in Table S1. We also found that the overall propensity of the enriched library to form secondary structure resembled that of the starting library (Figure S3), underscoring the importance of primary structure (i.e., nucleotide sequence) in ribosome binding. The lack of a strong pressure for low secondary structure in the RBS region may have resulted from compensatory low secondary structure in the first ∼40 nucleotides of the coding region.
Based on the observed C-rich trend and the complementarity to the G-rich 16S rRNA, we decided to perform a naïve motif search to reveal any interesting local patterns. We determined the frequency of all possible four-, five-, six-, seven-, and eight-base motifs within the 18 bases, independent of the 16S rRNA, and asked whether specific motifs were significantly overrepresented compared to what would be expected in the naïve library (i.e., N18). We considered all 4,863 18-base regions from the third-round products in this analysis, including both SD and non-SD sequences. As expected based on overall base frequencies, nearly all of the top sequences were highly C-rich. More striking was that the most frequent motifs from the motif search exhibited unexpected similarities to the Kozak consensus sequence found in vertebrates. To investigate these observed similarities in more detail, the most frequent motifs found in the 18 nucleotides prior to the start codon in human (NCBI TaxID 9606) from the Transterm database [38] were considered. Four of the top nine five-base motifs in our selected sequences were also present within the top 17 motifs in human: CCACC, CCGCC, CCCGC, and GCCCC (Table 1). The full results from this motif search are provided in Table S2.
Previous studies involving prokaryotic RBSs have not recognized the inherent ability of 70S ribosomes to efficiently translate from C-rich start sequences, including those resembling the Kozak consensus sequence, probably because those studies were not conducted in a minimal translation system. The Kozak sequence has been previously investigated for its complementarity to the rRNA of the small subunit in eukaryotes [39], much as we have done with our selected RBS sequences. The Discussion provides further insight into the parallels between our study and this previous analysis performed in a eukaryotic system, suggesting universal features of the ribosome.
All motifs found to be significant in the motif search (FDR<0.01) were given further consideration for their co-occurrence with other significant motifs within the same 18-base randomized RBS region. A co-occurrence metric was defined as the number of RBS regions that contained both motif 1 and motif 2 divided by the number of RBS regions that contained motif 2 only. Through this measure, we identified “enhancers” of canonical SD motifs. Variations of an AC dinucleotide repeat were found to correlate strongly with GGAGG. Interestingly, AC dinucleotide repeats downstream of the start codon have previously been reported to enhance translation [40]. Results from the co-occurrence analysis are provided in Table S3 for all pairs of significant motifs that had a non-zero co-occurrence metric. Co-occurrence of C-rich motifs with other C-rich motifs is also evident in Table S3.
We tested the poly-C consensus RBS against the WT pRDV RBS and one of our C-rich RBS clones in single-clone ribosome display. mRNA recovery was quantified by qRT-PCR (Figure 4A, top three sequences). Surprisingly, the poly-C consensus was not efficient. To determine which non-C nucleotides in a C-rich context enabled efficient translation, we performed single-clone ribosome display on a panel of our most C-rich clones (with cytosines at 15 of 18 positions). We considered clones from the basic selection scheme (three rounds: 30 min, 5 min, 3 min translation; “30-5-3”) as well as two alternate selection schemes (four rounds: 30 min, 30 min, 1 min, 1 min translation with or without an additional 1-min round; “30-30-1-1-1” and “30-30-1-1,” respectively). mRNA recovery from the alternate selection schemes, quantified by qRT-PCR, is presented in Figure S1. Most clones exhibited activity well above background (Figure 4A); however, highly similar clones exhibited greatly different activities, suggesting that the placement of non-C nucleotides in a C-rich context is crucial. We investigated two clones, 30-30-1-1 high C 1 (GCCCCCCCCGCCCCCUCC; ∼80% WT activity) and 30-5-3 high C 7 (CCGCCCCCCCGCCCCUCC; ∼10% WT activity) more closely. These two clones differ only in the position of two guanines: one near the 5′ end of the random region and one near the middle. To investigate the nucleotides responsible for the differential activity of these two clones, we performed single-clone ribosome display on an extended panel of mutant RBSs (Figure 4B). Mutation of the first G to A, C, or U in 30-30-1-1 high C 1 had no major effect, while mutation of the second G to A, C, or U greatly decreased activity. Mutation of the U to A, C, or G also decreased activity. Finally, shifting the first G from −18 to −17 or −16 or shifting the second G from −9 to −8 greatly decreased activity.
To investigate our base-pairing hypothesis experimentally, we performed single-clone ribosome display of WT and a C-rich clone (30-30-1-1 high C 1) in the presence of various ssDNA oligonucleotide competitors. We used five different 18-base competitors: random (N), clone 30-30-1-1 high C 1, a similar C-rich clone (30-5-3 high C 7), WT, and poly-C. This panel of competitors was designed to interrogate specificity of translational inhibition (if any). The activity of the WT clone was only moderately inhibited by a large excess of any oligonucleotide, while the activity of the C-rich clone was effectively eliminated by random or C-rich competitors. Even WT competitor strongly inhibited the C-rich clone, though to a lesser extent than the other competitors (Figure 5A).
Finally, we tested a panel of clones in vivo by fusing off7 to emGFP through a short linker (Figure 5B) and then monitoring green fluorescence in E. coli (Figure 5C). This panel of clones included five C-rich pre-AUG 18-base regions from E. coli (derived from the 5′ UTRs of thiI, bisC, gsk, nrdB, and uxuR), 15 clones from the 30-5-3 selection with maximal redundancy (two with four instances, 13 with three instances), three representative clones with high C content from the 30-5-3 selection, three of the most C-rich 18-base upstream regions present in phage annotated on EMBL-EBI, and the WT pRDV sequence. The average median fluorescence of these 31 clones from at least three independent experiments is provided in Figure 5C. The induced WT signal was over 580 times above that of 30-5-3 high C 7, while 5′ UTR mRNA levels were only about 14-fold different, which only accounts for a small fraction of the discrepancy in protein levels. This suggests that observed differences in the in vivo responses for WT and the C-rich clones can be primarily attributed to their differential translational efficiencies. The poor performance of C-rich upstream regions from phage was not unexpected, because the phage from which those 5′ UTRs were derived do not naturally infect E. coli. In support of a base-pairing mechanism, native hosts of phage having C-rich 5′ UTRs (e.g., Burkholderia cenocepacia, Mycobacterium tuberculosis H37Rv, and Synechococcus sp. WH 8109) clearly have more C-rich 5′ UTR profiles than E. coli (Figure S4). Although most of our selected clones performed poorly in vivo, at least two synthetic sequences (30-5-3 clones 11 and 12) exhibited activity >2-fold over background, on par with that of the native 18-base sequence immediately upstream of E. coli gsk. In light of our competition experiments in vitro, we conclude that the in vivo environment of E. coli contains a large quantity of endogenous RNA species that out-competes mRNA containing a C-rich RBS. However, given the two examples of synthetic sequences that retain some activity in vivo, the magnitude of this competition effect is likely to be sequence-specific.
Ribosome display, employed as a tool for investigating the non-coding regions of mRNA, particularly in a minimal translation system, has the potential to generate insights not available through previous studies. The large library sizes of ribosome display (easily up to ∼1014 with reasonable scale-up) allow much more exhaustive sampling than any technique requiring a transformation step. Coupling these selections with high-throughput sequencing enables the discovery of statistically relevant motifs in the selected sequences. Furthermore, a synthetic biology approach, in which a well-defined translation system is used, can elucidate inherent capabilities of the translational machinery and new insights into the function of natural biomolecules that may be difficult to uncover in a native biological context. In the present study, ribosome display and high-throughput sequencing were used to demonstrate that efficient translation in a minimal, well-defined, E. coli-based in vitro translation system can be mediated by C-rich RBSs which are postulated to base-pair to G-rich 16S rRNA motifs.
The identification of highly C-rich RBSs using ribosome display in the PURExpress system underscores the high structural and functional conservation of the ribosome and shows that, if given optimal conditions, ribosomes from one species can bind to mRNAs which are more frequent in other species in nature. Highly C-rich RBSs have been found in multiple diverse organisms, including non-E. coli phage, lower eukaryotes, plants, and vertebrates. A discussion of such natural examples as well as the notable lack of C-rich RBSs in E. coli genes is presented further below.
Interestingly, our selected sequences had an overall consensus of poly-C, although the poly-C sequence by itself was not efficient. The inability of this global consensus sequence to promote efficient translation in the PURExpress system provided an important insight for this study: the overall 18-base consensus does not describe the selected library well. Instead, shorter, significant (FDR<0.01) motifs that were analyzed independently of the 16S rRNA comprise many local consensus sequences. There was no striking position-dependence of individual local consensus sequences when viewed over the entire population; this contrasted starkly with the SD motifs, which were much more position-dependent.
Additionally, our consensus did not contain a “purine peak” at position -3, which is frequently found in humans and other vertebrates [6]. This purine peak may not be present in lower eukaryotes such as Encephalitozoon cuniculi, an intracellular eukaryotic parasite that frequently infects immunodeficient patients. This organism has short leaders but also contains a poly-C consensus prior to the start codon [24], much as we observed in our selections. The mechanism by which this parasite initiates translation is currently unknown, although the present study may provide some insight by demonstrating non-native functions of E. coli ribosomes that reflect the RBS preferences of other organisms.
The presence of C-rich sequences in phage 5′ UTRs suggests that some aspect of the host environment enables their fast translation. Based on our observations of the effect of competitor oligonucleotides, we propose that phage with C-rich 5′ UTRs best utilize these genes in an environment low in nucleic acids. Interestingly, the Burkholderia phage KS14 contains its most C-rich 5′ UTR prior to its gene for tail completion protein R. Therefore, at least one of the most C-rich motifs in phage precedes a highly-produced late protein (i.e., structural protein), although the general lack of annotation of phage genes limits our analysis. In late-stage infection, host mRNAs are often repressed, globally or locally [41]–[43], so highly efficient C-rich RBSs may also serve to temporally control the production of certain proteins (e.g., structural proteins should be abundantly synthesized, but only towards the end of phage assembly). Phage with C-rich 5′ UTRs may infect slow-growing organisms, such as M. tuberculosis [44], which may have lower basal mRNA content than other species, such as E. coli.
The co-occurrence of multiple short C-rich motifs within the 18-base RBS region suggests that multiple segments of the RBS may interact either sequentially or concurrently with the 16S rRNA, which has multiple binding sites itself. Fast binding and unbinding of these short mRNA motifs to various positions on the ribosome may help maintain a high concentration of ribosomes near the start codon while still permitting necessary mRNA repositioning for initiation and transition to elongation. The concept of multiple mRNA-rRNA interactions has been described as clustering for eukaryotic ribosomes [45], and we suggest that a similar mechanism may be at work here. In theory, the entire length of an mRNA molecule may be able to interact with the rRNA, but it is the initiation region that determines the accessibility of the start codon and the efficiency of forming the preinitiation complex [46].
mRNA-rRNA complementarity has also been found to enhance translation in plants. For example, the ARC-1 element (18S rRNA positions 1115–1124, GGGGGAGUAU) was shown to enhance translation when present in the leader or intercistronic region of model mRNAs [22]. This study also showed that linking three or more copies of this enhancer element augmented translation to levels directed by natural enhancers in tobacco mosaic virus and potato virus Y mRNAs. A subsequent investigation by the same group showed that enhancer activity was inhibited in the presence of competitor oligonucleotide and that the same oligonucleotide, when modified at the 5′ end with an alkylating group, hybridized to the ARC-1 element [23]. Intriguingly, part of the homologous E. coli 16S rRNA region was found to be a potential mRNA hybridization site in our study.
While it has been recognized for some time that the ribosome is, in fact, a broad-specificity ribozyme, there has not been much discussion of universally efficient RBSs in the literature. Recently, species-independent translational sequences have been reported [47]. These utilize a poly-A or UUUUA repeat to create a long, unstructured region prior to the start codon. The impressive efficiency of poly-A and (to a lesser extent) poly-U RBS constructs in vitro and in vivo is consistent with this report (Figure S5). An analysis of all eukaryotic start sequences has identified two distinct patterns, AAAAAA and GCCGCC, which supposedly work by distinct mechanisms [48]. S. cerevisiae, for example, prefers the former consensus, while human and other vertebrates generally use a sequence closer to the latter. Interestingly, the S. cerevisiae rRNA is rich in poly-U tracts, while vertebrate rRNAs are generally rich in poly-G tracts, further supporting the notion that transient rRNA-mRNA base-pairing may be a broad-specificity mechanism for translational regulation. Additionally, the base-pairing of Kozak sequences to the 18S rRNA has been proposed [39]. In this study, Sarge and Maxwell presented a competitive-displacement model for the initiation of translation involving the intermolecular base-pairing of 5S rRNA, 18S rRNA, and mRNA. They proposed that a particular segment of the 18S rRNA complementary to the Kozak sequence was able to lock the mRNA in place so that a 48S preinitiation complex could form. The 60S subunit would then join, and the 5S rRNA would displace the mRNA. Although the details of this model may not apply directly to the present study, there is indeed precedence in the literature for C-rich, Kozak-like sequences to show evidence of binding to the rRNA of the small subunit prior to initiation of translation [39]. More generally, the fact that ribosomes from distantly related organisms (i.e., E. coli and human) can use both poly-A and Kozak-like patterns to initiate translation provides interesting material for further research on the universality of the ribosome.
Because E. coli grows quickly and has large amounts of RNA compared to slower-growing bacteria, it is quite possible that competition for potential pairing sites on the ribosome from other nucleic acids or other molecules prevents translation of mRNAs containing C-rich RBSs. We make this assertion based on the fact that C-rich sequences are inhibited from facilitating translation in vitro when competitor oligonucleotides are added. Most E. coli genes are not C-rich, which highlights the fact that our results using E. coli ribosomes must be considered in the context in which they were selected. Our objective was to gain insight into the inherent capabilities of the ribosome, so we used a minimal in vitro translation system; by contrast, if the ultimate goal of a study is to simply increase in vivo expression, the selections should be performed in vivo. It is theoretically possible that C-rich mRNA sequences may have been selected in part because of their ability to outcompete other sequences for binding to ribosomes, not necessarily because they are the most efficient at promoting fast translation, which requires speed in forming the initiation complex and also in transitioning to elongation. However, the enriched libraries performed translation very well overall, suggesting that this should not be a major concern.
The computational analysis was performed without knowledge-based bias of where base-pairing occurs in available ribosomal crystal structures. Many of the potential pairing sites are at least partially base-paired in the crystal structure, but a large number of these sites may be vulnerable to displacement at the translation temperature. The ribosome is a highly dynamic macromolecule and surface-proximal potential pairing sites could easily be involved in transient complementary interactions.
Additionally, it is possible that the 23S and/or 5S rRNAs of the large ribosomal subunit may be involved in some of the interactions. The ribosomes in the PURExpress system are 70S complexes, although IF3 is able to separate them [49]. When an analysis identical to that shown in Figure 3 was performed with the 23S rRNA and 5S rRNA, we found 56 and 2 potential pairing sites, respectively. Based on what is known about the translation of leadered mRNAs, we would expect the 16S rRNA to play the major role; however, we cannot exclude the possibility of the large subunit rRNAs mediating mRNA-ribosome interactions, which, for example, could serve to increase the local mRNA concentration until a binding event resulting in translation initiation occurred.
Finally, based on the traditional model of prokaryotic translation, we assume that the 18-base randomized region before the coding region functions primarily in translation initiation, although it is possible that this region could exert some effects on elongation, perhaps if the C-rich sequences could interact with the ribosome in or near the exit tunnel to facilitate mRNA movement through the 70S ribosome. Differences in mRNA recovery could theoretically result from effects of the randomized RBS region on elongation, but current dogma suggests that this is less likely.
In the present study, we uncovered both expected SD sequences and unexpected C-rich non-SD sequences as efficient RBSs in a minimal, reconstituted E. coli system. All of these sequences appear to operate by base-pairing to the rRNA of the small subunit of the ribosome. This general design principle represents an inherent, broad-specificity mechanism for efficient translation in vitro that is further refined in vivo (Figure 6). Notably, the specific subset of RBSs that are utilized in vivo can be different for different hosts: E. coli does not appear to utilize C-rich RBSs in translating its native genes, likely due to the fact that SD sequences perform more robustly in its intracellular environment; bacteria such as Mycobacterium tuberculosis have more C-rich 5′ UTRs than E. coli, suggesting that both SD and C-rich RBSs play functional roles in these hosts; and human and other vertebrates widely use C-rich sequences (including Kozak-like motifs), but not SD-like sequences, for translation. Our results suggest the intriguing possibility that RBSs in different organisms that may appear unrelated by sequence may actually share a common mechanism for translation initiation based on broad-specificity mRNA-rRNA base-pairing.
Procedures for construction of the naïve RBS library, the single-clone constructs used for single-clone ribosome display, and the single-clone constructs used for the in vivo expression studies are provided in Text S1. All oligonucleotides specific to these procedures are listed in Table S4.
Ribosome display selection particles were generated using the well-defined PURExpress in vitro protein synthesis kit (New England Biolabs). Since the concentration of ribosomes in the standard PURExpress reaction is specified by the manufacturer (2.4 µM), we could accurately control the RNA∶ribosome ratio (∼10∶1 in the first round, ∼4∶1 in subsequent rounds) by using RNA, and not DNA, as the template. Kit components (Solution A and Solution B), RNA, RNasin ribonuclease inhibitor (Promega, Madison, WI) and water (if necessary for dilution) were mixed according to the manufacturer's instructions, except in cases where fewer ribosomes (found in Solution B) were required to achieve high RNA∶ribosome ratios. In the first round of selection, 18 µg mRNA (corresponding to ∼3.7×1013 molecules) was used in a total volume of 16 µL. The translation reaction was incubated at 37°C for 30 min in order to allow full translation of any mRNAs that contained an RBS. The translation was stopped using 400 µL cold WB buffer (50 mM Tris-acetate, pH 7.5 at 4°C, 150 mM NaCl, 50 mM magnesium acetate; [28]). Then, the stopped translation was subjected to ultrafiltration using a 100 kDa cut-off Microcon centrifugal filter unit (Millipore, Billerica, MA). The ultrafiltered translation was diluted up to 100 µL with WBT (WB plus 0.05% Tween-20) containing RNasin, mixed thoroughly, and used for binding in one well. Binding was performed using NUNC Maxisorp plates (Thermo Fisher Scientific, Rochester, NY) prepared as follows: plates were coated with 100 µL 66 nM NeutrAvidin (Thermo Fisher Scientific) for at least 16 h at 4°C, washed with TBS (50 mM Tris-HCl, pH 7.4 at 4°C, 150 mM NaCl), blocked with 25 mg/mL casein (Sigma-Aldrich, St. Louis, MO) or 10 mg/mL BlockAce (AbD Serotec, Raleigh, NC) in TBS at room temperature for at least 1 h with shaking, incubated with biotinylated maltose-binding protein of E. coli in blocking solution for at least 1 h at 4°C with shaking, and washed with TBS and WBT. Binding was performed for 1 h at 4°C with shaking. The plate was washed with WBT and then once with WB prior to reverse transcription.
Reverse transcription was performed using AffinityScript reverse transcriptase (Agilent Technologies, Santa Clara, CA) and reverse primer tolA_stops_HindIII_rev (5′-GGC CAC CAG ATC CAA GCT T-3′) that anneals just downstream of off7. An in situ reverse transcription protocol [50] was adapted as follows: 12 µL Solution 1 (11.375 µL water and 0.125 µL reverse primer tolA_stops_HindIII_rev) was pipetted into the well, incubated at 70°C for 10 min, and removed from heat for 5 min. 8 µL Solution 2 (3 µL dNTPs [5 mM each], 2 µL 10× AffinityScript buffer, 2 µL 0.1 M DTT, and 1 µL AffinityScript reverse transcriptase) was added and the reaction was incubated at 45°C for 1 h, then heat-inactivated at 70°C for 15 min. Half of the 20 µL reaction was taken as template for a 100 µL PCR with primers T7_ext_fwd (5′-ATA CGA AAT TAA TAC GAC TCA CTA TAG GGA CAC CAC AAC GGT TTC CCT AAT TGT GAG CGG ATA ACA ATA GAA ATA ATT TTG TTT AAC TT-3′) and tolA_stops_HindIII_rev. T7_ext_fwd anneals just before the 18-base randomized region to maximize recovery; additionally, by only recovering those sequences which contain enough bases upstream of the RBS region to facilitate primer annealing, we can be assured that potential nuclease processing near or within the RBS is not significantly influencing our results. The PCR product (624 bp) was gel-purified and digested with HindIII. The tolA spacer was made by amplifying pRDVstops-off7 with HindIII_tolA_stops_fwd (5′-TAC TGC AAC AAG CTT GGA TCT GGT GGC CAG AA-3′) and tolAk (5′-CCG CAC ACC AGT AAG GTG TGC GGT TTC AGT TGC CGC TTT CTT TCT-3′) [30] to form a 303 bp product. Both pieces were digested with HindIII, ligated, and gel-purified to generate the full-length construct (899 bp). This product was amplified with T7_no_BsaI (5′-ATA CGA AAT TAA TAC GAC TCA CTA TAG GGA CAC CAC AAC GG-3′) and tolAk to obtain enough product for transcription for the second round.
Different selection schemes were performed based on this first round with 30 min translation. In one scheme, two additional rounds (5 min and 3 min, respectively) were performed with no ultrafiltration (“30-5-3” selection). In an alternate scheme, three additional rounds (30 min, 1 min, and 1 min) were performed with ultrafiltration (“30-30-1-1” selection) followed by a final 1-min round without ultrafiltration (“30-30-1-1-1” selection). The volume in round 1 (16 µL) was chosen to be higher than in subsequent rounds because we expected few mRNAs in the original library to contain a functional RBS. After the initial round, the pool was highly enriched, so much smaller volumes could be used effectively. Pipetting errors were kept to a minimum by preparing translation reactions of at least 5 µL. After translation, the reactions were diluted, divided into four parts (each containing at least 1.25 µL translation), and used for binding in duplicate positive wells and duplicate negative wells. Thin-walled PCR tubes were used for incubation, so all volumes quickly reached the translation temperature (37°C). The products of all rounds were quantified by qRT-PCR on the Applied Biosystems 7300 Real-Time PCR System using TaqMan Universal PCR Master Mix (Applied Biosystems), off7_fwd (5′-TCC ATC GAC AAC GGT AAC GA-3′), tolA_stops_HindIII_rev, and off7_probe (6-FAM-5′-TGG CTG AAA TCC TG-3′). Products from all selection schemes were sequenced on a Roche/454 GS FLX sequencer at the University of Pennsylvania DNA Sequencing Facility. Sanger sequencing was also performed on the 30-30-1-1 selection. Sequences from the 30-5-3 selection were chosen for extensive sequence analysis. Highly C-rich clones from the 30-30-1-1 and 30-30-1-1-1 selections were also investigated. Prior to some rounds (5 min and 3 min rounds from 30-5-3 selection and final 1 min round from 30-30-1-1-1 selection), off7-tolA amplified with BsaI_FLAG_fwd2 (5′-ACT GAT TAG GTC TCA GAT GAC GAT GAC AAA GGA TC-3′) and tolAk was digested with BsaI and ligated onto the BsaI-digested library, made by PCR on the reverse transcription product using BsaI_FLAG_rev (5′-ACT GAT TAG GTC TCT CAT CTT TGT AGT CCG CCA T-3′) and T7_no_BsaI.
Sequence-verified minipreps were amplified with T7_no_BsaI and tolAk for in vitro transcription. Generally, ∼1 µL translation was used per well to make sure that the signal stayed in the linear range. The RNA∶ribosome ratio was 4∶1 in all experiments. Translation was performed for 10 min, which is optimal for WT. If applicable, DNA oligonucleotide at a concentration of 2.5 mM was added to the translation to a final concentration of ∼400 µM, which provided ∼40-fold molar excess compared to mRNA (∼9.6 µM). Five different DNA oligonucleotides were used: 18b_N, 5′-NNN NNN NNN NNN NNN NNN-3′; 18b_(30-30-1-1_high_C_clone_1), 5′-GCC CCC CCC GCC CCC TCC-3′; 18b_(30-5-3_high_C_clone_7), 5′-CCG CCC CCC CGC CCC TCC-3′; 18b_WT, 5′-TAA GAA GGA GAT ATA TCC-3′; and 18b_C, 5′-CCC CCC CCC CCC CCC CCC-3′. Oligonucleotides were added to the translation just prior to the mRNA.
Selected sequences were cloned into pET-3a (Novagen, Madison, WI) and sequence-verified minipreps were transformed into E. coli BL21(DE3)pLysS (Agilent, Santa Clara, CA) for expression. Individual colonies were inoculated into LB containing 100 µg/mL ampicillin (to maintain pET-3a) and 50 µg/mL chloramphenicol (to maintain pLysS) and grown for ∼16 h overnight at 37°C. Ampicillin was omitted from the negative control (background strain). The next morning, cultures were diluted 1∶50 in 1 mL LB without antibiotic and allowed to grow for 3 h at 37°C. Half of each culture was then induced with 1 mM isopropyl β-D-1-thiogalactopyranoside (IPTG). Cultures were grown for another 4 h at 37°C and analyzed on a Guava flow cytometer (Millipore). The average median fluorescence of three separate experiments was used to determine whether or not induction was appreciable (i.e., greater than two-fold over background fluorescence of the strain).
The 5′ UTRs of WT and 30-5-3 high C 7 were quantified using qRT-PCR with 5′_UTR_qPCR_fwd (5′-CCA CAA CGG TTT CCC TAA TTG T-3′), FLAG_qPCR_rev (5′-GTC ATC TTT GTA GTC CGC CAT-3′), and 5′_UTR_probe (6-FAM-5′-AGC GGA TAA CAA TAG AAA T-3′).
Raw sequences were filtered to make sure the randomized region was of the expected length (18 bases) and in the expected context (TGTTTAACTT upstream and ATGGCGGACT downstream). Sequences with an in-frame ATG present in the randomized region were excluded from analysis. For the rRNA comparison, a virtual library of 4,863 random 18-base sequences was generated (equal in size to the actual sequence pool analyzed). From each 18-base sequence, 19−k windows of length k were considered for k = 4–8. These 4,863×(19−k) windows were compared to E. coli 16S rRNA, and the number of reverse complements present in the virtual library for each window of length k on the 16S rRNA was recorded. Approximately 100,000 virtual libraries of this sort were generated to develop a probability distribution at each index of the 16S rRNA starting a k-base window. Bonferroni-corrected p-values are presented as P.rand in Table S1. The significance threshold was set at 0.01. For k = 7, significant windows neighboring at least one other significant window were considered to be part of a group of significant windows. PyMOL [51] was used to visualize these groups on the crystal structure. There appeared to be no correlation between the position of these groups on the crystal structure and the position of the complementary motif within the randomized region. Permuted (scrambled) 5′ UTRs were also used to calculate p-values (Bonferroni-corrected; P.perm in Table S1). P.rand allows us to recognize sequences that deviate from randomness in terms of their base composition and order of bases, while P.perm allows us to recognize the importance of the order of bases only. For the naïve motif search, all possible k-base motifs, k = 4–8, were generated. The virtual libraries (with random or scrambled 5′ UTRs) were again generated and the incidence of each k-base motif was assessed; to correct for multiple tests, FDR was applied, and the resulting q-values for the motif search are presented as Q.rand and Q.perm in Table S2. To analyze dependencies between motifs, each significant k-base motif (FDR<0.01) was assessed to determine if it was more likely to occur in a 5′ UTR context containing another particular motif. This dependency was quantified by a co-occurrence metric: [# 5′ UTRs having non-overlapping motifs 1 and 2]/[# 5′ UTRs having motif 2]. These values (when non-zero) are reported in Table S3.
mRNA secondary structure analysis was performed using the following procedure, which was adapted from previously published work [52]. Sequencing reads of selected library sequences were computationally trimmed to yield mRNA molecules consisting of a 26-base region immediately prior to the randomized region, the 18-base randomized region immediately prior to the start codon, and another 26-base region starting from the start codon. Each 70-base mRNA molecule was further processed to yield five overlapping 30-base windows using an offset of 10 bases. Finally, each 30-base window was assessed for secondary structure using the UNAFold suite (program melt.pl), and the corresponding ΔG values were recorded. For comparison, a library of 350,000 simulated mRNA molecules having random 18-base regions (probability of each base = 0.25) was assessed for secondary structure using the procedure described above.
|
10.1371/journal.pntd.0001006 | Use of Oral Cholera Vaccines in an Outbreak in Vietnam: A Case Control Study | Killed oral cholera vaccines (OCVs) are available but not used routinely for cholera control except in Vietnam, which produces its own vaccine. In 2007–2008, unprecedented cholera outbreaks occurred in the capital, Hanoi, prompting immunization in two districts. In an outbreak investigation, we assessed the effectiveness of killed OCV use after a cholera outbreak began.
From 16 to 28 January 2008, vaccination campaigns with the Vietnamese killed OCV were held in two districts of Hanoi. No cholera cases were detected from 5 February to 4 March 2008, after which cases were again identified. Beginning 8 April 2008, residents of four districts of Hanoi admitted to one of five hospitals for acute diarrhea with onset after 5 March 2008 were recruited for a matched, hospital-based, case-control outbreak investigation. Cases were matched by hospital, admission date, district, gender, and age to controls admitted for non-diarrheal conditions. Subjects from the two vaccinated districts were evaluated to determine vaccine effectiveness. 54 case-control pairs from the vaccinated districts were included in the analysis. There were 8 (15%) and 16 (30%) vaccine recipients among cases and controls, respectively. The vaccine was 76% protective against cholera in this setting (95% CI 5% to 94%, P = 0.042) after adjusting for intake of dog meat or raw vegetables and not drinking boiled or bottled water most of the time.
This is the first study to explore the effectiveness of the reactive use of killed OCVs during a cholera outbreak. Our findings suggest that killed OCVs may have a role in controlling cholera outbreaks.
| Simple measures such as adequate sanitation and clean water stops the spread of cholera; however, in areas where these are not available, cholera spreads quickly and may lead to death in a few hours if treatment is not initiated immediately. The use of life-saving rehydration therapy is the mainstay in cholera control, however, the rapidity of the disease and the limited access to appropriate healthcare units in far-flung areas together result in an unacceptable number of deaths. The WHO has recommended the use of oral cholera vaccines as a preventive measure against cholera outbreaks since 2001, but this was recently updated so that vaccine use may also be considered once a cholera outbreak has begun. The findings from this study suggest that reactive use of killed oral cholera vaccines provides protection against the disease and may be a potential tool in times of outbreaks. Further studies must be conducted to confirm these findings.
| Cholera is increasingly being reported, and more countries are now experiencing outbreaks [1], some lasting for several months. In 2001, the World Health Organization (WHO) recommended the use of oral cholera vaccines (OCV) in populations at risk in endemic areas but not reactively once an outbreak has begun [2]. While this recommendation has been updated in March 2010, to include reactive use of these vaccines [3], OCVs have only been used for reactive cholera control in 2000, when a live attenuated OCV (CVD-103HgR) was used in an outbreak in Micronesia [4]. The CVD-103HgR was assessed to be effective in this outbreak, although this was an observational study. In contrast, CVD-103HgR conferred no protection in the only randomized controlled efficacy trial of this vaccine [5], and this vaccine is no longer manufactured. There is one internationally licensed killed oral cholera vaccine, the recombinant B subunit killed OCV (rBS-WC, Dukoral, Crucell/SBL), but it has not been routinely adopted for public health use due to its high cost, limited duration of protection and logistic issues with vaccine administration. A variant of this oral vaccine, containing only killed whole cells (Vibrio cholerae O1 and O139) is manufactured in Vietnam following technology transfer from Swedish scientists. Vietnam is the only country in the world to use an OCV in its public health system for cholera control. Since 1997, this killed OCV (ORC-Vax) has been licensed and produced locally by the Company for Vaccine and Biological Production (VaBiotech) in Hanoi. The vaccine was found to confer 66% protection against an El Tor cholera outbreak occurring eight months following vaccination among all individuals aged 1 year and older [6] and 50% protection, three to five years after vaccination [7]. It is safe, inexpensive, and easy to administer [8]. Packaged in five-dose vials, each 1.5 ml liquid vaccine dose is drawn and squirted into the mouth by a syringe without a needle. Each dose contained: 5.0×1010 formalin-killed V. cholerae Inaba, El Tor strain Phil 6973; 2.5×1010 heat-killed V. cholerae Ogawa, classical strain Cairo 50; 2.5×1010 formalin-killed V. cholerae Inaba, classical strain 569B; and 5.0×1010 formalin-killed V. cholerae O139 strain 4260B. After oral administration, individuals are asked to drink water, but no oral buffer is required. Given in two doses, one to four weeks apart, it may be given to individuals aged one year and older.
Since the seventh pandemic reached Vietnam in 1964, cholera has been reported annually. A review of reported cases to the National Institute of Hygiene and Epidemiology (NIHE) from 1991 to 2001 showed that cholera is endemic in the central and southern provinces [9]. Compared with shigellosis and typhoid fever, cholera cases have decreased dramatically in 1997 to 2001. This decrease in cholera cases has been partly attributed to the extensive use of the killed OCV in Vietnam [10].
From 1997 to 2005, 9.2 million doses of the killed OCV have been used in the Expanded Programme of Immunization (EPI) of 20 cholera endemic provinces and metropolitan areas in Vietnam, mostly located in the central and southern areas (Figure 1). Vaccines are routinely provided in the endemic areas through regular monthly immunization sessions. In the routine EPI setting, depending on the commune, eligible children, aged 2–5 years are gathered for immunization on the same days for cholera vaccination. OCVs are provided 2 to 4 weeks apart. The killed OCV is also used preemptively in mass campaigns whenever an increase in the number of culture-confirmed cases are reported. National diarrheal disease surveillance is performed routinely and culture confirmation of organisms is available at the 61 provincial Centers for Preventive Medicine and in the national and four regional Institutes of Hygiene and Epidemiology. When cholera cases are detected in known endemic areas, mass vaccinations are arranged in designated locations such as schools, commune and district health facilities or government offices in the affected areas.
In October 2007, an increase in acute watery diarrhea cases was reported in Hanoi, caused by genetically altered Vibrio cholerae O1 Ogawa biotype El Tor producing classical biotype cholera toxin. Prior to this outbreak, the strain had never been isolated in Vietnam [11]. From 24 October to 4 December 2007, nearly 2,000 diarrhea cases were reported from Hanoi and neighboring provinces, of which 295 were laboratory confirmed. In response the Ministry of Health of Vietnam mandated the provision of free medical treatment for anyone suffering from acute diarrheal illness.
New cholera cases were identified on 24 December 2007 from Hanoi, thus, in the first week of January 2008, just prior to the Vietnamese Tet New Year, a decision was made to immunize two particularly hard hit districts of Hanoi – Hoang Mai and Thanh Xuan (combined population of ∼462,570). These districts are located close to waterways into which sewage drains. The Vietnam National Institute of Hygiene and Epidemiology (NIHE) together with the Ministry of Health launched the mass vaccination campaign on 16–28 January 2008, providing two doses of the killed oral cholera vaccine, spaced one week apart. Because of the absence of cases detected during the outbreak among children less than 10 years of age, vaccines were only provided to residents aged 10 years and older. Pregnant residents were also not eligible for vaccination. The campaign was announced in newspapers and radio and eligible residents were invited to proceed to commune health centers. Vaccination cards were provided to vaccinees and logbooks containing the names of vaccine recipients were maintained. It was estimated that ∼80% of the estimated 370,000 age-eligible individuals received one or more doses of the killed OCV. In addition, educational health campaigns were also conducted to inform the public of the signs of illness and to improve sanitary practices.
From 24 December 2007 to 6 February 2008, 59 diarrhea cases (33 culture confirmed V. cholerae O1) were identified, all cases coming from Hanoi. No cases were detected until 5 March 2008, when the number of diarrhea cases increased and V. cholerae O1 Ogawa was again identified as the causative agent. The NIHE requested the International Vaccine Institute (IVI) to assist in the outbreak investigation, specifically looking into the role of vaccines for control. This provided a unique opportunity to assess the effectiveness of reactive oral cholera vaccination in a cholera outbreak, as there has been little experience in the use of OCVs in cholera epidemics. Figure 2 shows the clinical cholera cases in Hanoi from 24 October 2007 to 15 July 2008.
A matched, hospital-based, case-control investigation was conducted from 8 April to 10 June, 2008. Hanoi has nine urban districts with a population of ∼2.9 million [12]. Hospitalized patients from the two vaccinated districts - Hoang Mai and Thanh Xuan, as well as the unvaccinated districts - Dong Da and Cau Giay were invited to participate in the outbreak investigation (Figure 3). These districts have a combined population of ∼1 million [12]. These districts have similar population characteristics, environmental conditions and epidemiological data from past cholera outbreaks. Residents of these districts are also served in common and have equal chances of attending five hospitals including the National Institute of Infectious and Tropical Disease (NIID) Hospital, Bach Mai District Hospital, Saint Paul Hospital, Dong Da District Hospital and Transportation Hospital. Case and control exposure histories of subjects from Hoang Mai and Thanh Xuan , were compared for evaluation of risk factors and effectiveness of killed OCV use during the outbreak, the results of which are presented here.
Patient admission logbooks at the five hospitals were reviewed daily to identify patients admitted for diarrhea. Hospital records of identified patients were then reviewed. Patients who met the clinical case definition for cholera were invited to participate. A cholera case was defined, a priori, as being hospitalized for diarrhea with illness onset of 8 April to 20 May 2008, with diarrhea defined as 3 or more loose, liquid or watery bowel movements in any 24 hour period; were 10 years of age or older and a resident of any of the 4 districts of interest. Cases were identified without knowledge of the vaccination status.
One matched control per case was recruited from wards of the same hospital, except for cases admitted to NIID, wherein controls were identified from the trauma and surgical wards of Bach Mai Hospital, an adjacent general hospital. Patient admission logbooks were reviewed to identify controls hospitalized for non-diarrheal conditions. Controls were matched for each case by the date of presentation (±5 days), age group (10–20 years old, 21–40 years old, 40+ years old), gender and district of residence. The first control in the logbook that fulfilled the matching characteristics to the case was identified and invited to participate. Controls were chosen by reviewers who were unaware of the vaccination status of the patients.
Data were obtained through transcription of clinical records and subject interviews using a standardized questionnaire. Demographic characteristics including occupation, water supply (tap water, public well), behavioral characteristic such as hand washing and sanitation (toilet with flush, latrine, none), as well as exposure factors (intake of raw vegetables, dog meat, shrimp paste; not drinking boiled or bottled water), were collected. Vaccination status including the number and date of dosing was verbally ascertained based on subject recall. When available the reported dosing dates were cross-checked against a vaccination card. In order to evaluate the use of the OCV in this outbreak setting we defined “vaccinated” a priori as receipt of one or two doses of OCV from 16–28 January 2008 without further consideration to dosing interval or interval between vaccination and date of selection into the study. Microbiological culture results, completed by and according to the standard operating procedures of the admitting hospital laboratory, were also obtained during the study when available.
To detect 50% vaccine protection, we assumed the following: 40% of controls would be vaccinated; the correlation of vaccine histories among matched cases and controls, phi, was .05; and with 80% power at P<.05 (2-tailed), at least 172 cases and 172 controls were required for the investigation.
Characteristics and exposures of hospitalized cases and controls from the vaccinated and unvaccinated districts were compared. To assess the effect of vaccination, we included diarrheal cases and controls hospitalized for non-diarrheal causes from the vaccinated districts. Baseline characteristics were statistically compared using McNemar's test for dichotomous variables and the paired Student t-test for continuous variables. Only complete pairs in which both the case and the control had exposure measurements were included, and the information contained in the incomplete pairs was ignored. The adjusted matched odds ratio (OR) and 95% confidence interval (CI) for calculation of vaccine effectiveness was determined using multivariate conditional logistic regression [13]. Statistical analysis was planned at the outset, to include all variables with p<0.05 in univariate analysis and the primary variable of interest (vaccination status) in the multivariable model. Vaccine effectiveness was calculated as: (1-matched OR)×100. All p values and 95% confidence intervals, estimated from the point estimates and standard errors for the coefficient for the vaccination variable in the models, were interpreted in a two- tailed manner. Statistical significance was designated as a p value<0.05. All statistical analyses were performed using Stata10 (StataCorp, College Station, TX).
The study qualified for exemption from review by the IVI Institutional Review Board and Ethical Review Committee of NIHE as the study was conducted as part of an outbreak investigation establishing risk factors and modifiers. Verbal consent was obtained in lieu of written consent from both cases and controls as the project was conducted as part of an outbreak investigation. Consent was documented in a logbook.
We enrolled 126 matched pairs of cases and controls for the outbreak investigation; one matched pair was excluded when on review the case definition was not met by the case (Figure 4). After exclusion of this matched pair, among cases, the ages ranged from 17 to 86 years old while the control age range was 15 to 80 years old. Thirty-seven percent of cases had vomiting and 76% had some or severe dehydration on admission. Among those with severe dehydration, only one was vaccinated. Of the 99 cases whose stools were tested, 74 subjects had culture confirmed V. cholerae O1 (75%). Only one vaccine recipient had culture confirmed cholera. Table 1 shows the causes of hospitalization for the controls.
Of the 125 matched pairs, 54 pairs (43%) were residents of districts where the mass vaccination campaign was carried out and were included in this evaluation of vaccine effectiveness. We compared the baseline characteristics of cases and controls from Huang Mai and Thanh Xuan, where the mass vaccination campaigns were carried out, and found no significant differences in demographic and socio-economic characteristics (Table 2). On comparing the exposure of cases with controls, intake of raw vegetables and not drinking boiled or bottled water were found to be significantly different (p<0.05). Similar results were obtained when comparing all cases and controls in the outbreak investigation, including patients from both the vaccinated and unvaccinated districts (data not shown). Because dog meat is customarily eaten with raw vegetables and 70% of those who ate dog meat also ate raw vegetables, we decided to combine these in the multivariate regression model.
Of subjects from the vaccinated districts, 8 of 54 cases (15%) and 16 of 54 controls (30%) were classified as vaccinated, having received at least one dose of the killed OCV from 16–28 January 2008. Seventy-five percent (6/8) of vaccinated cases and 63% of vaccinated controls (10/16) received two doses of killed OCV during the vaccination campaign. The unadjusted vaccine effectiveness (VE) was 54% (95% CI −31% to 84%; p-value = 0.144), however, after adjusting for factors which were found to be significantly associated with being a cholera case at P<0.05 in univariate analyses (intake of dog meat or raw vegetables and not drinking boiled or bottled water most of the time) (Table 3), the killed OCV was found to have an effectiveness of 76% (95% CI 5% to 94%, p = 0.04).
This is the first study to report on the use of killed OCV in an outbreak situation. While a significant association was detected between receipt of at least one dose of the killed OCV and protection against cholera, our study has several limitations.
Because there may be inherent differences in health care utilization and knowledge among those who presented for vaccination and those who refused vaccination [14], [15], bias may have been introduced in our assessment for vaccine protection, and may have exaggerated our results. The protective effect may have been augmented, as it has been shown that people refusing participation are more likely to engage in high-risk behaviors as compared to vaccines [15]. However, there were no differences in the baseline demographic, socioeconomic and exposure characteristics of vaccinated and non-vaccinated cases and controls. Moreover, there were several factors that may have decreased the true protective effect of the vaccine during this outbreak, namely: (1) individuals with a recent history of cholera-like diarrhea may not have participated in the campaign and were included in the control group (2) recipients of a single dose of the vaccine were included in the analysis (3) vaccinees may have been more likely to use the treatment centers for the care of diarrhea compared to refusers. A comparison of a partially immunized vaccine group to a control group with varying levels of natural immunity would tend to depress apparent vaccine protection against subsequent cholera.
Our evaluation was also limited by use of a clinical case definition without culture confirmation, however we used a strict case definition and random cases were culture confirmed. Moreover, inclusion of non-culture confirmed cases, if ever, would have depressed the protection afforded by vaccination as some cases may not be due to V. cholerae.
We did not reach the sample size required (54 instead of the desired 172) because of difficulty enrolling controls during this outbreak, throughout which most hospital beds were occupied by cholera cases. The smaller sample size may explain the unadjusted VE as being not statistically significant. We tried to limit selection bias by enrolling cases and controls without prior knowledge of their vaccination status. Moreover, in order to prevent interviewers from overzealously eliciting vaccination history, several exposure questions were included in the questionnaire.
Lastly, our study was initiated more than two months after the campaign, thus we were unable to include cases proximate to vaccination, however since the outbreak was prolonged and recurrent and vaccine effectiveness lasts for three to five years [7], measurement of the effectiveness of OCV use in this setting was still warranted.
To our knowledge, this is the first study that explored the reactive use of a killed OCV in an outbreak. In Hanoi, the outbreak was described as having occurred in three waves, each separated by 14 to 26 day intervals with no recorded cases in between each wave. Vaccination was performed while the second wave was ongoing (see Figure 3). Since the mass vaccination campaign was performed in the two districts that have been most affected in the previous waves of diarrheal cases, the characteristics of the residents in these districts may have been different from other areas that make them vulnerable to diarrheal outbreaks and may be more amenable to district specific interventions. However, comparison of baseline characteristics and exposures of patients from the vaccinated (Hoang Mai and Thanh Xuan) and unvaccinated districts (Dong Da and Cau Giay) showed no statistically significant differences (data not shown).
In the recently updated WHO recommendations, consideration for both preemptive and reactive use of OCVs is supported after assessment of local infrastructure and epidemiology. A model of a refugee camp based cholera outbreak in Africa compared the cost-effectiveness of several cholera controls strategies, including establishment of treatment centers and reactive vaccination. Based on duration of the hypothetical outbreak and the size of the hypothetical camp, reactive vaccination will only be a cost-effective option if the price of the vaccine falls below $0.22 per dose [16]. However, there were several limitations to this analysis [17] and this study did not account for large prolonged outbreaks such as those seen recently in Zimbabwe [18], [19], Angola [20], [21] and Vietnam [11], which would favor reactive vaccination.
Since 1996, extensive cholera outbreaks of this magnitude had not been reported in Vietnam, especially in areas where the killed OCV is routinely used. Between 5 March and 22 April 2008, the Vietnamese Ministry of Health reported 2,490 cases of severe acute watery diarrhea including 377 that were positive for V. cholerae O1 Ogawa [22]. Twenty provinces in the northern areas were affected in 2007 to 2008. No deaths were reported during these outbreaks indicating good case management. On the other hand, in Africa, cholera outbreaks are deadly. In Zimbabwe alone from August 2008 to May 2009, almost 100,000 cases have been identified with more than 4,000 deaths [18], 61% of whom did not reach a health facility for treatment [19]. Similarly in Angola, an outbreak from February to June 2006 with 46,758 cases and 1,893 deaths [20], [21] were reported with case fatality rates in some provinces tragically reaching up to 30% [20]. Provisions for clean water, adequate sanitation and good case management are necessary for controlling cholera, however, these are unlikely to happen in the near future in most of the developing world where cholera continues to cause significant hardship and misery. New measures need to be taken. Prior to the release of the March 2010 WHO position paper several groups were pressing for a rethink of the WHO stand on vaccine use for outbreaks [23]. The results of our study are consistent with earlier evaluations of the protective effects of OCV [6]. Microbiologic studies have shown that the outbreak was caused by the new strain of El Tor V. cholerae O1 producing classical cholera toxin [11]. This new strain has been increasingly reported in Asia and in parts of Africa [24]–[26] with some indications of increased severity [27]. The killed OCV provided protection against this new strain suggesting that there may be a role for reactive use of the killed OCV in future cholera outbreaks.
The Vietnamese killed OCV has now been extensively modified by the IVI to comply with WHO and current Good Manufacturing Practices (cGMP) standards. The modified vaccine was recently licensed in Vietnam (mORC-VAX). In order to expand its use internationally and to allow purchase by United Nations agencies, technology transfer of the vaccine production process was made by the IVI to Shantha Biotechnics in India where it is now licensed (Shanchol®). This modified vaccine with higher antigenic content than the previous versions has been found to be safe and protective in India [28] and resulted in comparable vibriocidal immune responses after one or two doses of the vaccine raising the possibility that it may be used as a single dose, which would greatly simplify vaccine delivery in times of outbreaks [29]. Further studies to confirm our findings are necessary; however, these results provide hope that the vaccine will be used not only for endemic cholera control but in times of outbreaks as well, when mortality may be higher [30].
|
10.1371/journal.pntd.0002411 | Phylogeography of Japanese Encephalitis Virus: Genotype Is Associated with Climate | The circulation of vector-borne zoonotic viruses is largely determined by the overlap in the geographical distributions of virus-competent vectors and reservoir hosts. What is less clear are the factors influencing the distribution of virus-specific lineages. Japanese encephalitis virus (JEV) is the most important etiologic agent of epidemic encephalitis worldwide, and is primarily maintained between vertebrate reservoir hosts (avian and swine) and culicine mosquitoes. There are five genotypes of JEV: GI-V. In recent years, GI has displaced GIII as the dominant JEV genotype and GV has re-emerged after almost 60 years of undetected virus circulation. JEV is found throughout most of Asia, extending from maritime Siberia in the north to Australia in the south, and as far as Pakistan to the west and Saipan to the east. Transmission of JEV in temperate zones is epidemic with the majority of cases occurring in summer months, while transmission in tropical zones is endemic and occurs year-round at lower rates. To test the hypothesis that viruses circulating in these two geographical zones are genetically distinct, we applied Bayesian phylogeographic, categorical data analysis and phylogeny-trait association test techniques to the largest JEV dataset compiled to date, representing the envelope (E) gene of 487 isolates collected from 12 countries over 75 years. We demonstrated that GIII and the recently emerged GI-b are temperate genotypes likely maintained year-round in northern latitudes, while GI-a and GII are tropical genotypes likely maintained primarily through mosquito-avian and mosquito-swine transmission cycles. This study represents a new paradigm directly linking viral molecular evolution and climate.
| Although Japanese encephalitis virus (JEV) is a major cause of death and disability throughout tropical and temperate Asia, little is known about the evolution, geographical distribution and epidemiology of the five JEV genotypes (genetically distinct groups). To address this gap in our knowledge, we performed a genetic-based geographical analysis using the largest JEV sequence dataset assembled to date, including 487 viral sequences sampled from 12 countries over 75 years. We showed that both the newly and previously dominant genotypes of JEV are associated with temperate climates and are maintained throughout the cold winter months in northern Asia, likely by hibernating mosquitoes (survive throughout the winter), vertical transmission in mosquitoes (female to offspring), cold-blooded vertebrates and/or bats.
| Japanese encephalitis virus (JEV) belongs to the JEV serocomplex within the genus Flavivirus, family Flaviviridae. Recurrent epidemics of summer encephalitis suggestive of JE were recorded in Japan from 1871 onwards and major epidemics occurred in 1924 (6,000 cases, with a 60% case fatality rate), 1929, 1935 and 1937 [1]. The prototype Nakayama strain of JEV was isolated in mice from the brain of a male that died of summer encephalitis in Tokyo, Japan in 1935 [1]. The seasonal occurrence of epidemic encephalitis coupled with the abundance of culicine mosquitoes led to suggestions that JEV was transmitted by mosquito vectors, leading to the subsequent recovery of the virus from rice-paddy breeding Culex tritaeniorhynchus mosquitoes in 1938 [2]. A series of ecological studies performed in Japan in the late 1950s established waterbirds as maintenance hosts of the virus, domestic swine as major amplifying hosts, and Cx. tritaeniorhynchus as the principal vector between these vertebrate hosts and the incidental, dead-end human host [3], [4], [5], [6], [7], [8], [9], [10], [11], [12].
JEV circulates throughout most of Asia, with the northern limit of virus activity extending north into maritime Siberia. In recent years the geographical distribution of JEV has expanded, reaching east into Saipan in 1990 [13], west into Pakistan in 1992 [14] and south into the Torres Strait between Papua New Guinea and Australia in 1995 [15]. JEV epidemics occur in temperate zones, with the majority of cases occurring in summer or monsoon season months. In contrast, JEV is endemic in tropical regions and transmission occurs year-round at lower rates [16]. Despite the availability of effective vaccines against JEV, the virus is still considered the most important etiologic agent of epidemic encephalitis worldwide, causing an estimated 68,000 cases and a reported 10,000–15,000 deaths annually [17]. Of symptomatic infections, 20–30% are rapidly fatal, 30–50% develop long-term neurologic and/or psychiatric sequelae, and only 20–50% fully resolve the disease [17].
Like other flaviviruses, JEV possesses an 11 kilobase, single-stranded, positive-sense RNA genome containing 5′ and 3′ untranslated regions, and a single open reading frame (ORF) encoding a polyprotein that is co- and post-translationally cleaved by viral and host proteases into three structural proteins: the capsid (C), the precursor of the membrane (prM), and the envelope (E), as well as seven non-structural proteins [18]. The E protein represents the major constituent of the mature virion surface and is the dominant antigen involved in the elicitation of virus neutralizing antibodies [19].
Phylogenetic studies have divided JEV into five genotypes. GI includes isolates collected in northern Australia, northern Cambodia, China, India, Japan, Korea, Laos, Malaysia, Taiwan, Thailand and Vietnam between 1967 and present. GII includes isolates collected sporadically in northern Australia, Indonesia, Korea, Malaysia, Papua New Guinea and southern Thailand between 1951 and 1999. GIII has been the source of annually occurring epidemics of encephalitis and includes isolates collected in China, India, Indonesia, Japan, Korea, Malaysia, Myanmar, Nepal, Philippines, Sri Lanka, the former Soviet Union, Taiwan, Thailand and Vietnam between 1935 and present. GIV includes seven isolates collected in Indonesia between 1980 and 1981 from mosquitoes only. GV includes three isolates collected in Malaysia, China and South Korea between 1952 and 2010. Previous investigations noted GI and GIII viruses were collected mostly in temperate zones, while GII and GIV isolates were collected mostly in tropical zones [20], [21]. However, the statistical significance of this observation has never been tested using a comprehensive dataset of JEV isolates, with information regarding the isolates' genotypes and locations of collection.
The geographical distribution of JEV has expanded in recent years, causing outbreaks of encephalitis in immunologically naïve populations. In addition, the molecular epidemiology of the virus has changed over this period of time. From the isolation of the prototype Nakayama strain of JEV in 1935 until recently, GIII was the most frequently isolated genotype throughout Asia. However, over the past two decades, multiple reports have indicated that GI has displaced GIII as the most frequently isolated virus genotype in a number of Asian countries including China [22], Thailand [23], South Korea [24], Japan [25], Malaysia [26], Vietnam [27], India [28] and Taiwan [29]. Further, following the isolation of the GV Muar isolate [30] in 1952 from an encephalitic patient originating in Malaysia, the genotype remained undetected for almost 60 years until a pool of Cx. tritaeniorhynchus collected in the Tibetan Province of China in 2009 yielded the GV XZ0934 isolate [31] and a pool of Culex bitaeniorhynchus collected in South Korea in 2010 yielded the GV 10-1827 isolate [32]. A recent evolutionary study utilizing sequence information derived from the ORF of 35 JEV isolates (22 GIII isolates) revealed that JEV originated from its ancestral virus around 1500 [33] and an earlier evolutionary study using 18 genomic JEV sequences (14 GIII) proposed that this evolutionary event occurred in the Indonesia-Malaysia region [34]. Due to the small viral sequence sample sizes, neither of these studies were able to robustly examine the evolution, epidemiology or geographical distribution of the genotypes of JEV. In disagreement with the results of previous studies [33], [34], a recent study utilizing 98 genomic sequences, 76 of which were derived from Chinese virus isolates, estimated that JEV originated from its ancestral virus around 300 AD [35]. This difference was likely due to the slow evolutionary rate estimated for the Chinese JEV study [35] relative to previous JEV evolutionary studies [33], [34], as well other flavivirus evolutionary studies [36], [37], [38], [39]. Prior to the work presented here, no studies have utilized a comprehensive dataset of molecular sequences to examine the phylogeography and epidemiology of the virus genotypes.
Although there is a paucity of ORF sequences of wild-type isolates, extensive sequencing of the phylogenetically-informative E gene of both old and new JEV isolates in recent years has resulted in a large, spatiotemporally distributed collection of viral sequence data. Therefore, we performed a phylogeographic analysis on a dataset consisting of E gene sequence information derived from 487 JEV isolates (largest collection of JEV sequences assembled to date) to address the following key questions: 1) When and where did the virus and its genotypes originate, and what is their geographical range? 2) Is there an association between genotype and climate of virus collection (temperate versus tropical zones)? 3) What amino acid sites within the E protein were involved in the phylogenetic divergence of JEV and were any of these sites subject to diversifying and/or directional selection?
All available sequences for the E gene of JEV isolates were retrieved from GenBank in July 2011. The initial JEV E gene dataset was pruned of sequences representing non wild-type virus isolates, duplicate isolates, and isolates absent of information regarding the date and country of collection. The pruned dataset consisted of 489 sequences. The E gene sequences of two JEV isolates (M859/Cambodia/1967/Mosquito and KE-93-83) obtained from the World Reference Center for Emerging Viruses and Arboviruses (WRCEVA) at the University of Texas Medical Branch (UTMB), were determined for analysis in this study utilizing previously described methods [40], [41], [42].
Recombination can invalidate the results of coalescent analyses. Therefore, the nucleotide sequence alignment file was analyzed for potential recombination events using RDP [43], GENECONV [44], Chimaera [45], MaxChi [46] and Bootscan [47] methods implemented in RDP3 v Beta 41 [48]. Common program settings were to perceive sequences as linear, require phylogenetic evidence, refine breakpoints and check alignment consistency, while all method-specific program settings remained at their default values. The highest acceptable p-value was set at 0.05, after considering Bonferroni correction for multiple comparisons. Potential recombination events were those that were identified by at least two methods. The breakpoint positions and recombinant sequence inferred for the potential recombination events were manually confirmed using the phylogenetic and recombination signal analysis features in RDP3. The K82P01 and K91P55 sequences were confirmed as recombinants (Table S1). These two isolates were not available from the WRCEVA at UTMB to re-sequence; therefore, the two corresponding sequences were removed from the dataset, leaving a final dataset of 487 sequences.
To make an initial identification of the genotype of the JEV E gene sequences, neighbor-joining (NJ) and maximum-likelihood (ML) phylogenies were generated using SeaView v 4.2.12 [49] and PhyML v 3.0 on the South of France bioinformatics platform [50], respectively.
The final dataset of 487 JEV E gene sequences included information regarding the year, host and country of collection of the corresponding virus isolates. Sequences derived from isolates collected north of the Tropic of Cancer (23.5°N) were classified as temperate, while sequences derived from isolates collected south of the Tropic of Cancer were classified as tropical. The climate corresponding to five Taiwanese sequences could not be ascertained and therefore these sequences were not included in the climate phylogeographic analysis described below.
To estimate the date and location of the most recent common ancestor (MRCA) of the five genotypes and the overall rate of molecular evolution, time-scaled Bayesian phylogenies (country and climate) were inferred from the JEV E gene sequence dataset using a Bayesian Markov Chain Monte Carlo (MCMC) method implemented in BEAST v 1.6.1 [51].
An SDR06 nucleotide substitution model [52], a relaxed-uncorrelated exponential molecular clock and a piecewise constant Bayesian skyline demographic model with 20 coalescent-interval groups [53] were used in all analyses. The relaxed-uncorrelated exponential molecular clock was found to best-fit the data when Bayes factor (BF) values were calculated (Tracer v 1.5.1) [54] to evaluate the relative fit of strict and relaxed molecular clock models to the data by determining the natural logarithm of the ratio of the marginal likelihoods of the competing models [55]. The good fit of this relaxed clock model to the data has recently been shown to be an artifact of the harmonic mean estimator [56]. However, preliminary analyses showed that the selection of a particular relaxed molecular clock model had little effect on the results.
To infer the probable geographic origin of the MRCA of the genotypes of JEV, the BEAST input files (country and climate) created in BEAUti v 1.6.1 [51] were edited to include the Bayesian stochastic search variable selection procedure [57].
The Bioportal at the University of Oslo [58] was used to execute the MCMC analyses for 600 million generations. This was achieved by using LogCombiner v 1.6.1 [51] to compile 12 independent runs of 50 million generations (sampled every 1,000th state) to attain convergence, which was assessed by examining the trace and effective sample size statistics for each model parameter in Tracer v 1.5 [54]. TreeAnnotator v 1.6.1 [51] was used to summarize the posterior tree distribution and annotate both country and climate maximum clade credibility (MCC) phylogenies, which were viewed in FigTree v 1.3.1 [59]. Each of the nodes of the Bayesian MCC phylogenies were annotated with posterior probability (PP) values, estimated median dates of the MRCA with corresponding 95% HPD values, and state PP values for each plausible geographic location of origin (country and climate). In addition, BOA v 1.15 [60] implemented in R v 2.15.1 [61] was used to calculate a 50% HPD interval for the date of the root of the phylogeny.
Maps showing the distributions of sequences according to sampling location (country and climate) were created using GIMP v 2.6.12 from a blank map of Asia.
To test the null hypothesis of no association between genotype and climate, a Fisher's exact test was performed at α = 0.05 (IBM SPSS Statistics v 20). Post-hoc analyses were then performed to determine which cell(s) in the table of genotype versus climate contributed the most to the statistically significant Fisher's exact test. Adjusted standardized residuals (z-scores) were calculated and the Bonferroni method was used to correct for multiple comparisons. The adjusted standardized residual values were then compared against the critical z-value (±1.96) for α = 0.05 (IBM SPSS Statistics v 20). Only GI-a, GI-b, GII and GIII were considered in these analyses, as the dataset included only three sequences each for GIV and GV.
The null hypothesis of no phylogeny-trait association was further evaluated at α = 0.05 using the association index (AI), parsimony score (PS), unique fraction (UniFrac), nearest taxa (NT), net relatedness (NR), phylogenetic diversity (PD) and maximum exclusive single-state clade size (MC) statistics calculated from the posterior set of trees generated by BEAST in Befi-BaTS v 0.1.1 [62].
Nonsynonymous substitutions involved in the phylogenetic divergence of the five genotypes of JEV were identified within the E protein alignment. The E gene alignment was evaluated for statistically significant evidence of positive selection (ratio of nonsynonymous to synonymous nucleotide substitutions [dN/dS] >1; p<0.05) using the single-likelihood ancestor counting (SLAC), fixed effects likelihood (FEL) and internal FEL (IFEL) methods [63] available on the Datamonkey webserver [64]. All analyses of positive selection utilized a NJ phylogeny and the reversible nucleotide substitution model. Evidence of directional selection within the E protein alignment was evaluated using the directional evolution of protein sequences (DEPS) [65] method implemented in HyPhy v 2.0 [66]. The DEPS method utilized a Bayesian phylogeny and the Jones, Taylor, Thorton amino acid substitution model to assess for the presence of statistically significant shifts in amino acid residue frequencies (p<0.05) and/or a statistically significant large number of substitutions toward a particular residue (BF>100).
Of the 487 isolates in the JEV dataset (Table S2), the majority belonged to GI-b and GIII, the most common viral host was the mosquito, the most frequent decade of isolation was the 2000s, the most common country of virus isolation was Japan and the majority of isolates were collected in temperate climates (Table 1).
The overall median evolutionary rates estimated from the JEV E gene country and climate datasets were 5.33×10−4 (95% HPD: 3.92×10−4, 6.52×10−4) and 5.51×10−4 (95% HPD: 4.24×10−4, 6.67×10−4) substitutions/site/year, respectively. These estimates are slightly higher and the 95% HPD intervals are slightly wider compared to estimates previously obtained from a dataset of 35 JEV ORF sequences (mean: 4.35×10−4 substitutions/site/year, 95% HPD: 3.49×10−4, 5.30×10−4) [33]. This variation was likely due to the fact that more temporal signal can be extracted from longer alignments.
Figures 1 and 2 show the geographical distribution of the JEV sequences included in this study according to the country and climate of collection, respectively. Country and climate Bayesian MCC phylogenies are shown in Figures 3 and 4, respectively. As expected, the topologies of the BEAST phylogenies are supported by the NJ and ML phylogenies (Figures S1 and S2), and are similar to recently published phylogenies generated from both ORF and E gene sequence information for GI-V of the virus [31], [32], [33]. All four of the phylogenies inferred in this study support the division of GI into two clusters, GI-a and GI-b, where the GI-a clade consists of 15 isolates sampled in Cambodia, Thailand and Australia between 1967 and 2005 and GI-b includes 219 isolates sampled from Vietnam, Thailand, Japan, Korea, China and Taiwan between 1979 and 2009. Estimated dates of the MRCA and state PP values in support of each of the 12 countries are presented in Table 2 for the key nodes within the country Bayesian MCC phylogeny (Figure 3), and estimated dates of the MRCA and state PP values in support of tropical and temperate climates of divergence are presented in Table 3 for the key nodes within the climate Bayesian MCC phylogeny (Figure 4).
The country Bayesian MCC phylogeny (Figure 3) and the country map (Figure 1) show that GV includes three isolates sampled in China, South Korea and Malaysia between 1952 and 2010, GIV includes three isolates sampled in Indonesia between 1980 and 1981, GIII includes 234 isolates sampled in China, India, Indonesia, Japan, Korea, Sri Lanka, Taiwan and Vietnam between 1935 and 2009, GII includes 28 isolates sampled in Australia, Indonesia, Korea and Malaysia between 1951 and 1999, GI-a includes 15 isolates sampled in Cambodia, Thailand and Australia between 1967 and 2005, and GI-b includes 219 isolates sampled from Vietnam, Thailand, Japan, Korea, China and Taiwan between 1979 and 2009.
Phylogeographic analysis estimated that the date of the MRCA of JEV lies between 1506 and 1704 with a posterior probability of 50%, and between 1022 and 1800 with a posterior probability of 95% (median: 1553, 50% HPD: 1506, 1704; 95% HPD: 1022, 1800). These estimates are consistent with those recently inferred using a dataset of 35 JEV ORF sequences (mean: 1559; 95% HPD: 1509, 1635 [33]. The increased width of the 95% HPD intervals for the date of the MRCA of JEV for the E gene sequence dataset compared to the ORF gene sequence dataset can again be attributed to the fact that more temporal signal can be extracted from longer alignments. In agreement with previous suggestions regarding the origin of JEV [34], the root (MRCA of JEV) state PP values for all locations range between 0.03 and 0.19, with the highest state posterior probability values corresponding to Malaysia and Indonesia (0.19 and 0.18, respectively).
Of the five genotypes of JEV, the MRCA of GIII occurred earliest in time (median: 1894; 95% HPD: 1857, 1916) possibly in Japan (state PP: 0.57), followed by the MRCA of GV (median: 1902; 95% HPD: 1813, 1938) possibly in Malaysia (state PP: 0.40), the MRCA of GII (median: 1913; 95% HPD: 1867, 1939) possibly in Indonesia (state PP: 0.40), the MRCA of GI (median: 1936; 95% HPD: 1908, 1957) possibly in Vietnam (state PP: 0.44) and, most recently, the MRCA of GIV (median: 1971; 95% HPD: 1948, 1977) possibly in Indonesia (state PP: 0.98) (Figure 3, Table 2).
Within GI, the MRCA of GI-a occurred first (median: 1949; 95% HPD: 1927, 1962) possibly in Thailand (state PP: 0.43), followed by the MRCA of GI-b (median: 1961; 95% HPD: 1941, 1971) possibly in Vietnam (state PP: 0.56) (Figure 3, Table 2).
The MRCA of the recently emerged GV isolates (XZ0934 [China, 2009] and 10-1827 [South Korea, 2010]; node PP: 1.00) occurred recently (median: 1997; 95% HPD: 1982, 2006) possibly in Korea (state PP: 0.51) (Figure 3, Table 2).
The most striking observation from the climate Bayesian MCC phylogeny (Figure 4) and the climate map (Figure 2) is that GV includes isolates sampled from temperate and tropical locations, GIV includes isolates sampled from only tropical locations, GIII and GI-b include isolates sampled primarily from temperate locations, and GII and GI-a include isolates sampled primarily from tropical locations. The posterior probabilities for a tropical or temperate climate at the root of the tree were approximately equal (tropical state PP: 0.53, temperate state PP: 0.47) (Figure 4, Table 3). The MRCA of GIII (state PP: 0.97) and the recently emerged GV isolates (state PP: 0.99) was most likely in temperate Asia, while the MRCA of GV (state PP: 0.65), GII (state PP: 0.77), GI (state PP: 0.87), GI-a (state PP: 0.97), GI-b (state PP: 0.67) and GIV (state PP: 0.99) was most likely in tropical Asia (Figure 4, Table 3).
A Fisher's exact test was used to statistically evaluate the observed relationship between genotype and climate. Based on α = 0.05, we rejected the null hypothesis of no genotype-climate association and concluded that there was a statistically significant relationship between genotype and climate (Fisher's exact test: 173.48; exact two-sided p-value: 0.000). Post-hoc analysis revealed that GIII included significantly more isolates sampled from temperate climates than expected (adjusted standardized residual: 4.0), GII included significantly more isolates sampled from tropical climates than expected (adjusted standardized residual: 12.4), GI-a included significantly more isolates sampled from tropical climates than expected (adjusted standardized residual: 9.3) and GI-b included significantly more isolates sampled from temperate climates than expected (adjusted standardized residual: 5.2). The phylogeny-trait association test of genotype-climate phylogenetic structure also failed to reject the null hypothesis of no association between genotype and climate (Table S3); thereby, providing further evidence that JEV genotype is associated with climate.
Forty-four genotype-defining nonsynonymous substitutions were identified within the E protein (Table S4), 36 of which were GV-specific (13 non-conservative). Selection analyses were then performed to determine if any of these genotype-defining nonsynonymous substitutions might have played a role in the adaptation of the viral genotypes to their respective environments. No positively selected sites were identified using the SLAC, FEL and IFEL methods. However, the DEPS method revealed elevated substitution rates towards seven residues (Table S5) and thirteen sites were identified to be involved in this directional evolution (Table S6). One of these directionally selected sites (129; preferred residue: M) corresponded to the site of a genotype-defining nonsynonymous substitution (129 [GV: I, GIV: T, GIII: T, GII: T, GI: M]).
In accordance with previous analyses, we estimated that JEV originated in the Indonesia-Malaysia region [34] around the 16th century [33]. However, as expected for a node so far back in time, the posterior probability values in support of the origination of JEV in the Indonesia-Malaysia region were low. Nevertheless, as emphasized previously, all virus genotypes have been found in the Indonesia-Malaysia region and large epidemics suggestive of JEV have never been reported to occur in this region [34]. These lines of evidence are consistent with the virus having evolved in the Indonesia-Malaysia region [34]. Interestingly, based on the results of an amino acid signature analysis, others have suggested that Asian JEV and Australian Murray Valley encephalitis virus may have evolved from a virus related to the African Usutu virus in the Southeast Asia-Australasia region [67].
Phylogeographic analysis estimated that GIII evolved in temperate Asia (Japan) around the late 19th century. These estimates are coincident with the first reported summer epidemics of encephalitis suggestive of JE, which occurred in 1871 in Japan [1]. Following the isolation of the prototype Nakayama strain of JEV (a GIII virus) from Japan in 1935, GIII has been found throughout most of Asia including China, India, Indonesia, Korea, Japan, Sri Lanka, Taiwan and Vietnam.
Statistical analyses indicated that GIII did indeed include significantly more temperate isolates than expected under the null hypothesis of no association between genotype and climate. The paucity of GIII viruses sampled from tropical regions and the genetic relatedness of GIII viruses sampled years apart suggests that the annual re-introduction of GIII viruses from tropical regions to temperate regions by migratory birds or wind-blow mosquitoes does not seem to play a large role in the epidemiology of GIII. Rather, GIII is most likely maintained year-to-year by hibernating mosquitoes, vertical transmission in mosquitoes, poikilothermic vertebrates and/or bats. In support of this hypothesis, JEV has been isolated from overwintering Culex sp. in temperate Asia [68], virus transmission by Culex spp. was shown following experimental hibernation [69], and vertical transmission of JEV in Culex spp. and Armigeres sp. has been experimentally demonstrated [70], [71]. Furthermore, antibody to JEV has been detected in several poikilothermic vertebrates [71], [72], [73], [74], experimentally JEV-infected lizards were able to maintain the virus throughout the winter [72], and experimental transmission of JEV has been shown from infected mosquitoes to uninfected lizards and from infected lizards to mice through mosquitoes [72]. In temperate Asia, JEV has been isolated from several bat species [73], [74] and JEV-infected bats subjected to experimental hibernation were able to maintain their viremias for over 100 days [75].
The MRCA of the three GV sequences was estimated to have existed in the early 20th century. Although the node states were largely influenced by the small number of GV sequences (n = 3), the mass of posterior probability supporting the location of GV evolution corresponds to tropical Asia, specifically Malaysia. In Malaysia, JEV was first described in the 1940s when an outbreak occurred during the Second World War among British prisoners of war [76]. It is possible that GV may have circulated undetected in tropical Asia for much longer, causing only sporadic cases of encephalitis that were mistaken for cerebral malaria or other encephalitic diseases.
Surprisingly, after almost 60 years of undetected virus circulation, a pool of Cx. tritaeniorhynchus collected in the Tibetan Province of China in 2009 yielded the GV XZ0934 isolate [31] and a pool of Culex bitaeniorhynchus collected in South Korea in 2010 yielded the GV 10-1827 isolate [32]. The MRCA of the XZ0934 and 10-1827 isolates was estimated to have occurred sometime within the last 27 years in temperate Asia. Interestingly, despite surveillance neither JEV nor Cx. tritaeniorhynchus had been detected in Tibet prior to 2009 [77], suggesting that GV of JEV may have entered Tibet shortly before it was initially isolated in 2009. It is possible that GV arrived in Tibet via JEV-infected migratory birds or perhaps by wind-blown mosquitoes.
The three GV viruses shared 36 genotype-defining nonsynonymous substitutions within the E protein, 13 of which were non-conservative. This is consistent with the Muar strain's distinct serological classification based on its reactivity with a set of monoclonal antibodies [78]. None of the GV isolates have been characterized using polyclonal antibodies derived from other members of the JEV serocomplex. Such studies may provide interesting information regarding the antigenic relationships between this ancestral JEV genotype and other closely related viruses, such as Murray Valley encephalitis virus and Usutu virus.
We estimated that GII evolved in tropical Asia around the early 20th century. The Bennett isolate, made in Korea circa 1951, represents the only example of a GII virus collected outside of tropical Asia [40]. As extensive surveillance for JEV has been performed in temperate Asia, this single genotype isolation event likely coincided with a GII epidemic focus that quickly died off [40]. Therefore, as predicted, statistical analyses demonstrated that GII included significantly more isolates sampled from temperate regions than expected.
It was estimated that GI, GI-a and GI-b all emerged in tropical Asia around the mid 20th century. Statistical analyses demonstrated that GI-a included significantly more isolates sampled from tropical regions, and GI-b included significantly more isolates sampled from temperate regions, than expected under the null hypothesis of no association between climate and genotype. Like GIII, GI-b may be maintained in temperate Asia throughout the winter months in hibernating mosquitoes, vertical transmission in mosquitoes, poikilothermic vertebrates, and/or bats. This suggests that the spread and establishment of GI-b throughout Asia may have been due to its ability to efficiently overwinter in temperate Asia.
The phylogenetic divergence of GI is defined by a non-conservative threonine to methionine substitution at site 129 of the E protein. This GI-defining substitution was also found to be under directional selection. Although substitutions at this site within the E protein of JEV have not yet been associated with phenotypic alterations, this substitution alone or in combination with substitutions in other regions of the genome may have provided a phenotypic advantage to GI viruses that led to the spread and establishment of this genotype throughout Asia.
The MRCA of the three GIV sequences was estimated to have existed in the late 20th century. GIV includes seven isolates (only three of these include E gene sequence information) collected from mosquitoes only on three islands encompassing the Indonesian archipelago between 1980 and 1981. The reasons why GIV appears to be confined to Indonesia are unknown, but could be due to a number of reasons. For example, there could be a narrow host/vector range for GIV, the vector competence of Cx. tritaeniorhynchus for GIV may be low, the primary vector of GIV may be a mosquito that is confined to Indonesia, the replicative ability of GIV in birds may be low, and/or the GIV transmission cycle may involve a non-migratory amplifying host [42].
Although valuable information was obtained from our phylogeographic study, the data should be interpreted in light of its limitations. Isolations of JEV prior to the 1970s were primarily made from humans residing in China and Japan in response to epidemic transmission of the virus. After 1970, the majority of JEV isolates were made from mosquitoes and swine throughout many prefectures of Japan (due to rare cases of JEV following introduction of an effective vaccination program) and provinces of China as part of yearly routine surveillance for JEV. Therefore, it is unlikely that the dataset is largely biased by sequences sampled from local endemic foci that would have confounded the reported genotype-climate association. Due to the small number of ORF sequence data for JEV, we utilized a dataset consisting of sequence information derived from the E gene. While the E gene of JEV was found to be a good evolutionary proxy of the ORF, we were not able to assess whether diversifying and/or directional selection in other regions of the genome may have played a role in the adaptation of the viral genotypes to their respective environments. Finally, phenotypic characterization of the five genotypes of JEV has yet to be performed and would provide further evidence to support the genotype-climate association.
By applying Bayesian phylogeographic, categorical data analysis and phylogeny-trait association techniques to a large JEV sequence dataset we have demonstrated that GIII and GI-b are temperate genotypes maintained year-round in northern latitudes likely by either hibernating mosquitoes, vertical transmission in mosquitoes, poikilothermic vertebrates and/or bats. In contrast, GI-a and GII are tropical genotypes likely maintained via mosquito-avian and/or mosquito-swine transmission cycles. This suggests that the spread and establishment of GI-b throughout Asia may have been due to its ability to efficiently overwinter in temperate Asia. As highlighted by the recent emergence of West Nile virus into the western hemisphere [79] and Usutu virus into the European continent [80], the invasion of JEV into previously unoccupied regions is a real threat. Many areas of the world have JEV-competent vectors and waterbirds, and unlike West Nile and Usutu viruses, JEV also utilizes domestic swine as amplifying hosts, which can drive epidemics by producing an abundance of infected mosquitoes.
|
10.1371/journal.ppat.1005950 | An Epigenetic Compound Library Screen Identifies BET Inhibitors That Promote HSV-1 and -2 Replication by Bridging P-TEFb to Viral Gene Promoters through BRD4 | The human HSV-1 and -2 are common pathogens of human diseases. Both host and viral factors are involved in HSV lytic infection, although detailed mechanisms remain elusive. By screening a chemical library of epigenetic regulation, we identified bromodomain-containing protein 4 (BRD4) as a critical player in HSV infection. We show that treatment with pan BD domain inhibitor enhanced both HSV infection. Using JQ1 as a probe, we found that JQ1, a defined BD1 inhibitor, acts through BRD4 protein since knockdown of BRD4 expression ablated JQ1 effect on HSV infection. BRD4 regulates HSV replication through complex formation involving CDK9 and RNAP II; whereas, JQ1 promotes HSV-1 infection by allocating the complex to HSV gene promoters. Therefore, suppression of BRD4 expression or inhibition of CDK9 activity impeded HSV infection. Our data support a model that JQ1 enhances HSV infection by switching BRD4 to transcription regulation of viral gene expression from chromatin targeting since transient expression of BRD4 BD1 or BD1/2 domain had similar effect to that by JQ1 treatment. In addition to the identification that BRD4 is a modulator for JQ1 action on HSV infection, this study demonstrates BRD4 has an essential role in HSV infection.
| The human HSV-1 is associated with cold sore, while HSV-2 is considered a pathogen of sexually transmitted infection. Lytic infection by HSV-1 and HSV-2 triggers cellular responses as the virus strives to express its own genes and to replicate. To investigate host factors involved in the lytic infection cycle, we screened a chemical library of epigenetic regulation and identified several BET bromodomain inhibitors that enhanced both HSV-1 and HSV-2 infection. Using JQ1, a well-defined BRD4 inhibitor, as a model we showed that JQ1 increases HSV infection by allocating BRD4 to viral gene promoters. We also showed that BRD4 regulates HSV-1 and HSV-2 lytic infection by recruitment of factors for transcription elongation. The study expands the knowledge on viral replication regulation and identifies novel targets for antiviral agents.
| Herpes simplex virus-1 and -2 (HSV-1, HSV-2) are important pathogens of human diseases [1,2]. HSV-1 infection is mainly associated with cold sores and blisters, while HSV-2 is a major factor of sexually transmitted infections [3,4]. Patients acquire HSV-1 at relatively young ages, while initial HSV-2 infections occur mainly after puberty, often transmitted after intimate contact [5]. It has been estimated that two thirds of adult population aged 15–49 are infected with HSV-1, while over 550 million individuals aged 15–49 have genital infection with HSV-1 or HSV-2 [1,2].
HSV-1 and HSV-2 are double-stranded DNA viruses that are genetically similar and share many common features in infection and replication. The viruses are acquired initially by direct contact and replicate within mucosal epithelial cells. In the meantime, the virion can enter the nerve termini of sensory neurons and travel transgradely to the cell bodies and establish latency. Latent infection serves as a reservoir of virus for recurrent infection and transmission to other individuals. Although immeasurable advances have been made towards our understanding of HSV infection, the molecular machinery responsible for HSV replication regulation remains elusive and largely mystified.
Multiple viral and cellular factors are involved in HSV replication [6]. Upon HSV infection of epithelial cells, more than 80 viral genes are sequentially expressed in a temporal cascade, including the immediate early genes, early genes and late genes. Meanwhile, the HSV genome is rapidly incorporated into nucleosomes bearing histone modifications that resemble characteristics of heterochromatic structures [7,8]. Histone modifications have an essential role in HSV lytic and latent infections. For example, chemicals that inhibit histone deacetylase activity are reported to enhance viral replication [9,10]. Inhibition of the histone demethylase LSD1 blocks virus lytic replication and reactivation from latency [11,12]. Whether other factors of epigenetic regulation have a role in HSV infection is not well studied.
We took an approach by screening a chemical library of epigenetic regulation to identify factors affecting HSV infection. The library consists of well-defined inhibitors of HDAC, methyltransferase, the aurora kinase, among other categories. In addition to TSA, a known HDAC inhibitor that has been reported to enhance HSV-1 and HSV-2 infectivity, we discovered several structurally different BRD4 inhibitors that promoted HSV-1 and HSV-2 infection. BRD4 is a member of the bromodomain and extraterminal (BET) family, which includes BRD2, BRD3, BRD4 and BRDT in mammals. BRD4 is an epigenetic reader and recruits transcriptional regulatory complexes to acetylated chromatin and therefore participates in host gene regulation [13] and has multiple functions in HPV transcription activation and infection [14–17]. BRD4 interacts with HIV Tat protein to negatively regulate HIV-1 replication [18]. There has been no previous report on BRD4 participation in HSV infection. We therefore performed detailed studies and discovered that bromodomain inhibition enhances HSV infection by promoting transcription factor association with HSV gene promoters.
To investigate whether an epigenetic factor(s) regulates HSV infection, we screened an epigenetic compound library. We first determined the toxic effect against Vero cells, a host cell line for the initial screening, even though the concentrations for those compounds to cause cytotoxic effect in tumor cells are well documented. The effect on HSV-2 infection was then tested by pre-treatment of Vero cells using the maximal non-toxic concentrations. In those studies, we also included trichostatin A (TSA) and acyclovir as controls since those compounds are known to either enhance or inhibit HSV-2 infection. If a compound showed an effect that resembled viral infection, we then performed secondary infection assays to quantitatively measure their effect on virus infection. We identified several candidates from a library of 129 compounds that promoted the cytolytic effect associated with HSV-2 infection, while no compound from this library showed inhibitory effect in the infection assay. In addition to HDAC inhibitors like TSA that are known to enhance HSV-2 infection, we also identified several BET bromodomain inhibitors, including JQ1, I-BET-762, PFI-1, and TG101348, with enhancement effect on HSV-2 infection. The effect was confirmed by the detection of increased infectious virus production using both HSV-1 and HSV-2 (Fig 1A). Concomitant with increases in viral production, plaque sizes in those samples were obviously increased, compared to mock-treated controls (Fig 1B). On average, the plaque sizes in JQ1-treated samples were increased by over 250%, while those in TSA-treated samples by approximately 55–60%. The names, PubChem CID, and their effect on HSV-1 and HSV-2 infection are listed in Table 1.
A literature review indicated that the abovementioned BET compounds are developed as anticancer and anti-inflammatory agents. They compete selectively for acetylated lysine residues against BET proteins, particularly BRD4, by occupying the acetyl-lysine binding pocket of the BD1, thus inhibit BET protein binding to acetylated histones and disrupt the formation of the chromatin complexes essential for host gene expression [19–22]. To delineate a mechanism of bromodomain inhibition on HSV infection, we focused on JQ1 since the target of this compound is well defined [19] and the inactive enantiomer (-)-JQ1 is also commercially available. To preliminarily determine whether BD domain inhibition was responsible for enhanced HSV infection, Vero cells were treated with JQ1, RVX-208, a BD2-specific inhibitor of BRD4 [23], or with the inactive (-)-JQ1, prior to HSV-1 or HSV-2 infection. Treatment with 300 nM JQ1, but not RVX-208 or (-)-JQ1, resulted in increased production of HSV-1 and HSV-2 (Fig 2A). The effect of JQ1 on HSV-1 and HSV-2 infection was dose dependent (Fig 2B), a conclusion validated by increases in viral protein expression (Fig 2C). Additionally, we found that the effect was not restricted to Vero cells since treatment of HeLa, HEp-2, SK-N-SH, or primary mouse embryonic fibroblast (MEF) cells with JQ1 or with PFI-1 also increased HSV-1 and HSV-2 production (Fig A in S1 Text). Since both Vero and HeLa cells responded to JQ1 equally well, we then used HeLa cells for subsequent studies.
Next, we determined the time effect of JQ1 addition on HSV-1 and HSV-2 infection using HeLa cells. We found that addition of JQ1 at 2 hr prior to (-2 hr), or within the first 6 hr of inoculation had similar effect on HSV infection. The effect became less significant when JQ1 was added at 12 hr and 24 hr PI (Fig 2D). In contrast, treatment of HSV (Tx-V) with JQ1 prior to inoculation had little effect on HSV infection, indicating the compound did not target the virion for its enhancement effect.
We also measured the course of virus production in the presence or absence of JQ1 (Fig 2E). During the course of the infection, HSV-2 titers in JQ1-treated samples were higher compared to those in vehicle-treated controls. Although the highest titers in JQ1 treated and untreated groups were at comparable levels, the highest titer in JQ1-treated samples was reached by approximately 12 hr than that in the controls. The titers started to drop, probably by cell lysis due to the infection. These results indicated that BD1 domain inhibition promoted infectious virus production.
Among the bromodomain proteins, JQ1 has been shown to target the BD domains of BRD4 with high affinity [19]. Despite this strong selectivity, we would like to verify that JQ1 promoted HSV infection through BRD4. In this regard, we performed RNAi studies to determine whether suppression of BRD4 expression would have similar effect on HSV infection. HeLa cells were treated with 3 different siRNAs targeting BRD4 expression or with a scrambled control (siCtrl). siRNA treatment significantly suppressed BRD4 expression (Fig 3A). When tested for HSV infection using those cells, we found unexpectedly that suppression of BRD4 expression inhibited virus production as was determined by titration and immunoblotting studies (Fig 3A and 3B). The effect was specific since knockdown of BRD4, but not BRD2 or BRD3, suppressed HSV-1 and HSV-2 infection (Fig 3C, 3D and 3E). Importantly, we found knockdown of BRD4 expression also ablated the enhancement effect of JQ1 on HSV infection (Fig 3F). It is known that c-Myc is among the downstream genes regulated by BRD4 and JQ1 treatment selectively down regulated c-Myc expression [24,25]. Indeed, c-Myc expression was down regulated in JQ1-treated and BRD4 siRNA treated samples (Fig B in S1 Text), suppression of c-Myc by siRNA did not impact HSV infection. Therefore, the results clearly showed that BRD4 expression had an essential role in HSV infection and JQ1 exerted its enhancement effect on HSV-1 and HSV-2 infection through BRD4. It is therefore interesting to delineate a mechanism governing BD1 inhibition on HSV infection.
To firmly establish a role of BRD4 in HSV infection, we tested whether overexpression of BRD4 would lead to increases in HSV infection. To this end, the readily transfectable 293T cells were transfected with an empty vector or with a plasmid for BRD4 expression. Transient expression of full length BRD4 (BRD4 wt) did not increase HSV-1 production, even though the protein was expressed at high levels and localized correctly to the nuclei (Fig 4A, 4B and 4C). Since most cells we tested had high levels of endogenous BRD4 protein expression, we suspected that endogenous BRD4 was at sufficient levels in supporting HSV infection.
BRD4 is a reader protein and interacts with acetylated lysine residues using its BD domains. JQ1 binds to the BD domains and interrupts BRD4 interaction with acetyl lysine residues [19]. We reasoned that the BD domains might function similarly to JQ1 on HSV infection since BD domains may compete with BRD4 for acetyl lysine interactions. We therefore generated vectors for BD1, BD1/2 expression (Fig 4A). We also prepared BRD4 with BD1 deletion (ΔBD1) to potentially eliminate BRD4 chromatin targeting ability, and tested whether those mutants would support HSV infection. The truncated protein were mainly detected in the nuclei (Fig 4B) since they retain the putative nucleus localization signals. Similar to JQ1 treatment, BD1 or BD1/2 expression resulted in increased HSV infection as was demonstrated by immunoblotting and titration studies (Fig 4D and 4E). Impressively, BRD4 with BD1 deletion (ΔBD1) showed similar effect to JQ1 on HSV infection, suggesting the BD1-modulated events compete against BRD4 function on HSV infection. The results suggested to us that JQ1 promoted HSV infection by potentially diverting BRD4 function due to BD1 blocking.
In addition to participation in epigenetics regulation, BRD4 also regulates the transcription of cellular genes by recruiting of the positive transcription elongation factor P-TEFb [13,26,27] and transcriptional activators and repressors [28]. Next, we addressed whether JQ1 promoted HSV infection by allocating BRD4 complex more selectively to viral gene expression. We first studied whether HSV infection induced complex formation involving BRD4, P-TEFb, and RNAP II by immunoblotting for CDK9 and Rpb-1, subunits of P-TEFb and RNAP II, respectively. As shown in Fig 5A, a weak association between BRD4 with CDK9 and Rpb-1 was detected in the uninfected cells. HSV infection markedly induced BRD4 association with the two proteins (Fig 5A), an observation that was substantiated with immunostaining studies for protein co-localization (Fig 5B and 5C), indicating HSV infection promoted complex formation involving BRD4 and CDK9 or Rpb-1 proteins.
We then performed a modified ChIP assay to determine whether the complex was recruited to viral gene promoters. We used HSV-1 strain F for those studies since the genome sequence is readily accessible. There was selective association between BRD4 with viral gene promoters of IE, early and late genes that was demonstrated by PCR (Fig 5D) and quantitatively measured by qPCR (Fig 5E). The pattern of their association with BRD4 paralleled that with RNAP II (Fig 5F and 5G). The association was selective and specific, since primer pairs that anneal further upstream or within the coding region failed to detect viral DNA from the anti-BRD4 immunocomplexes (Fig C in S1 Text).
Next, we addressed whether JQ1 promoted protein association with viral gene promoters. Compared with HSV-1 infected controls, increased amount of CDK9 and Rpb-1 proteins was detected in the anti-BRD4 immunocomplexes from JQ1-treated samples (Fig 6A), an observation that was supported by immunofluorescence studies (Fig 6B). Concomitantly, there was increased recruitment of BRD4 to viral gene promoters, including ICP0, ICP4, ICP22, and gB genes in the JQ1-treated samples (Fig 6C and 6D). The results therefore indicated that JQ1 promoted protein complex formation required for HSV gene transcription.
To provide further evidence on the findings, we studied whether JQ1 effect on viral DNA synthesis by performing EdU incorporation study (Fig 6E). EdU incorporation remained scarce in serum starved cells, while HSV-1 infection promoted EdU incorporation. The staining became more intense and profound in JQ-treated samples. More importantly, BRD4 staining showed strong co-localization of viral DNA synthesis with BRD4 staining in those samples, indicating that JQ1 reallocated BRD4 and, likely, the transcriptional complex to viral gene transcription.
BRD4 plays a diverse role in modulating viral infection by RNA and DNA viruses. Chiefly, Brd4 regulates CDK9 and RNAP II phosphorylation during viral infection [29–32]. Inhibition of CDK9 has been shown to block DNA virus infection, including HSV-1 [33,34]. The role of BRD4 in HSV infection was therefore finalized by the demonstration of CDK9 in Rpb-1 phosphorylation using a specific antibody for phospho-Ser2/Ser5 of Rpb-1 CTD. HSV infection promoted Rpb-1 phosphorylation at 3 hr PI. Treatment with JQ1 caused more intense phosphorylation of Rpb-1 which correlated with increased HSV infection in those samples (Fig 7A). Treatment with LDC000067, a CDK9 specific inhibitor, dose dependently decreased Rpb-1 phosphorylation and HSV infection (Fig 7B), which was consistent with results from RNAi studies. Moreover, we found that depletion of CDK9 expression by RNAi obliterated HSV-induced Rpb-1 phosphorylation as well as HSV infection (Fig 7C). Importantly, both treatments also annulled the enhancement effect of JQ1 on HSV infection (Fig 7D and 7E). Those results together demonstrated that BRD4 regulates HSV infection by recruiting transcriptional factors for viral gene expression. BD1 inhibitors augment HSV infection by devoting those factors more efficiently to sites of viral gene transcription.
Human herpes simplex viruses are members of Herpesviridae that have the ability to have a productive infection cycle in epithelial cells. The viruses also have the ability to establish latency where they persist in the sensory neurons. Upon infection of an epithelial cell, the viral nucleocapsid is transported along the microtubules to the nuclear pore to release the viral genome into the nucleus. The linear viral DNA circularizes rapidly in the nucleus and undergoes chromatinization and modifications characteristic of euchromatin [7,8,35]. The modifications are related to viral gene transcription because inhibition of protein methylation reduced viral gene expression [11,36], while reagents that promote histone acetylation resulted in enhanced viral replication [37]. By screening an epigenetics compound library, we discovered several bromodomain inhibitors that promote HSV infection. Those compounds modulate chromatin acetylation process since they prevent chromatin readers from binding to acetylated chromatin and interfere with histone modification. Unlike HDACi that promote HSV replication by increasing histone acetylation, JQ1 treatment did not affect the pattern of histone modifications (Fig D in S1 Text). The effect was unrelated to c-Myc suppression by c-Myc (Fig B in S1 Text). Instead, our data support a model of BRD4 regulation of HSV replication (Fig 8). HSV infection induces BRD4 association with components of transcriptional regulation machinery for viral gene transcription. BD1 domain inhibitors like JQ1 dislodge BRD4 from cellular chromatin, therefore can facilitate this association, resulting in protein complex relocation to viral gene promoters for more effective replication.
Transcription elongation has been recognized as a rate-limiting step for the expression of signal-inducible genes. By recruitment of positive transcription elongation factor P-TEFb, the bromodomain-containing protein BRD4 plays a critical role in regulating transcription elongation of a vast array of genes for multiple cellular processes and responses [38]. Here we show that bromodomain inhibition results in enhanced association of BRD4 with viral gene promoters, underscoring the importance of BRD4 in HSV infection, which was consistent with previous observations since the virus has the ability to repress host transcription by diversion of the host Pol II transcription machinery to the viral genome [39,40]. The ubiquitously expressed BRD4 is an epigenetic reader and transcriptional regulator that bookmarks active genes during mitosis and serves as a scaffold for transcription factors [41–45]. We found that BRD4 recruits transcription factors including CDK9 to viral gene promoters, which may lead to optimization of HSV-1 gene expression [33,46,47]. BRD4 was shown to function as a cellular adaptor in retaining HPV genome during mitosis [16] and is essential for viral DNA replication [15]. BRD4 blocks the recruitment of transcriptional regulators to viral gene promoters for HPV and HIV-1 latency [43,48–50]. BRD4 depletion or inhibition increases HIV-1 replication [18]. JQ1 reactivates HIV from latency by antagonizing BRD4 inhibition of Tat-transactivation [18,32,51–53]. It will be interesting to determine whether JQ1 or BRD4 plays a role in HSV reactivation. Nonetheless, this study demonstrate BRD4 is dual functional protein in epigenetic regulation and viral replication.
It is known that hexamethylene bisacetamide (HMBA), a potent inducer of differentiation of tumor cells, is capable to enhance HSV replication [54–56]. The mechanism has yet to be defined. One peculiar feature of its purported mode of action on virus gene expression kinetics is that the effect is transient and requires a short exposure (1.5 to 5 h) to the agent early after infection using a VP16-null mutant [55,57]. Various signaling events, UV light, or chemicals such as HMBA cause transiently release of P-TEFb from its inhibitory apparatus, resulting in host gene transcription [58,59]. P-TEFb is sequestered by small nuclear ribonucleoprotein (7SK snRNP) and remains inactive in quiescent cells [60,61]. Free P-TEFb, which is required for acute response and cell growth, can be recruited to RNAP II via the super elongation complex (SEC), BRD4, or transcription factors. We show that JQ1 treatment enhances complex formation involving BRD4-P-TEFb-RNAP II. It is likely that HMBA and JQ1 share a common mechanism in promoting HSV replication.
The compounds that promote HSV infection were discovered by screening an epigenetics compound library. An obvious advantage of screening a compound library with defined targets is that it allows rapid but definitive dissection of biological events like HSV infection with ease since the compounds, structurally diverse, serve as probes to cross-validate each other. The identification of BRD4 as an essential factor in HSV lytic infection uncovers a novel target for antiviral drug research. BET inhibitors are currently developed for tumor therapy [62,63]. Our results raise questions that those agents may exacerbate existing herpes infection. With recent approval of an oncolytic herpes virus for the treatment of certain tumors, JQ1 and similar compounds nonetheless may have potential application in tumor therapy with oncolytic HSV vectors.
Vero cell line (African green monkey kidney epithelial cells, ATCC CCL-81), HeLa cell line (ATCC, CCL-2), HEp-2 cell line (ATCC, CCL-23) and human neuroblastoma SK-N-SH cells (ATCC, HTB-11) were purchased from ATCC (Manassas, VA) or from Cell Bank of Chinese Academy of Sciences (Shanghai, China). Mouse embryo fibroblast (MEF) cells (isolated from C57BL/6J mouse embryos) were prepared in the lab following published protocol [64]. The cells were cultured in DMEM (high glucose) supplemented with 10% heat-inactivated fetal bovine serum (FBS), sodium pyruvate and non-essential amino acids (Life Technologies, Carlsbad, CA) in a humidified incubator at 37°C with 5% CO2. HSV-1 strain F and HSV-2 strain G were propagated and titrated on Vero cells.
An epigenetics compound library (Catalog No. L1900) was purchased from Selleck Chemicals China (Shanghai). This library contains cell permeable inhibitors of epigenetic enzymes including histone deacetylases (HDACi), lysine demethylases, histone acetyltransferases (HATs), DNA methyltransferases (Dnmts), and the epigenetic reader domain inhibitors. Trichostatin A (TSA), PFI-1, JQ1, TG101348, RVX-208, and LDC000067 (CDK9 inhibitor) were also purchased from Selleck Chemicals. (-)-JQ1 was purchased from MedChem Express. Antibodies to BRD4 (Abcam, ab128874), CDK9 (Santa Cruz, sc-13130), Rpb-1 CTD (Cell Signaling Technology, 2629), phospho-Rpb1 CTD (Cell Signaling Technology, 13546, Ser2/Ser5 specific), c-Myc (Bioworld Technology, BS2462), histone H3 (Beyotime Institute of Biotechnology, AH433), phospho-histone H3 (Cell Signaling Technology, 3377, ser10), acetylated-histone H3 (Millipore, DAM1776434, Lys9/Lys14), GAPDH (Bioworld Technology, MB001) were obtained commercially. Antiserum to ICP4 (HSV-1 and -2; GenScript, Nanjing), and to ICP0 (HSV-1 and -2; ABmart, Shanghai) were prepared in New Zealand rabbits using commercial sources. Horse radish peroxidase (HRP)-conjugated secondary antibodies were purchased from Bio-Rad. Alexa Fluor 488 (Green) conjugated anti-mouse IgG and Alexa Fluor 568 (Red) conjugated anti-rabbit IgG were purchased from Life Technologies. Protein G agarose beads were purchased from Roche Diagnostics.
We first determined cytotoxic effect of the compounds on Vero cells using an MTT assay as we previously described [65]. To screen for their activity on HSV infection, a DMSO stock of a tested compound was freshly diluted into culture medium and tested in triplicate at the maximal non-toxic concentration at 2 hr prior to inoculation. HSV-2 at 1 MOI was used for the screening. The compounds or DMSO at 0.1–0.2% were left in the medium throughout the infection process. We measured cell viability at 48–60 hr PI using the MTT assay to quickly measure the cytopathic effect by HSV infection. A plaque forming assay (PFU) was then performed if a compound showed selective effect on HSV infection.
pCMV2-FLAG-BRD4 was purchased from Addgene (#22304). The plasmid encodes full length BRD4 (NM_058243) for mammalian expression [49]. The DNA corresponding to BD1 (aa 1–196), BD1/2 (aa 1–640), and BD1 deletion (ΔBD1, aa 197–1362) was inserted into p3xFLAG-CMV-24 (Sigma) using standard protocols of molecular biology. For transfection studies, the DNA were transfected using Lipifectamine-2000 (Life Technologies). The cells were used at 24 hr post transfection for different assays.
The oligos of small interfering RNA (siRNA) targeted human BRD4 or CDK9 were synthesized by GenePharma (Shanghai, China). For gene knockdown experiments, cells were plated in 6- or 24-well plates (Corning) 24 hr before transfection. Cells were transfected using Lipofectamine-2000. The cells were used at 48–72 hr after transfection for further experiments.
BRD4-siRNA #1: 5’-CUCCCUGAUUACUAUAAGATT-3’ and
5’-UCUUAUAGUAAUCAGGGAGTT-3’
BRD4-siRNA #2: 5’-GGAGAUGACAUAGUCUUAATT-3’ and
5’-UUAAGACUAUGUCAUCUCCTT-3’
BRD4-siRNA #3: 5’-GCACAAUCAAGUCUAAACUTT-3’ and
5’-AGUUUAGACUUGAUUGUGCTT-3’
BRD2-siRNA #1: 5’-CACGAAAGCUACAGGAUGU-3’ and
5'-ACAUCCUGUAGCUUUCGUG-3’
BRD2-siRNA #2: 5’-GGGCCGAGUUGUGCAUAUA-3’
5'-UAUAUGCACAACUCGGCCC-3’
BRD3-siRNA #1: 5’-AAUUGAACCUGCCGGAUUA-3’ and
5'-UAAUCCGGCAGGUUCAAUU-3’
BRD3-siRNA #2: 5’- CGGCUGAUGUUCUCGAAUU-3’ and
5'-AAUUCGAGAACAUCAGCCG-3’
c-Myc-siRNA #1: 5'-ACGGAACUCUUGUGCGUAA-3’ and
5'-TTACGCACAAGAGUUCCGU-3’
c-Myc-siRNA #2: 5’-GAACACACAACGUCUUGGA-3’ and
5'-UCCAAGACGUUGUGUGUUC-3’
CDK9-siRNA #1: 5’-GGAGAAUUUUACUGUGUUUdTdT-3’ and
5’-AAACACAGUAAAAUUCUCCdTdT-3’
CDK9-siRNA #2: 5’-CCGCUGCAAGGGUAGUAUAdTdT-3’ and
5’-UAUACUACCCUUGCAGCGGdTdT-3’
CDK9-siRNA #3: 5’-UAGGGACAUGAAGGCUGCUAAdTdT-3’ and
5’-UUAGCAGCCUUCAUGUCCCUAdTdT-3’
Cell lysates were collected by centrifugation after cell lysis using a buffer containing 50 mM Tris-HCl (pH 7.4), 150 mM NaCl, 1% NP-40, and a cocktail of protease inhibitors (Roche). For immunoprecipitation studies, cell lysates were incubated at 4°C for 2 hr with a capture antibody or a control antibody, followed by overnight incubation with protein G-agarose beads. The immunocomplexes were collected by centrifugation, then washed with ice-cold PBST (PBS-0.02% Tween-20), and separated by SDS-PAGE.
A modified ChIP assay was performed to determine protein and viral DNA interactions [66]. Briefly, HSV-1-infected cells (2x107) were washed with ice-cold PBS for three times and consequently cross-linked using 1% formaldehyde. The reaction was quenched by adding glycine (125 mM). After rinse with ice-cold PBS, the cells were collected into a cell lysis buffer containing 5 mM PIPES (pH 8.0), 1% SDS, 1 mM EDTA, and protease inhibitors. The mixture was incubated on ice for 15 min followed by sonication at an amplitude of 30 on a 15-sec on and 10-sec off cycle for 20 min to shear the DNA. The supernatants were collected by centrifugation and were used immediately or stored at -80°C. For immunoprecipitation, the supernatants were diluted using a ChIP dilution buffer (Beyotime) and precleared using salmon sperm DNA/protein G agarose beads (Roche). The protein-DNA complexes were precipitated with an indicated antibody and protein G beads at 4°C for overnight. The immune complexes were washed with a low salt buffer containing 150 mM NaCl in buffer A (2 mM EDTA, 1% Triton X-100, 0.1% SDS, 20 mM Tris-HCl, pH 8.1), followed by a high salt buffer (500 mM NaCl in buffer A), and then LiCl wash buffer (250 mM LiCl, 1 mM EDTA, 1% NP-40, 1% deoxycholate, 10 mM Tris-HCl, pH 8.1). The complexes were eluted with an elution buffer (1% SDS, 0.1 M NaHCO3). Contaminating RNA was removed by treating with 10 μg/ml RNase A. The complexes were then incubated at 65°C for overnight to reverse the cross-linking and then with proteinase K at 55°C for 1 hr to remove proteins. The DNA was purified by phenol chloroform extraction, followed by ethanol precipitation, and used for analysis of DNA-protein association using real-time and PCR. The primers are listed in S1 Table.
For RT-PCR studies, total RNA was extracted using TRIzol reagent (Life Technologies). One microgram RNA was reverse transcribed into cDNA using AMV reverse transcriptase. Real-time PCR was performed using SYBR Green PCR Master Mix (Q141-02/03, Vazyme, Nanjing, China) on an ABI 7300 real-time PCR system (Applied Biosystems) and data was analyzed using the 2-ΔΔCt method to obtain relative abundance [67]. The GAPDH Ct level was used as an internal control for value normalization.
Cells cultured on coverslips were fixed with 4% paraformaldehyde for 10 min at room temperature (RT). The cells were permeabilized at RT by treating with 0.2% Triton X-100 for 10 min. The cells were blocked with normal goat serum for 30 min and then stained by incubation with antibody to BRD4 (1:500 dilution), to Rpb-1 (1:700 dilution), or to CDK9 (1:500 dilution) at 4°C for overnight. Alexa Fluor conjugated secondary antibodies (1:1000) were used for visualization. Cell nuclei were stained with 1 μg/ml DAPI for 10 min. The images were collected on an Olympus FluoView FV10i confocal microscope.
We used EdU incorporation method to measure viral genome replication [28]. Briefly, HeLa cells grown on glass coverslips were serum-starved in DMEM containing 0.5% FBS for 24 hr to arrest cells at G0 stage. The cells were then infected with HSV-1 at 10 PFU/cell for 4 hr, at which time the medium was replaced with fresh DMEM containing 2% FBS and 5 μM EdU and cultured for another 4 hr. EdU-labeled DNA was conjugated by reaction with Alexa Fluor 488 azide using the Click-iT EdU imaging kit (Life Technologies). The nuclei were stained with DAPI. BRD4 redistribution was detected with confocal microscope.
Statistical analysis was performed using SPSS 17.0 software package. Data were analyzed by paired T test. P values equal to or less than 0.05 were considered statistically significant.
|
10.1371/journal.ppat.1006792 | Effect of analytical treatment interruption and reinitiation of antiretroviral therapy on HIV reservoirs and immunologic parameters in infected individuals | Therapeutic strategies aimed at achieving antiretroviral therapy (ART)-free HIV remission in infected individuals are under active investigation. Considering the vast majority of HIV-infected individuals experience plasma viral rebound upon cessation of therapy, clinical trials evaluating the efficacy of curative strategies would likely require inclusion of ART interruption. However, it is unclear what impact short-term analytical treatment interruption (ATI) and subsequent reinitiation of ART have on immunologic and virologic parameters of HIV-infected individuals. Here, we show a significant increase of HIV burden in the CD4+ T cells of infected individuals during ATI that was correlated with the level of plasma viral rebound. However, the size of the HIV reservoirs as well as immune parameters, including markers of exhaustion and activation, returned to pre-ATI levels 6–12 months after the study participants resumed ART. Of note, the proportions of near full-length, genome-intact and structurally defective HIV proviral DNA sequences were similar prior to ATI and following reinitiation of ART. In addition, there was no evidence of emergence of antiretroviral drug resistance mutations within intact HIV proviral DNA sequences following reinitiation of ART. These data demonstrate that short-term ATI does not necessarily lead to expansion of the persistent HIV reservoir nor irreparable damages to the immune system in the peripheral blood, warranting the inclusion of ATI in future clinical trials evaluating curative strategies.
| While we have made considerable advancements in the treatment of HIV, most infected individuals require life-long treatment to suppress plasma viremia, underscoring the need for the development of additional therapeutic strategies that would allow durable virologic remission in the absence of antiretroviral therapy (ART). While a definitive cure has not yet been identified, the field is moving in a promising direction, and with continued efforts we may arrive at a clinically acceptable alternative to ART. Clinical validation of new treatment options likely requires patients to stop therapy while monitoring for viral rebound, but the effect of treatment interruption and its precise impact on immunologic and virologic parameters in HIV-infected individuals has not been fully delineated. In this work, we measured a significant increase of HIV burden in the CD4+ T cells of infected individuals who underwent ATI with subsequent plasma viral rebound. However, the size of the HIV reservoirs as well as immune parameters returned to pre-ATI levels 6–12 months after the participants resumed ART. These data suggest ATI does not lead to expansion of the persistent HIV reservoir nor irreversible damages to the immune system in the peripheral blood.
| Sustained suppression of human immunodeficiency virus (HIV) and dramatic improvements in health outcomes have been achieved in infected individuals receiving antiretroviral therapy (ART) [1]. Nonetheless, the vast majority of HIV-infected individuals experience plasma viral rebound upon cessation of therapy [2], underscoring the need for developing additional therapeutic strategies that would allow durable virologic remission following the interruption of ART. Considerable efforts have been made in recent years to develop interventional approaches aimed at eliminating viral reservoirs and/or enhancing host immune responses against the virus in an effort to achieve durable suppression of HIV following discontinuation of ART [3]. In this regard, the effectiveness of such interventions has been typically evaluated ex vivo by measuring the impact on the size of persistent HIV reservoirs in CD4+ T cells of infected individuals [4]. However, these assays have proven to be inadequate for predicting whether a specific therapeutic intervention will lead to eradication of replication-competent virus or long-term suppression of HIV in the absence of ART [5–7]. Therefore, the incorporation of short-term analytical treatment interruption (ATI) into the clinical trial design has been employed to determine the efficacy of immune-based therapies in suppressing and/or eradicating HIV. Short-term ATI conducted under close virologic monitoring has been considered to be clinically safe [8]; however, its precise impact on immunologic and virologic parameters in HIV-infected individuals has not been well defined. We conducted the present study to address these issues.
We first investigated the effect of rebounding virus following ATI on the dynamics of HIV reservoirs using longitudinal specimens collected from 10 HIV-infected individuals who previously participated in a passive antibody transfer study (Table 1) [9]. Time points analyzed include: 1) prior to discontinuation of ART (referred to as “Pre-ATI”), 2) during ATI (referred to as “ATI”, the time point at which plasma viremia was at or closest to the highest level), and 3) following reinitiation of ART to suppress plasma viremia (referred to as “Post-ATI”). The median duration of the ATI phase was 57 days (range 22–115). All patients rebounded with observed peak viremia of 30,950 copies of HIV RNA per ml of plasma (median, mean 50,911, range 340–273,221). All study participants reinitiated ART under the pre-defined criteria of the ATI protocol and had been resumed on ART for a median of 363 days (range 140–418) post ATI at the time of analysis. Of note, one participant (N10) had received ART for 140 days when post-ATI analyses were conducted. Upon rebound, the level of total HIV DNA in the CD4+ T cell compartment increased significantly compared to that at the pre-ATI level (P = 0.002, Fig 1A–1C). There was a strong correlation between the level of plasma viremia and the percent increase of HIV DNA burden following cessation of ART (P = 0.005, Fig 1B). However, following reinitiation of ART, HIV DNA burden declined and subsequently reached pre-ATI levels (P = 0.002, Fig 1C), resulting in no statistically significant difference in HIV DNA between the pre-ATI and post-ATI time points. A similar pattern was observed when the level of cell-associated HIV RNA was measured prior to and during ATI as well as following reinitiation of ART (Fig 1D). There was a significant increase in the ratio of HIV-1 RNA to DNA during ATI compared to that at the pre-ATI level. However, following reinitiation of ART, the ratio declined and subsequently reached pre-ATI levels, resulting in no statistically significant difference in the ratio between the pre-ATI and post-ATI time points (Fig 1E). Given that the vast majority of infected CD4+ T cells carry replication-defective HIV, we then sought to determine the impact of ATI on the persistent viral reservoir carrying replication-competent HIV. As show in Fig 1F, there were no significant differences with respect to the number of CD4+ T cells carrying replication-competent HIV prior to ATI and following reinitiation of ART (P = 0.13). To obtain a more comprehensive view of the impact of ATI on the viral reservoir landscape, we used single-genome, near full-length viral, next-generation sequencing to assess the relative frequency of intact and defective HIV proviral sequences at the pre-ATI and post-ATI time points in subjects who underwent longitudinal leukapheresis (N01, N02, N04, N06, N08, N09, and N10). Analysis of 758 at pre-ATI (N01, n = 118; N02, n = 78; N04, n = 134; N06, n = 102; N08, n = 86; N09, n = 143; and N10 = 97) and 820 individual sequences at post-ATI (N01, n = 182; N02, n = 77; N04, n = 97; N06, n = 76; N08, n = 190; N09, n = 68; and N10, n = 130) revealed that the frequency of intact HIV proviral sequences was remarkably similar at the pre-ATI and post-ATI time points (mean of 40 vs. 33 intact proviral sequences per 106 CD4+ T cells, respectively), as was the relative proportion of intact proviruses within the total pool of analyzed HIV DNA sequences (7.4% vs. 6.4% of all sequences, respectively) (Fig 2A–2C). In addition, the contribution of viral sequences with defined structural defects to the total pool of reservoir sequences was not significantly different between the two time points, although there was a trend toward a reduced proportion of hypermutated, near full-length sequences following reinitiation of ART (Fig 2B and 2C). Of note, clusters of intact proviral sequences that were completely identical and likely derived from clonally-expanded HIV-infected CD4+ T cells [10–12] accounted for 69.6% and 62.3% of all intact sequences at the pre-ATI and post-ATI time points, respectively (Fig 2D and 2E), indicating that ATI had no substantial impact on the relative proportion of clonally-expanded CD4+ T cells harboring intact HIV-1 (P = 0.42, two-tailed Fisher’s exact test). Interestingly, an analysis of antiretroviral drug-induced viral sequence polymorphisms did not show any evidence for an accumulation of known drug escape mutations after ATI in intact proviruses (S1 Table). However, intact proviral sequences at the pre-ATI and post-ATI time points displayed significant differences in the amino acid composition at several residues within recognized VRC01 contact regions, specifically at viral envelope position 363, a site known to influence sensitivity to VRC01 [13] (S1 Fig); this is consistent with the previously-described evolution of VRC01-resistant viruses during this study [9]. Collectively, our data demonstrate that the increased level of cell-associated HIV in CD4+ T cells during short-term treatment interruption is transient and subsequent reinitiation of ART returns the viral burden to its pre-ATI level.
It is well established that active HIV replication leads to numerous immunologic abnormalities in infected individuals. In order to further investigate the effects of ATI followed by reinitiation of ART, we examined multiple immune parameters in the study participants. Although there was a transient decrease in CD4+ and CD8+ T cells during the ATI phase, there were no significant differences in the levels of CD4+ and CD8+ T cells comparing pre-ATI and post-ATI time points (Fig 3A and 3B). Additionally, there were no significant differences in the percentages of B (CD19+, Fig 3C), NK (CD16+56+, Fig 3D), or activated CD8+ T (CD38+HLA-DR+, Fig 3E) cells between pre-ATI and post-ATI time points. We also investigated the impact of ATI and reinitiation of ART on gene expression in highly purified CD4+ and CD8+ T cells (S2 Fig). A small number of genes (38 of >20,000 genes) were shown to be differentially expressed predominantly in the CD8+ T cells of the study participants prior to ATI and following reinitiation of ART only when unadjusted analyses were performed (S2 Fig and S2 Table). However, no significant changes in expression of genes associated with HIV life cycle and pathogenesis, interferon-regulated genes, or regulation of immune activation were found in the CD4+ and CD8+ T cells when measured prior to ATI and following reinitiation of ART (S2 Fig and S3 Table). In addition, analyses of longitudinal plasma samples revealed no change in the level of the vast majority of cytokines, chemokines, and inflammation markers between the two time points (S4 Table). Of note, a significant increase in the plasma level of RANTES following reinitiation of ART was observed (S3 Fig) although the significance of this finding requires further investigation. Collectively, our data suggest that short-term ATI and subsequent reinitiation of ART does not contribute to permanent immunologic or inflammatory abnormalities associated with the active HIV replication that results from ATI.
Finally, in order to further determine the functional status of T cells prior to and during the ATI as well as after reinitiation of ART, we longitudinally analyzed expression of exhaustion and activation markers on the CD8+ T cells of the study participants. We specifically focused on CD38+CD8+ T cells, a subset of CD8+ T cells that are over-represented in HIV-infected individuals during periods of active viral replication [14]. In addition, we measured the expression of immune exhaustion markers TIGIT and PD-1 on CD8+ T cells during the various phases of the study. These markers are expressed during chronic viral infection and are associated with decreased CD8+ T cell function [15]. During the short ATI phase, TIGIT expression, which is up-regulated on exhausted T cells in cancer as well as during chronic viral infection including HIV-infection [16–18], did not change significantly (Fig 4A–4C). The expression of PD-1 on CD8+ T cells increased in a majority of patients during the ATI phase, with the changes between the ATI phase and the post-ATI time point being statistically significant; however, the difference between pre-ATI and post-ATI was not significant (Fig 4A and 4C). Notably, the level of CD38 expression on CD8+ T cells elevated transiently during the ATI phase compared to that of the pre-ATI time point (P = 0.004) and subsequently returned to baseline during the post-ATI period (Fig 4A and 4C). Of interest, CD8+ T cells expressing a high level of CD38 (CD38hi) cells emerged in the majority of study participants during the ATI phase (Fig 4B and 4C). Further flow cytometric characterization revealed the majority of CD38hiCD8+ T cells co-expressed elevated levels of PD-1 and TIGIT during the ATI phase concomitant with plasma viral rebound that was significantly increased compared to the pre-ATI level (P = 0.03, Fig 4C). Following reinitiation of ART, the degree of CD38hiCD8+ T cells expressing PD-1 and TIGIT decreased significantly relative to that of the ATI phase (P = 0.002) and was not statistically different from the pre-ATI level (Fig 4C). Of note, the level of the above immune parameters on the CD4+ T cells of the study participants did not significantly change over time (S4 Fig). Taken together, these data indicate that active HIV replication/rebounding plasma viremia contributes to transient expansion of a dysfunctional subset of CD8+ T cells during the ATI phase that return to baseline after reinitiation of ART.
A major focus of current HIV research is the development of therapeutic strategies to achieve sustained virologic remission following discontinuation of ART either by eradication of viral reservoirs or enhancement of host immunity against HIV [19]. These efforts towards durable viral suppression are being driven by the fact that HIV reservoirs persist despite clinically effective treatment [20–22] and subsequent rebound occurs in virtually all infected individuals upon discontinuation of therapy [2, 23]. The efficacy of such strategies is typically assessed by laboratory-based assays ex vivo; however, the majority of such assays lack physiologic relevance and are unable to predict the likelihood of clinical outcome. Although the ultimate evaluation of the efficacy of a therapeutic agent in achieving sustained viral suppression would require discontinuation of ART and monitoring of plasma viremia, treatment interruption studies, especially those with repeated and prolonged cycles of ATI and infrequent monitoring, have shown adverse immunologic and virologic consequences. Despite this past experience, it remains unclear whether short-term ATI accompanied by frequent monitoring and strict ART restart criteria would have similar consequences in infected patients [8]. A thorough evaluation of this issue could potentially have a major impact on the future design of therapeutic studies in HIV-infected individuals. In the current study, we have demonstrated that short-term ATI causes transient expansion of the HIV reservoir in CD4+ T cells; however, the frequency of infected cells, including those carrying replication-competent HIV, returned to the pre-ATI level after reinitiation of ART. Notably, the relative frequency of near full-length viral sequences classified as genome-intact and the relative proportions of proviral sequences with defined structural defects were similar prior to ATI and following reinitiation of ART although it may require inclusion of larger copy numbers of genome-intact full-length HIV sequences obtained from a large number of participants who are not receiving interventional agents in order to validate our findings. Of note, ATI was not associated with an accumulation of intact proviral sequences that encode for antiretroviral drug resistance mutations that could potentially compromise treatment responses upon reinstitution of ART. It is not clear whether the diminution of the size of HIV reservoir following reinitiation of ART was due to decay of labile, infected cells (i.e., those with unintegrated HIV DNA) or due to death of productively infected cells. Future experiments involving a larger cohort of study participants and extensive phylogenetic analyses of subsets of CD4+ T cells could address this issue. Our data also demonstrate that short-term ATI was not associated with irreversible immune system abnormalities, such as a decrease in the level of CD4+ T cells or an increase in the markers of immune exhaustion and activation on CD8+ or CD4+ T cells. Of note, ATI/rebound was associated with emergence of CD38+, particularly CD38hi, CD8+ T cells. A significant proportion of these CD38hi CD8+ T cells co-expressed TIGIT and PD-1, markers expressed on activated and exhausted cells, during the ATI phase. However, similar to what was observed with the transient increase in the HIV reservoir associated with ATI, the level of expression of these markers normalized to the pre-ATI levels following reinitiation of ART. With recent studies demonstrating that co-blockade of the TIGIT and PD-1/PD-L1 pathway is a potential target for immune restoration in HIV-infected participants [16], the blunting of emergent CD38hiCD8+ T cells co-expressing TIGIT and PD-1 could potentially contribute to better virologic suppression in the absence of ART. Finally, the expression of TIGIT and PD-1 on CD4+ T cells did not change significantly before, during, or after the ATI phase as shown in the supplemental data. One of the potential confounding factors of our study is that the participants received VRC01 prior to and following discontinuation of ART that may have influenced virologic and immunologic outcomes [24]. Although our data suggest VRC01 was unable to neutralize all infectious viral isolates examined, it is plausible that it may have exhibited partial antiviral effects as evidenced by emergence of antibody-resistant virus. However, it is unlikely that VRC01 had any significant impact on the parameters examined in the present study for the following reasons: 1) all study participants experienced substantial levels of plasma viral rebound (median 30,950 copies of HIV RNA/ml), comparable to that seen in a previous study [25], following discontinuation of ART accompanied by a statistically significant increase in the frequency of CD4+ T cells carrying both proviral HIV DNA as well as cell-associated HIV RNA that ultimately returned to baseline levels following reinitiation of ART; 2) the majority of the study participants were found to carry VRC01-resistant replication-competent HIV prior to administration of the antibody [9]; 3) following administration of VRC01, antibody-resistant HIV emerged in the majority of the study participants during the course of ATI, which led to seeding of the CD4+ T cell compartment with replication-competent viruses that were unable to be neutralized by VRC01 ex vivo [9] and exhibited features of sequence diversification at VRC01 contact residues; 4) CD8+ T cells that displayed dysfunctional and exhausted characteristics typically associated with active HIV replication appeared during the ATI period; and 5) the last time points measured occurred one year after reinitiation of ART, well past the half-life of the VRC01 antibody, making it highly unlikely to be a source of virologic and immunological influence. Further investigations addressing longitudinal examination of tissue compartments and meta-analytic explorations may be necessary to formally evaluate the impact of ATI and re-initiation of ART on the size of the persistent HIV reservoir as well as immune parameters in a larger cohort of study participants who are not receiving other investigational anti-HIV agents (i.e., broadly neutralizing HIV-specific antibodies). Nonetheless, our findings support the use of antiretroviral treatment interruption in the setting of close monitoring of plasma viremia and concomitant strict ART restart guidelines as an integral part of determining the in vivo efficacy of therapeutic strategies aimed at achieving sustained ART-free virologic suppression in HIV-infected individuals.
Blood and leukapheresed products were collected from the study participants in accordance with clinical protocols approved by the Institutional Review Boards of the National Institute of Allergy and Infectious Diseases at the National Institutes of Health. All study participants were adults and provided written informed consent. All samples were anonymized.
Ten HIV-infected individuals from the NIH cohort (Participants N01-N10) who previously participated in a passive antibody transfer study (VRC01) were studied (Table 1). Participants received infusions of VRC01 (40mg/kg) intravenously three days before and 14 and 28 days after the discontinuation of ART followed by monthly thereafter for up to 6 months. Participants who met any of the following criteria discontinued VRC01 infusions and resumed ART: a decrease of more than 30% in the baseline CD4 T-cell count or an absolute CD4 T-cell count below 350 cells per cubic millimeter, a sustained (≥2 weeks) HIV plasma viremia greater than 1,000 copies per milliliter, any HIV-related symptoms, or pregnancy [9].
In order to measure the frequency of cells carrying HIV DNA, genomic DNA was isolated from purified CD4+ T cells using the Puregene DNA Extraction kit (Qiagen). The extracted DNA was digested with MscI (New England BioLabs) and subjected to droplet digital PCR (Bio-Rad Laboratories) according to the manufacturer’s specifications. The PCR reaction was carried out using HIV-specific and housekeeping gene RPP30-specific primers and probes in triplicate. The following primers were used for amplification of HIV LTR: 5’- GRAACCCACTGCTTAAGCCTCAA -3’ (5’ primer) and 5’- TGTTCGGGCGCCACTGCTAGAGA -3’ (3’ primer) along with the fluorescent probe 5’-6FAM-AGTAGTGTGTGCCCGTCTGTT-IABkFQ-3’. The following primers were used for amplification of RPP30: 5’-GATTTGGACCTGCGAGCG-3’ (5’ primer) and 5’-GCGGCTGTCTCCACAAGT-3’ (3’ primer) along with the fluorescent probe 5’-HEX-TTCTGACCTGAAGGCTCTGCGC-IABkFQ-3’.
In order to determine the level of cell-associated HIV RNA, RT-PCR was carried out using RNA isolated from purified CD4+ T cells (RNeasy Mini kit, Qiagen). Subsequently cDNA was synthesized from 2μg of RNA using qScript XLT cDNA Master Mix (Quanta Biosciences) using the following incubation steps: 5 minutes at 35°C, 60 minutes at 42°C, and 5 minutes at 85°C. cDNA was then subjected to droplet digital PCR (Bio-Rad) using HIV-specific and housekeeping gene, TATA box binding protein (TBP)-specific primers and probes in triplicate. The following primers were used for amplification of HIV 5’- TCTCTAGCAGTGGCGCCCGAACA -3’ (5’ primer) and 5’- TCTCCTTCTAGCCTCCGCTAGTC -3’ (3’ primer) along with the fluorescent probe 5’-6FAM- CAAGCCGAGTCCTGCGTCGAGAG -IABkFQ-3’. The following primers were used for amplification of TBP: 5’- CACGAACCACGGCACTGATT -3’ (5’ primer) and 5’- TTTTCTTGCTGCCAGTCTGGAC -3’ (3’ primer) along with the fluorescent probe 5’-HEX- TGTGCACAGGAGCCAAGAGTGAAGA/3-IABkFQ-3’. HIV RNA copy numbers were normalized per 1x106 copies of TBP.
In order to determine the level of total CD4+ T cells carrying replication-competent/infectious virus, serially diluted (1x106, 200,000, 40,000, 8,000, 1,600, 320) and replicates of 5x106 CD4+ T cells were subjected to quantitative co-culture assays. The cultures were then incubated with irradiated PBMCs (6–8 x106 cells per well) obtained from healthy sero-negative donors and anti-CD3 antibody. 1x106 CD8-depleted and anti-CD3 stimulated PBMCs from HIV-negative donors were added to each well the following day followed by periodic removal of cell suspensions and replenishment with fresh media containing IL-2. HIV p24 ELISA was conducted on the culture supernatants between days 14 and 21. The infectious units per million cells (IUPM) from quantitative co-culture assays were determined as described [26].
Genomic DNA was extracted from highly enriched CD4+ T cells using the QIAGEN DNeasy Blood & Tissue kit (Qiagen). HIV-1 gag DNA was then quantified using digital-droplet PCR (ddPCR). DNA diluted to single viral genome levels (<25% of wells being positive for HIV DNA PCR products) based on Poisson distribution statistics and ddPCR results was subjected to amplification using Platinum Taq DNA polymerase (ThermoFisher Scientific) and nested primers [12] spanning near full-length HIV-1 (HXB2 coordinates 638–9632). PCR products were visualized by agarose gel electrophoresis. Near full-length sequences (>8000 bp) were subjected to Illumina MiSeq sequencing with a median of approximately 2500 reads per base. Resulting short reads were de novo assembled and aligned to HXB2 using MUSCLE [27] to identify premature/lethal stop codons, internal inversions, or packaging signal deletions, using an automated in-house pipeline written in R scripting language that analyzes all open reading frames [28]. Presence/absence of APOBEC-3G/3F-associated hypermutations was determined using the Los Alamos HIV Sequence Database Hypermut 2.0 program [29]. Viral sequences that lacked all mutations listed above and had less than 15 base pairs deletions in the sequenced 5-LTR region were classified as “intact” [30]. If a near-full length sequence showed a mapped 5’ LTR deletion with an absent or incomplete primer binding site, but otherwise displayed no lethal sequence defects, the missing 5’ LTR sequence was inferred to be present, as described elsewhere [30] and the sequence was termed “inferred-intact”. Sequences identified as intact or inferred-intact by this algorithm were selected for manual verification of all open reading frames [12]. Phylogenetic distances between sequences were examined using ClustalX-generated neighbor joining algorithms [31]. The assay was validated by repeated (50 consecutive times) single genome amplification and sequencing of near full-length HIV DNA from the 8E5 cell line, which resulted in completely identical sequences in each case. Sequencing products with two different viral contigs, which would suggest amplification of two different viral DNA products in the same well, were discarded and not included in the analysis.
Peripheral blood mononuclear cells (PBMC) were isolated from peripheral blood and leukapheresis by Ficoll-Hypaque density gradient centrifugation and cryopreserved in liquid nitrogen. Cryopreserved PBMCs were thawed, washed, and stained with the following fluorophore-conjugated antibodies: CD3-APC-H7 (clone SK7, BD#560176), CD4-APC (clone SK3, BD#340443), CD38-BV421 (clone HIT2, BD#562444), TIGIT-PE (clone MBSA43, ebioscience#12–9500), PD-1- PE-Cy7 (clone eBioJ105, ebioscience#25–2799), LAG-3-PerCP-eF710 (clone 3DS223H, ebioscience#46–2239), TIM-3-FITC (clone F38-2E2, ebioscience#11–3109). Data were acquired on a BD FACSCanto II flow cytometer using the FACSDiva software (Becton Dickinson) and analyzed using Flow Jo version 10.1r5. Flow cytometry was repeated on specimens from select patients to ensure reproducibility.
RNA from CD4+ and CD8+ T cells was isolated using RNeasy mini kit (Qiagen) according to manufacturer’s specifications with on-column DNase I digestion to remove genomic DNA. 100 ng of total RNA was amplified and labeled using GeneChip 3’IVT PLUS Reagent Kit (Affymetrix), according to manufacturer’s instructions. 6μg of labeled cRNA was hybridized for 16h in 45°C to Affymetrix Human Genome U219 Array Strip, which contains more than 530,000 probes covering more than 36,000 transcripts and variants, which represent more than 20,000 genes mapped through UniGene or via RefSeq annotation. The arrays were washed and stained using GeneAtlas Hybridization, Wash, and Stain Kit for 3’IVT Arrays. CEL files retrieved from the GeneAtlas Software were normalized by Robust Multiarray Average (RMA), and further analyzed for differential gene expression by one-way analysis of variance (ANOVA), using Partek Genomics Suit 6.6.
Levels of 61 cytokines and chemokines (Human Cytokine Group I kit, Bio-Rad) and inflammation markers (Human Inflammation Panel I kit, Bio-Rad) were measured in undiluted plasma samples collected from the study participants prior to ATI and following reinitiation of ART according to the manufacturer’s instructions. Immunoassays were performed using the Bio-Plex 200 instrument (Bio-Rad) and analyzed with the Bio-Plex Manager software (Bio-Rad) using standard curves generated from the provided recombinant standards.
Three-way comparisons were performed using the Friedman test followed by pair-wise comparisons with the Wilcoxon signed rank test if significant. Correlations were determined by the Spearman rank method. The statistical tests used for each experiment are indicated in the figure legends. P < 0.05 was considered significant.
|
10.1371/journal.pgen.1006635 | Retrotransposon activation contributes to neurodegeneration in a Drosophila TDP-43 model of ALS | Amyotrophic lateral sclerosis (ALS) and frontotemporal lobar degeneration (FTLD) are two incurable neurodegenerative disorders that exist on a symptomological spectrum and share both genetic underpinnings and pathophysiological hallmarks. Functional abnormality of TAR DNA-binding protein 43 (TDP-43), an aggregation-prone RNA and DNA binding protein, is observed in the vast majority of both familial and sporadic ALS cases and in ~40% of FTLD cases, but the cascade of events leading to cell death are not understood. We have expressed human TDP-43 (hTDP-43) in Drosophila neurons and glia, a model that recapitulates many of the characteristics of TDP-43-linked human disease including protein aggregation pathology, locomotor impairment, and premature death. We report that such expression of hTDP-43 impairs small interfering RNA (siRNA) silencing, which is the major post-transcriptional mechanism of retrotransposable element (RTE) control in somatic tissue. This is accompanied by de-repression of a panel of both LINE and LTR families of RTEs, with somewhat different elements being active in response to hTDP-43 expression in glia versus neurons. hTDP-43 expression in glia causes an early and severe loss of control of a specific RTE, the endogenous retrovirus (ERV) gypsy. We demonstrate that gypsy causes the degenerative phenotypes in these flies because we are able to rescue the toxicity of glial hTDP-43 either by genetically blocking expression of this RTE or by pharmacologically inhibiting RTE reverse transcriptase activity. Moreover, we provide evidence that activation of DNA damage-mediated programmed cell death underlies both neuronal and glial hTDP-43 toxicity, consistent with RTE-mediated effects in both cell types. Our findings suggest a novel mechanism in which RTE activity contributes to neurodegeneration in TDP-43-mediated diseases such as ALS and FTLD.
| Functional abnormality of TAR DNA-binding protein 43 (TDP-43), an aggregation-prone RNA and DNA binding protein, is observed in the vast majority of both familial and sporadic ALS cases and in ~40% of FTLD cases, and mutations in TDP-43 are causal in a subset of familial ALS cases. Although cytoplasmic inclusions of this mostly nuclear protein are a hallmark of the disease, the cascade of events leading to cell death are not understood. We demonstrate that expression of human TDP-43 (hTDP-43) in Drosophila neurons or glial cells, which results in toxic cytoplasmic accumulation of TDP-43, causes broad expression of retrotransposons. In the case of glial hTDP-43 expression, the endogenous retrovirus (ERV) gypsy causally contributes to degeneration because inhibiting gypsy genetically or pharmacologically is sufficient to rescue the phenotypic effects. Moreover, we demonstrate that activation of DNA damage-mediated programmed cell death underlies hTDP-43 and gypsy mediated toxicity. Finally, we find that hTDP-43 pathology impairs small interfering RNA silencing, which is an essential system that normally protects the genome from RTEs. These findings suggest a novel mechanism in which a storm of retrotransposon activation drives neurodegeneration in TDP-43 mediated diseases such as ALS and FTLD.
| RTEs are “genomic parasites”–“selfish” genetic elements that are coded within our genomes and that replicate themselves via an RNA intermediate. After transcription, an RTE-encoded reverse transcriptase generates a cDNA copy, and this cDNA is inserted into a new genomic location at the site of double stranded DNA breaks created by an endonuclease activity encoded by the RTE [1]. Unrestrained RTE activity has been demonstrated to be highly destructive to genomes, resulting in large-scale deletions and genomic rearrangements, insertional mutations, and accumulation of DNA double strand breaks [2]. RTE-derived sequences constitute ~40% of the human genome, a quantity which encompasses a surprisingly large number of functional RTE copies. Although multiple interleaved, highly conserved gene silencing systems have evolved to protect the genome by blocking RTE expression, certain RTEs are nevertheless expressed in some somatic tissues [3, 4] and can replicate in a narrow window during neural development, leading to de novo genomic insertions in adult brain [5–12]. Moreover, a gradual deterioration of RTE suppression–and resultant increase in RTE activity–has been documented with advancing age in a variety of organisms and tissues [13–20], including the brain [21]. One or more of the gene silencing systems that normally block genotoxic RTE expression may therefore be weakened with age. We advance the novel hypothesis that a broad and morbid loss of control of RTEs contributes to the cumulative degeneration observed with TDP-43 protein aggregation pathology that is observed in a variety of neurodegenerative disorders, including ALS and FTLD, and that this loss of control of RTEs is the result of a negative impact of TDP-43 pathology on general RTE suppression mechanisms that are most prevalently relied upon in somatic tissue such as the brain.
TDP-43 is a member of the hnRNP family that homodimerizes to bind single stranded RNA and DNA with UG/TG-rich motifs [22]. This pleiotropic protein was originally identified as a transcriptional repressor that binds to the TAR element of the HIV-1 retrovirus to repress transcription [23]. TDP-43 is capable of shuttling back and forth from the nucleus to the cytoplasm but is predominantly found in the nucleus in healthy cells. In cells that are experiencing TDP-43 protein pathology, the protein accumulates in dense cytoplasmic inclusions that include full-length protein, caspase cleavage products and C-terminal fragments of TDP-43, as well as abnormally phosphorylated and ubiquitinated protein [24–26]. TDP-43 protein pathology is currently thought to involve toxicity incurred by cytoplasmic aggregates, interference with normal cytoplasmic function, depletion of normal nuclear TDP-43 stores, or some combination thereof [27].
Functional abnormality of TDP-43, an aggregation-prone RNA binding protein, is commonly observed in a spectrum of neurodegenerative diseases that spans motor neuron deterioration and progressive paralysis in ALS to dementia and cognitive decline in FTLD [28]. 90% of ALS cases and a large fraction of FTLD cases are considered to be genetically sporadic, in the sense that no known genetic lesion precipitates pathology. However, loss of nuclear TDP-43 and accumulation of TDP-43 immunoreactive cytoplasmic inclusions is observed in nearly all ALS and almost half of FTLD cases [28–30]. The mechanism that initiates the nucleation of TDP-43 protein pathology in apparently genetically normal individuals is not understood [29]. However, TDP-43 contains a low complexity domain in its C-terminal region, which is a common feature of RNA binding proteins that exhibit aggregation pathology in a variety of neurodegenerative disorders. Indeed a recent literature has established that cellular stress can induce such low complexity domain proteins, including TDP-43, to undergo a concentration dependent phase separation to form liquid droplets that over time can drive fibrilization [31–33]. TDP-43 protein also is known to accumulate in cytoplasmic stress granules in response to cellular stress [34]. Importantly, nuclear TDP-43 protein normally regulates splicing of TDP-43 mRNA, leading to nonsense mediated decay of its own message [35]. Thus the formation of cytoplasmic inclusions and clearance from the nuclear compartment that is observed in patients is also associated with loss of this feedback inhibition onto TDP-43 mRNA, leading to increased accumulation of cytoplasmic TDP-43 mRNA [36], which likely exacerbates formation of cytoplasmic inclusions.
Animal models of TDP-43 related disorders–and neurodegenerative disorders in general–have taken advantage of the concentration dependence of low complexity domain protein aggregation [37]. Most animal models of neurodegenerative diseases therefore have involved transgenic expression to increase protein concentration above endogenous levels, and reproduce many of the signatures of human disease, which in the case of TDP-43 includes aggregation of TDP-43 protein in cytoplasmic inclusions and downstream neurological defects [28, 38, 39]. Although such animal models are imperfect representations of what is largely a sporadically occurring disorder, they have enabled the delineation of a myriad of cellular roles for TDP-43 [28, 40] and have provided the means to uncover genetic interactions between TDP-43 and other genes that are implicated in neurodegenerative disorders [41–44]. TDP-43 pathology in animal models is now understood to cause global alterations in mRNA stability and splicing, de-repression of cryptic splicing, and biogenesis of some microRNAs (miRNAs) [28, 29, 38, 39, 45–47]. In principle, any of the cellular impacts of TDP-43 protein pathology could contribute to disease progression either alone or in combination. However, no clear consensus has yet emerged regarding the underlying causes of neurodegeneration in TDP-43 pathologies.
The RTE hypothesis investigated here is motivated by a series of prior observations. First, as mentioned above, LINE 1 RTEs are expressed in some somatic tissue [3, 4] and can actively replicate during normal brain development, leading to de novo genomic insertions in adult brain tissue [5–12], although the frequency of de novo insertions per cell is still hotly debated [48, 49]. Second, increased RTE activity occurs in the brain during aging [21]. Moreover, elevated expression of RTEs has been detected in a suite of neurodegenerative diseases [50–57] and reverse transcriptase biochemical activity of unknown origin has been shown to be present in both serum and cerebrospinal fluid (CSF) of HIV-negative ALS patients [58–61]. More recently, a specific RTE, the human ERV HERV-K, was found to be expressed in post-mortem cortical tissue of ALS patients and their blood relatives [50, 55, 58] and transgenic expression of the HERV-K Envelope (ENV) protein in mice is sufficient to cause motor neuron toxicity [55]. Finally, we have previously predicted via meta-analysis of RNA Immunoprecipitation (RIP) and Crosslinked RIP (CLIP) sequencing data that TDP-43 protein binds broadly to RTE–derived RNA transcripts in rodent and human brain tissue and that this binding is selectively lost in cortical tissue of FTLD patients [54]. However to date, no studies exist which address whether TDP-43 pathology causes endogenous RTEs to become expressed in vivo, no reports have probed the functional impact of TDP-43 pathology on the natural mechanisms of RTE suppression employed by somatic tissue such as the brain, and no studies have investigated the toxic effects of endogenous RTE activation on nervous system function.
The destructive capacity of RTEs has been extensively documented in many other biological contexts, including a wealth of seminal data from Drosophila melanogaster [62–64]. To test whether RTEs play a role in TDP-43 mediated neurodegeneration, we used an established Drosophila transgenic model which afforded the means to examine whether RTE activation causally contributes to TDP-43 mediated toxicity and cell death. We found that several hallmarks of TDP-43-induced degeneration are the result of activation of the gypsy ERV, an RTE that is structurally related to HERV-K, and that this activation leads to DNA damage mediated cell death. Moreover, we uncovered an inhibitory effect of TDP-43 expression on small interfering RNA (siRNA) mediated silencing, leading to broad activation of a panel of RTEs. These findings strongly suggest a broad impact of TDP-43 pathology on general RTE activity.
In order to determine whether RTEs contribute to TDP-43 pathological toxicity, we implemented an established animal model in which hTDP-43 is transgenically expressed in Drosophila. As with other animal models, including mouse, rat, fish, and C elegans, such expression reproduces many neuropathological hallmarks of human disease, likely via interference with endogenous protein(s) function [27, 38, 39, 65, 66]. In Drosophila, there is an endogenous putative ortholog of TDP-43, TBPH. Null mutations in TBPH in flies are lethal [67], as is the case with mammalian TDP-43. Hypomorphic loss of TBPH results in neurodevelopmental defects as with the mammalian gene. Overexpression-mediated toxicity has formed the basis of the preponderance of studies on TDP-43 in animal models, and has revealed much of what is known regarding TDP-43 protein function and dysfunction, leading to the dominant hypotheses regarding mechanisms of pathogenesis wherein toxic cytoplasmic aggregates are thought to contribute to disease progression [27–29, 39, 40, 66, 68]. To test the impact of expressing hTDP-43 on RTE expression, we first used RNA sequencing (RNA-seq) to profile transcript abundance. In patient tissue, TDP-43 protein pathology is observed in both neurons and glial cells [29] and an emerging literature has implicated glial cell toxicity in ALS [69–71]. Toxicity of TDP-43 in glia has similarly been documented in animal models, including in Drosophila [72–75]. We therefore examined the effects of transgenic hTDP-43 expression in the neuronal versus glial compartments of the brain.
We conducted paired-end Illumina based RNA-seq on head tissue of flies expressing either pan-neuronal (ELAV > hTDP-43) or pan-glial (Repo > hTDP-43) hTDP-43 compared with control flies that carried the hTDP-43 transgene alone with no Gal4 driver (hTDP-43 / +). We generated two independent sequencing libraries for each genotype from a population of animals that were 8–10 days post-eclosion. We generated a total of ~900 million reads, or about 150 million reads per sample (S1 Table), and conducted differential expression analysis (see methods). In order to identify effects both on gene transcripts and RTE transcripts (Fig 1A–1D; S2A and S3B Tables), we included reads that map to repetitive elements using an analysis pipeline that we have previously reported [54, 76]. Both glial (Repo > hTDP-43) and neuronal (ELAV > hTDP-43) expression of hTDP-43 caused differential expression of a number of cellular transcripts (Fig 1A and 1C; S2A and S3A Tables) and transposons, most of which were RTEs or Class I elements (Fig 1B and 1D; S2B and S3B Tables). In the case of differentially expressed genes, a broad spectrum of cellular processes were represented (see S2A and S3A Tables), with both increases and decreases in expression level seen for many genes. This is broadly consistent with previously reported transcriptome analysis using tissue from ALS patients [77]. In fact, the differentially expressed transcripts identified in our RNAseq experiments were significantly enriched for orthologs of genes that are implicated in ALS (ALS KEGG gene list; S1 Fig and S2C Table). In contrast with differentially expressed genes, when examining transposon transcripts the majority of those that were differentially expressed exhibited elevated levels in response to hTDP-43 expression. This was particularly striking for glial TDP-43 expression (Repo > hTDP-43; Fig 1D), where 23 of 29 differentially expressed transposons showed higher levels relative to controls. The majority of the differential effects were observed in RTEs (Class I elements), although a few Class II elements were also represented (Fig 1B and 1D).
It is also notable that while some RTE expression was elevated with both neuronal and glial hTDP-43 expression, there were several cases where effects were uniquely detected with only glial or only neuronal hTDP-43 expression. For example, the HeT-A LINE RTE and the mdg3, HMS-Beagle, gtwin and 3S18 LTR RTEs were elevated with either glial or neuronal expression of hTDP-43. However, the TART and TAHRE LINE RTEs and the Stalker2 and mdg1 LTR RTEs were only elevated in response to neuronal hTDP-43 expression, while a broad host of RTEs’ expression was elevated specifically in response to hTDP-43 expression in glia. Notable among these is the gypsy element, which we have previously demonstrated to be progressively de-repressed and even actively mobile with advanced age in brain tissue [21]. We cannot formally rule out the possibility that some of the differences between differentially expressed RTEs in Repo > hTDP-43 vs Elav > hTDP-43 may result from variation in copy number of specific TEs between the two Gal4 strains. But we think this is unlikely to be a major contributing factor because all of the strains were backcrossed to the same wild type strain for a minimum of 5 generations prior to the experiments. In the case of gypsy, expression levels are significantly increased in response to pan-glial hTDP-43 expression (Repo > hTDP-43) relative to controls (hTDP-43 / +) but no significant effect was observed with pan-neuronal expression of hTDP-43 (ELAV > hTDP-43) (Fig 1B, 1D, and 1E).
We selected the gypsy RTE as a candidate of interest to test the functional impact of loss of endogenous RTE suppression in response to hTDP-43 expression for several reasons. First, although gypsy was not the most abundantly expressed RTE in the RNA seq data, this element is known to be one of the most active natural transposons in Drosophila melanogaster, and is responsible for a high fraction of the spontaneous mutations that have been identified. Second, we have previously documented that gypsy is capable of replicating and generating de novo insertions in brain during advanced age [21]. Third, gypsy is an ERV with functional similarity to HERV-K, which is expressed in some ALS patients [50, 55]. And finally, because of intense prior investigation of the biology of this RTE, extant molecular genetic reagents provided the means to both perturb and detect gypsy function. We began by performing quantitative RT-PCR (qPCR) for both ORF2 (Pol) and ORF3 (ENV) of gypsy on head tissue of flies expressing either pan-neuronal (ELAV > hTDP-43) or pan-glial (Repo > hTDP-43) hTDP-43. Because disease risk is age dependent and symptoms in ALS patients are progressive, we also examined the compounding effects of age. At two relatively young ages (2–4 and 8–10 days post-eclosion) we observe a dramatic increase in expression of both ORFs (Fig 2A and 2B) of gypsy specifically in flies expressing hTDP-43 in glia. In contrast, flies expressing neuronal hTDP-43 experience a wave of gypsy expression at the population level that occurs much later in age (S2A Fig for ORF3; similar effects seen for ORF2) in a similar manner to genetic controls that do not express hTDP-43 (see also: [21]). These flies do not show a significant impact of hTDP-43 expression on gypsy transcript levels. This is entirely consistent with the RNA-seq analyses (Fig 1B and 1D), where gypsy expression was found to be increased in head tissue specifically in response to glial hTDP-43 expression, but not to expression of hTDP-43 in neurons. Importantly, different genomic copy number or basal levels of gypsy expression between the parental Elav-Gal4 and Repo-Gal4 lines are unlikely to underlie the separate effects that we observe on gypsy when driving hTDP-43 expression in either neurons or glia (S2A.5 and S2A.6 Fig). Taken together, the RNA-seq and qPCR experiments confirm that gypsy RTE RNA levels are significantly and precociously elevated in response to pan-glial hTDP-43 expression.
Whole mount immunolabeling of brains using a monoclonal antibody directed against the gypsy ENV glycoprotein [21, 78] likewise shows early (5–8 days post-eclosion) and acute accumulation of strongly immunoreactive puncta particularly in brains of flies expressing glial hTDP-43 (Fig 2C; for quantification see S2B Fig). These intense puncta are observed throughout the superficial regions, which contain the majority of cell somata, as well as in deeper neuropil (Fig 2C) and persist into older ages. In contrast, we do not observe neuronal hTDP-43 expression to cause elevated gypsy levels above that seen in wild type flies at any time point with either qPCR or immunolabeling (Fig 2C and S2A Fig). Given that effects of glial hTDP-43 expression on gypsy ENV immunoreactivity were so robust in 5–8 day old animals, we examined ENV at earlier time points. We found that in animals expressing hTDP-43 in glia, there is little detectable gypsy ENV protein expression at 0 days (immediately following eclosion). In brains from animals 3 days post eclosion, we observe regional puncta with a variable intensity and spatial location (S2C Fig) although this effect was difficult to quantify because of its variability.
We next examined the relative impact of glial and neuronal hTDP-43 expression on the physiological health of the animal. As previously documented [73–75], we see effects with either neuronal or glial expression. However we observe differing severity and time courses, with effects of glial expression being more acute than those observed with expression in neurons. Flies expressing hTDP-43 in neurons exhibit significant locomotor impairment at 1–5 days post-eclosion, and flies expressing glial hTDP-43 show more severe locomotor impairment at this same age. This effect is further exacerbated by 5–10 days post-eclosion; at which point the animals expressing hTDP-43 in glia are largely immobile (Fig 3A). As previously reported [68, 73, 74, 79–82], flies expressing neuronal hTDP-43 exhibit reduced lifespan in comparison to genetic controls. But flies expressing hTDP-43 in glia display an even more severely reduced lifespan (Fig 3B). We further observe rampant cell death as detected by terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL) in the brains of flies expressing hTDP-43 in glia as early as 5 days post-eclosion (Fig 3C). Similarly, with transmission electron microscopy (TEM; Fig 3D) we observe profuse proapoptotic nuclei in brains of 12 day-old flies expressing glial hTDP-43. In contrast, driving expression of hTDP-43 in neurons under the OK107-Gal4 driver, which provides high levels of expression in the well-defined and easily imaged population of central nervous system (CNS) neurons that constitute the mushroom body, results in little to no increase in TUNEL labeling (consistent with a previous report: [81]) even when the flies were aged to 30 days (S3A Fig). The relative expression of hTDP-43 under the two major Gal4 drivers we are using, Repo-Gal4 (glia) and ELAV-Gal4 (neurons), does not differ with age, suggesting that divergent age effects on expression level cannot account for the observed differences in toxicity and impact on physical health (S3D and S3E Fig; respectively). Furthermore, we do not observe any effect of hTDP-43 expression on levels of the endogenous fly ortholog, TBPH, regardless of cell type of expression (S3F Fig; S4 Table). Thus, the phenotypes that we observe are not caused by indirect effects on TBPH transcript abundance but instead derive from toxicity of the hTDP-43 transgene itself. The levels of expression of the hTDP-43 transgene relative to the endogenous TBPH gene also are similar to what has been reported in rodent models (S4 Table). As is true in other animal models and in human patients, we cannot readily distinguish whether the effects we observe are due to toxic gain of function, dominant interference with an endogenous protein, or some combination thereof. Importantly, however, we can detect a disease specific phosphorylated isoform of hTDP-43 (S3B Fig) as well as cytoplasmic accumulation and nuclear clearance of the protein (S3C Fig), implying that the human protein is being processed in the CNS of the fly as it is thought to be in the disease state in human tissue.
Our observation that a panel of RTEs are expressed in response to hTDP-43 transgene expression, along with the extensively documented toxic effects of loss of control of RTEs in other biological contexts [21, 62–64] and our observation that the gypsy RTE itself can actively replicate and generate de novo insertional mutations in brain tissue during aging [21], and in response to the hTDP-43 transgene, suggested the possibility that loss of gypsy silencing might in fact account for a portion of the physiological toxicity observed with hTDP-43 expression in glia. To test whether the Drosophila ERV gypsy causally contributes to the harmful effects of hTDP-43, we used a previously published inverted repeat (IR) “RNAi” construct [57] directed against gypsy ORF2 (gypsy(IR)) that is sufficient to reduce the expression of gypsy by approximately 50% in head tissue of 28-day old animals (S4A Fig). We found that co-expression of this gypsy(IR) substantially ameliorates the lifespan deficit induced by glial hTDP-43 expression (Fig 4A). This effect is not observed when a control IR construct is co-expressed with hTDP-43 in glial cells (Repo > hTDP-43 + GFP(IR); Fig 4B), and neither the gypsy(IR) nor the GFP(IR) constructs, when expressed alone under Repo-Gal4 (S4B Fig) or ELAV-Gal4 (S4C Fig) or when present without a Gal4 driver (S4D Fig), has such an effect on lifespan. Therefore activation of gypsy is responsible for a substantial portion of the toxicity that we observe when hTDP-43 is expressed in glia, which results in drastically premature death in these animals. In contrast, co-expression of gypsy(IR) does not rescue the lifespan deficit exhibited by animals expressing hTDP-43 in neurons (Fig 4C). This is in accordance with our observations from RNA-seq (Fig 1B and 1D), qPCR (Fig 2A and 2B; S2A.1–S2A.3 Fig) and immunolabeling (Fig 2C) that neuronal expression of hTDP-43 also does not elevate gypsy expression above wild type levels at any given time point over the course of lifespan. The glial specificity of gypsy(IR) lifespan rescue is consistent with our observation that gypsy expression is induced specifically when TDP-43 is expressed in glia, lending credence to the conclusion that gypsy is causally participating in the resulting degenerative phenotype We of course cannot rule out the possibility that the gypsy-RNAi construct may also impact gypsy family RTEs that share sequence homology to gypsy. The conclusion that RTEs contribute to TDP-43 toxicity is further supported by the mild but significant lifespan extension that we observe with pharmacological inhibition of the reverse transcriptase activity that is essential for all RTE replication (S5E, S5F, S5G and S5H Fig).
RTE replication involves reverse transcription, generation of chromosomal DNA breaks, and integration of the RTE cDNA copy. DNA damage is therefore associated with either abortive or successful attempts at RTE replication and such DNA damage is thought to be a major source of cellular toxicity caused by RTE activity because it activates Chk2 signaling, which leads to programmed cell death. To test whether the harmful effects of hTDP-43 are in fact mediated by DNA damage, we capitalized on the previously documented ability of mutations in Chk2 to mask the toxic effects of RTE-induced DNA damage [83, 84]. Importantly, mutations in Chk2 do not prevent accumulation of DNA damage; rather they prevent the signaling required for the cell to recognize that DNA damage has occurred and respond by committing to a programmed cell death pathway [85]. We therefore employed an IR construct directed against loki (loki(IR)), the Drosophila ortholog of chk2, which is sufficient to significantly reduce levels of endogenous loki mRNA in head tissue of 28-day old animals (S5A Fig). Remarkably, co-expression of loki(IR) with hTDP-43 is able to fully rescue the lifespan deficit caused by hTDP-43 expression in glia (Fig 5A) or neurons (Fig 5B). These findings support the conclusion that DNA damage makes a major contribution to the loss of lifespan induced by either neuronal or glial expression of hTDP-43. This conclusion is supported by our RNA-seq findings, in which we observe that in each case the expression of a panel of RTEs is activated. Although neuronal hTDP-43 expression does not impact the levels of the gypsy RTE specifically, several other RTEs exhibit elevated expression (see: Fig 1B). Importantly, this extension of lifespan is not seen with co-expression of a control IR construct (Repo > hTDP-43 + GFP(IR); Fig 4B), and neither the GFP(IR) nor loki(IR) constructs when expressed individually under Repo-Gal4 or ELAV-Gal4 or present without a Gal4 driver (S4B–S4D Fig), have such an effect on lifespan on their own. These data suggest that Loki/Chk2 activity makes a major contribution to the pathological toxicity of hTDP-43 that we observe with both glial and neuronal hTDP-43 expression.
The brains of flies expressing hTDP-43 in glia display rampant cell death, seen both with TUNEL staining (Fig 3C) and at the level of TEM (Fig 3D). To test whether the decision of cells to commit to a programmed cell death pathway in response to hTDP-43 expression is mediated by Loki (Chk2), we co-expressed the loki(IR) that was so effective in suppressing hTDP-43 toxicity in survival analyses (Repo > hTDP-43 + loki(IR)) and found that this was sufficient to abolish the dramatic accumulation of TUNEL-positive nuclei induced by glial expression of hTDP-43 (Fig 5C and 5D). Moreover, we found that the gypsy RTE contributes at least in part to the decision of cells to undergo programmed cell death in response to hTDP-43 expression in glia, as knocking down gypsy (Repo > hTDP-43 + gypsy(IR)) also significantly reduces the TUNEL labeling observed in the CNS of these animals (Fig 5C and 5D). These effects are specific to loki(IR) and gypsy(IR) as co-expression of an unrelated UAS-(IR) construct with hTDP-43 in glia (Repo > hTDP-43 + GFP(IR)) does not significantly alter the number of TUNEL positive cells compared to brains of flies expressing hTDP-43 alone under Repo-Gal4 (S5B Fig). Importantly, co-expression of the GFP(IR), loki(IR), and gypsy(IR) constructs with hTDP-43 under Repo-Gal4 also does not significantly reduce the expression of hTDP-43 (TARDBP; S5C Fig). Differences in the level of hTDP-43 expression between these experimental groups therefore cannot account for the phenotypic rescue observed with loki or gypsy knock down in either the survival or cell death assays. Taken together, these data support the conclusion that the cell death induced by hTDP-43 is mediated predominantly via Loki/Chk2 activity in response to DNA damage, and that this DNA damage is likely induced by RTE activity. For both the physiological toxicity and cell death induced by hTDP-43 expression in glial cells, this effect is in large part due to the activity of one particular RTE, the gypsy ERV. These observations are in agreement with the well-documented accumulation of DNA double strand breaks induced by unleashing RTEs [86], as well as reports that transgenic expression of the HERV-K ENV protein in mice results in loss of volume in the motor cortex and DNA damage [55]. While the impact of gypsy appears to be restricted to the case where hTDP-43 is expressed in glial cells, our RNA-seq data demonstrate that expression of hTDP-43 causes the induction of a panel of RTEs that normally would be silenced. Such results lead us to postulate that hTDP-43 pathology might be impacting the natural mechanisms by which RTEs in general are normally kept suppressed. We therefore designed a reporter assay to detect the effect of hTDP-43 expression on the siRNA system, which provides the primary silencing mechanism to keep RTEs in check in somatic tissues such as the brain.
The major post-transcriptional RTE silencing system available in somatic tissue such as the brain is the siRNA pathway [87–92]. siRNAs with sequence complementarity to RTEs have been detected in many species, including mammals [1, 88, 93], and RTE-siRNA levels have been demonstrated to affect RTE activity [1, 94–96]. Moreover, disruptions in the siRNA pathway result in increased TE transcript levels [21, 91, 97] as well as novel insertions in the genome [21, 98]. Indeed, we have previously shown that disruption of the major siRNA pathway effector Argonaute 2 (Ago2) leads to precocious gypsy expression in Drosophila head tissue and this is accompanied by rapid age-dependent neurophysiological decline [21]. We therefore engineered a genetically encoded sensor system to inform us as to whether hTDP-43 expression impairs the efficiency of Dicer-2 (Dcr-2)/Ago2-mediated siRNA silencing in the Drosophila nervous system in vivo.
Our reporter system relied on three components. We co-expressed a Dcr-2 processed IR construct directed against GFP (GFP(IR)) with a GFP transgenic reporter. By selecting an effective GFP(IR), we were able to generate substantial silencing of the GFP reporter (Fig 6A and 6B). To test the effects of hTDP-43 on siRNA mediated silencing, we then co-expressed our third component: either hTDP-43 or an unrelated control transgene (tdTomato). This tripartite system was expressed either in all glial cells using the Repo-Gal4 driver (Fig 6A) or in mushroom body neurons using the OK107-Gal4 driver (Fig 6B). Brains of young (2–4 day) and middle aged (10–12 days) flies were imaged using confocal microscopy. In the case of neuronal expression we were able to carry the experiment out to old age (45–47 days), but this was not possible with glial expression of hTDP-43 as it results in dramatic reduction in lifespan (see Fig 3B). What we observed was conspicuously reminiscent of hTDP-43’s impact on gypsy expression. Glial expression of hTDP-43 causes a marked reduction of siRNA silencing efficacy, resulting in easily detectable expression of the GFP reporter. Such expression is dramatic and significant in brains of 2–4 day old flies and persists out to 10–12 days of age (Fig 6A). Brains are obviously deteriorated by the 10–12 day time-point, which likely explains why GFP levels appear to drop off somewhat. Neuronal expression of hTDP-43 in the mushroom body has a similar but more slowly progressing effect on siRNA-mediated silencing of our GFP reporter, with a somewhat later onset (Fig 6B). Indeed, when we perform an analogous experiment using an endogenous reporter of siRNA mediated silencing in a separate structure we observe a similar effect. The GMR-Gal4 driver, which drives high levels of expression in the fly eye, was used to express an IR construct directed against the endogenous white+ pigment gene in place of GFP as a reporter (Fig 6C). As with mushroom body neurons in the CNS, expression of hTDP-43 in the eye causes a progressive de-repression of the silenced reporter. It is noteworthy that the erosion of siRNA efficacy caused by hTDP-43 expression in the eye manifests as clusters of red-pigmented cells, a phenotype which is evocative of the stochastic clusters of ENV immunoreactivity observed early in response to glial hTDP-43 expression (Fig 6C and S2C Fig). In contrast, simply turning on expression of white+ after development results in a uniform darkening of the eye with age (S6A and S6B Fig). Taken together, these findings demonstrate that hTDP-43 expression interferes with siRNA-mediated silencing in several tissue types, resulting in de-suppression of reporter expression. In neurons hTDP-43 expression causes age-dependent progressive erosion of siRNA efficacy, while glial expression of hTDP-43 results in more acute siRNA silencing impairment.
Although we have yet to identify which step of the siRNA pathway is disrupted by hTDP-43 expression, it is not simply due to loss of expression of Dcr-2 or Ago2, the two major effectors of siRNA-mediated silencing in Drosophila [90–93]. qPCR of whole head tissue demonstrated that hTDP-43 expression in both neurons and glial cells does not affect absolute expression levels of Dcr-2 (S6C.1 Fig) or Ago2 (S6C.2 Fig) at either 2–4 or 8–10 days post-eclosion, therefore down-regulation of these molecules is not responsible for the observed de-suppression of gypsy. In fact, in the case of genetic controls and flies expressing hTDP-43 in neurons, Dcr-2 and Ago2 levels actually increase with age beginning at 21–23 days post-eclosion and persisting into old age (40–42 days old), suggesting that down-regulation of Dcr-2 and Ago2 likewise cannot explain the later elevation of gypsy expression observed in these genotypes (S6D.1–S6D.3 Fig). On the other hand, small-RNA seq from head tissue of flies expressing hTDP-43 under the glial Repo-Gal4 driver reveals a relative reduction specifically in levels of antisense siRNAs among the subset that target RTEs whose expression is elevated in the RNAseq data (S5 Table; S7A and S7B Fig). This is suggestive of a defect in either biogenesis or stability of the siRNAs that target these RTEs. We favor a model (Fig 7) in which TDP-43 protein pathology interferes with siRNA biogenesis and/or function, resulting in deterioration of siRNA-mediated silencing accompanied by activation of RTE expression. The resulting increase in RTE expression may lead to accumulation of DNA damage resulting from RTE activity induced by TDP-43 pathology, in turn activating Loki/Chk2 signaling and leading to programmed cell death (Fig 7).
We previously reported bioinformatic predictions of a physical link between TDP-43 protein and RTE RNAs in rodent and in human cortical tissue [54]. Here we provide in vivo functional evidence in Drosophila that TDP-43 pathological toxicity is the result of RTE activity generally and, in glial cells, expression of the gypsy ERV specifically. This finding is parsimonious with reports of high levels of reverse transcriptase activity in serum and CSF of HIV-negative ALS patients and their blood relatives [59–61], and of accumulation of transcripts and protein of HERV-K, a human ERV of the gypsy family, in the CNS of ALS patients [50, 55]. It also is notable that accumulation of virus-like inclusions have been detected by electron microscopy in both neurons and glia of the frontal cortex of one ALS patient with extended prolongation of life via artificial lung ventilation [99]. Furthermore, our findings are complementary to those documenting progressive motor dysfunction in transgenic mice expressing HERV-K ENV protein, one of the three major open reading frames of this human RTE [55]. However, the findings reported here provide the first demonstration that an endogenous RTE causally contributes to physiological deterioration and cell death in TDP-43 protein pathology. Additionally, our findings indicate that reverse transcriptase enzymatic activity contributes to the toxicity of TDP-43 induced RTE expression, and that toxicity is largely mediated by DNA damage-induced cell death. Finally, we demonstrate that TDP-43 pathology leads to erosion of the post-transcriptional gene silencing mechanisms that are broadly responsible for RTE repression, which is accompanied by elevated expression of a panel of RTEs. These findings are in agreement with our previous observations that TDP-43 protein normally exhibits widespread interactions with RTE transcripts in rodent and human cortical tissue and that these interactions are selectively lost in cortical tissue of FTLD patients [54], as well as a report that knocking out the C. elegans ortholog of hTDP-43 results in broad accumulation of transposon-derived RNA transcripts and double stranded RNA [100].
TDP-43 has frequently been reported to co-localize with the major siRNA pathway components, DICER and Argonaute, in both cell culture and human patient tissue. Indeed such co-localization is commonly detected in stress granules (SGs), cytoplasmic foci for modulating mRNA translation that materialize in response to cellular stress [34, 47, 101–104]. SGs are observed in pathological ALS and FTLD patient tissue, and can be induced in neuronal cell culture via overexpression of mutant and wild-type hTDP-43 as well as two other ALS linked genes, SOD1 and FUS, suggesting that they may represent a common downstream mechanism of pathological progression [47]. And SG formation in response to cellular stressors, or overexpression of ALS-linked genes including hTDP-43, inhibits DICER processing of pre-miRNAs to mature miRNAs [47]. This signature is detectable in both sporadic and familial ALS spinal column motor neurons as a dramatic global reduction in mature miRNAs in comparison to control tissue [47]. These findings are in accordance with previous work by Kawahara and Mieda-Sato (2012), which showed that loss of hTDP-43 function itself inhibits cytoplasmic miRNA processing by DICER for at least a subset of miRNAs [45]. In mammals, the same DICER and Argonaute proteins process both miRNAs and siRNAs [105]. Therefore, the effects of SG formation and hTDP-43 manipulation on DICER function may affect siRNAs just as dramatically as miRNAs, however the effects of TDP-43 expression on siRNA function in mammals have as yet to be investigated. In contrast with the mammalian system, siRNAs and miRNAs in Drosophila are processed largely via distinct pathways–Dcr-1/Ago1 and Dcr-2/Ago2, respectively [105]. This disparity provided an opportunity for us to engineer an in vivo sensor to investigate the effects of TDP-43 on the siRNA system separate from its effects on miRNA biogenesis. While production of some miRNAs is disrupted in both animal models of ALS and human patient tissue, our data clearly demonstrate that in Drosophila, pathological TDP-43 expression disrupts the siRNA function of the DICER/Ago pathway. This finding dovetails with a report that the C. elegans TDP-43 ortholog impacts accumulation of double stranded RNA, which is the substrate of the DICER enzyme [100]. Our findings support the conclusion that the disruption of siRNA silencing contributes to cellular toxicity, dramatic physiological deterioration, and premature death via loss of control of RTEs.
Unregulated RTE expression is known to be highly toxic in other biological contexts for a number of reasons, including accumulation of toxic RNAs, creation of harmful mutations, and accumulation of DNA damage. In the case of gypsy we demonstrate that this RTE has a causal impact on cell death and physiological decline in the animal’s health. We also identify DNA damage-induced cell death, mediated by Chk2 activity, as a major contributing factor in the toxicity of TDP-43 both at the cellular and organismal level. Importantly, gypsy is not the only RTE whose expression we found to be increased in response to hTDP-43 expression. We in fact observe a panel of RTEs that exhibit elevated expression, with some variation in this profile when hTDP-43 is expressed solely in neurons or glia. This is consistent with the observation that knocking down gypsy expression only partially suppresses the toxicity of TDP-43, whereas blocking loki (chk2) expression leads to a near complete suppression of the effects of hTDP-43 expression on cell death and lifespan reduction. The involvement of DNA damage-induced cell death suggest that gypsy and likely other RTEs may be successfully or abortively inserting into genomic DNA, although we are mindful of the fact that increased levels of RTE proteins and RNAs may themselves be cytotoxic, as is observed with the Alu RTE in macular degeneration [52].
In the case of HERV-K, it has recently been shown [55] that TDP-43 binds directly to the LTR at the DNA level, thereby activating transcription of HERV-K. Our results establish, however, that TDP-43 pathology also compromises the siRNA-mediated gene silencing system, which is the major post-transcriptional genomic defense against RTEs in somatic tissues. The mechanisms by which TDP-43 protein pathology disrupts siRNA silencing remain to be investigated, but we favor the idea that it involves direct interactions between TDP-43 and the siRNA protein machinery [45, 47], and our previous findings also suggest direct interaction with RTE RNAs [54]. The disruptive impact of TDP-43 on the siRNA system points to a general loss of RTE silencing—as opposed to activation of a specific element such as the gypsy ERV (or HERV-K)—as the major contributing factor in hTDP-43-related neurophysiological deterioration. This conclusion is supported by our RNA-seq data which shows a broad and general increase in RTE expression in head tissue of Drosophila expressing either neuronal or glial hTDP-43, as well as a pronounced reduction specifically in antisense siRNAs which target the RTEs we observe to exhibit increased expression in response to hTDP-43 expression. In accordance with this notion, we have previously shown in Drosophila that mutation of Ago2, a major effector protein of the siRNA system, results in activation of several different RTEs in brain tissue and causes rapid age-related cognitive decline and shortened lifespan [21].
Like Drosophila, the human genome contains more than one type of functional RTE. In addition to HERV-K, the human genome contains on the order of 100 fully active copies of L1 RTEs, and a far higher number of non-autonomous elements that replicate in trans by capitalizing on the protein machinery encoded by L1s [106]. Moreover, abnormally high levels of expression of several different RTE families has been reported across a suite of neurodegenerative diseases [50–57], and there is accumulating evidence suggesting RTEs generally become active with advanced age in a variety of organisms and tissues [13–18], including the brain [21]. Previous studies make the case that this effect may result from age-related loss of transcription-level heterochromatic silencing [16, 18]. Our finding that TDP-43 erodes post-transcriptional, siRNA-mediated RTE silencing therefore raises an intriguing hypothesis regarding the synergy between age and TDP-43 pathology on RTE activation, particularly when the reinforcing action of siRNAs on heterochromatin is taken into consideration [1, 18]. This potential synergy, in conjunction with the replicative capacity encoded by RTEs, leads us to posit the “retrotransposon storm” hypothesis of neurodegeneration. We envision that loss of control of RTE expression and replication leads to a feed-forward mechanism in which massive levels of activity drive toxicity and degeneration in the nervous system. Our findings are not in conflict with a wealth of data that have implicated effects of TDP-43 pathology on splicing, RNA stability, translation, and miRNA biogenesis [28, 29, 38, 39, 45–47]., and it will be important to conceptually integrate our findings with these other aspects of TDP-43 pathology. But the direct impact we observe on cell death highlights the importance of investigating the contribution of siRNA dysfunction and RTE toxicity in TDP-43-mediated pathogenesis, and may indicate a promising common avenue for novel therapeutic targets in both familial and sporadic cases of ALS.
All transgenic fly stocks used, with the exception of w(IR) and GMR-Gal4, were backcrossed into our in-house wild type strain, the Canton-S derivative w1118 (isoCJ1) [107], for at least five generations to homogenize genetic background. The GFP, OK107-, ELAV-, and Repo-Gal4 lines [108], as well as the hTDP-43 [68] and gypsy(IR) [57] lines, are as reported previously. The GMR-Gal4, Gal80ts, w(IR), GFP(IR), and tdTomato lines were acquired from the Bloomington Drosophila Stock Center (stock numbers: 43675, 7019, 25785, 9331, and 32221; respectively), and the loki(IR) line was acquired from the Vienna Drosophila Resource Center [109] (stock number: v44980). Flies were cultured on standard fly food at room temperature unless otherwise noted.
All fly stocks used for lifespan analysis and longitudinal qPCR experiments were double dechorionated by bleach treatment in order to remove exogenous viral infection [21]. Briefly, 4-hour embryos were collected and treated with 100% bleach for 30 min to remove the chorion. Treated embryos were washed and subsequently transferred to a virus-free room equipped with ultraviolet lights to maintain sterility. This was repeated for at least two successive generations and expanded fly stocks were tested via qPCR of whole flies to ensure Drosophila C Virus (DCV) levels were below a threshold of 32 cycles.
Fly heads were collected for each genotype and total RNA was purified with Trizol (Invitrogen). RNA-Seq libraries were constructed using the NuGEN Ovation Drosophila RNA-Seq kit including DNase treatment with HL-dsDNase (ArcticZymes Cat. # 70800–201) and cDNA fragmentation using the Covaris E220 system according to manufacturer specifications. After amplification, library quality was measured using the Agilent Bioanalyzer system and quantity was determined using Life Technologies Qubit dsDNA HS Assay kit (for use with the Qubit 2.0 Fluorometer). Prior to sequencing, pooled libraries were quantified using the Illumina Library Quantification kit (with Universal qPCR mix) from Kapa Biosystems and a 1.2 pM loading concentration was used for PE101 on the Illumina NextSeq500 platform. Small RNAs were cloned using TruSeq (Illumina) approach with modifications described in Rozhkov 2015 [110]. Briefly, all small RNA libraries were constructed from 15–20 ug of TRIzol isolated total RNA. 18–29 nucleotide long small RNAs were size selected on 15% PAGE-UREA gel. After 3’- and 5’-adapter ligation, subsequent gel purification steps and reverse transcription cDNAs were PCR amplified with barcoded primers. PCR products were size selected on 6% PAGE gel, and quantified on Bioanalyzer. Loading concentrations were determined using the NEBNext Library Quant kit for Illumina (NEB). The libraries were sequenced on NEXTSeq platform.
RNA-seq libraries were run on an Illumina NextSeq (paired end 101). Reads were mapped to the Drosophila dm3 genome with STAR [111] allowing up to 4 mismatches and a maximum of 100 multiple alignments. To estimate the pileup along gypsy element, reads were mapped to the gypsy consensus sequence (GenBank accession: M12927) using Bowtie [112] with up to 2 mismatches. Reads were annotated based on genomic locations against ribosomal RNAs, transposable elements (FlyBase), and RefSeq genes (UCSC genome database RefSeq track). The GEO accession number for all RNAseq and smallRNAseq data is GSE85398 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE85398).
Reads mapped to ribosomal RNAs were removed from each library. For the remaining reads, expression abundance estimation and differential expression analysis were performed using the TEtranscripts package [76]. Reads for each library were normalized based on library size, e.g., reads per million mapped (RPM). Statistically significant differences were taken as those genes/transposons (TEs) with a Benjamini-Hochberg corrected P-value < 0.05, as calculated by DESeq [113]. Biological replicates were averaged for the purpose of estimating pileup along consensus TEs.
The ortholog pairs of fly and human were predicted with the Drosophila Integrated Ortholog Prediction Tool (DIOPT), which integrates the ortholog predictions from 11 existing tools [114]. The pairs with the “high” rank, defining as the best match with both forward and reverse searches and the DIOPT score is at least 2, were selected. Then we performed Fisher's exact test to check if the differential expression fly genes are enriched in the KEGG ALS human genes with identifiable fly orthologs.
Male flies were used for all lifespan assays since the majority of glial-expressing hTDP-43 flies that escape their pupal cases are male. Flies were housed 15 to a vial with a total of 75 flies per genotype and flipped into fresh food vials every other day. All vials were kept on their side in racks for the duration of the experiment. Lifespan experiments were performed blind.
Tenofovir disoproxil fumarate (TDF; Selleck Chemicals, CAS 147127-20-6), Zidovudine (AZT; Sigma-Aldrich, CAS 30516-87-1), and Stavudine (d4T; Sigma-Aldrich, CAS 3056-17-5) were prepped as solutions using dimethyl sulfide (DMSO) as solvent. Standard fly food was melted and cooled to a liquid, and NRTI solutions were added just before solidification to give final concentrations of 0 (vehicle alone control), 1, 5, 10, and 15 μM of each NRTI in a total volume of 0.2% DMSO, then stored at 4°C for a maximum of 2 weeks until use. Lifespan was monitored in all male flies with 20 flies per vial for a total of 100 flies monitored per genotype per treatment. Lifespans were performed as above.
To quantify and determine the consistency of feeding across treatments, Capillary Feeder (CAFE) Assays [115] were performed. Each replicate consisted of five male wild-type flies contained in a 1.5 mL microfuge tube chamber with 1% agarose in bottom to maintain humidity and two 5 μL disposable calibrated pipets (VWR, 53432–706) providing the media solution (5% sucrose / 5% autolyzed yeast extract, Sigma-Aldrich), inserted though the tube cap. Flies were acclimated with untreated solution ad libitum for 24 hours. Measurements were initiated one hour following the switch to experimental solutions and monitored for a total duration of 24 hours. Experimentally treated solutions were prepared to match the concentrations of each NRTI treatment used in solid food for lifespan analysis, as well as an additional solution of high concentration for each NRTI (100 μM) and a solution of vehicle alone control (0.2% DMSO). Displacement due to evaporation was controlled for by subtracting measurements from fly-less CAFE chambers with the vehicle solution set up in parallel to the experimental assays. CAFE assays with untreated solution were also set up in parallel and controlled for evaporation.
Locomotion behavior was assayed using the classic Benzer counter current apparatus as in Benzer, S., 1967 [116], with the following modifications: freshly eclosed flies were transferred into glass bottles with food and a paper substrate and plugged with foam stoppers. Flies were transferred to fresh bottles every 48 hours until they reached the appropriate age for locomotion assays. The Benzer assay was conducted in a horizontal position with a fluorescent light source to measure phototaxis. Locomotion assays were performed blind.
Tissue preparation, cDNA synthesis and qPCR were performed as previously described [21] using the Applied Biosystems StepOnePlus Real-Time PCR System. Heads of 75–100 flies were used for each biological replicate unless otherwise noted. All TaqMan Gene Expression Assays were acquired from Applied Biosystems and used the FAM Reporter and MGB Quencher. The inventoried assays used were: Act5C (assay ID Dm02361909_s1), Dcr-2 (assay ID Dm01821537_g1), Ago2 (assay ID Dm01805433_g1), TARDBP (assay ID Hs00606522_m1), TBPH (assay ID Dm01820179_g1), and loki (assay ID Dm01811114_g1). All custom TaqMan probes were designed following the vendor’s custom assay design service manual and have the following assay IDs and probe sequences: gypsy ORF2 (assay ID AI5106V; probe: 5’–AAGCATTTGTGTTTGATTTC-3’), gypsy ORF3 (assay ID AID1UHW; probe: 5’-CTCTAGGATAGGCAATTAA-3’), and DCV (assay ID AIPAC3F; probe: 5′-TTGTCGACGCAATTCTT-3′). For genomic QPCR, DNA was isolated from 2–4 day old flies (equal numbers per sample, 50% male/female). After RNAase treatment, equal input levels of DNA were used for each PCR reaction, and each reaction was performed in triplicate. The CT values for Gypsy ORF2 were normalized to those for Actin within each sample. The values shown S2A.5 Fig were then further normalized to the levels in the wild type strain in order to display relative fold change over our wild type strain.
Dissection, fixation, immunolabelling, and confocal imaging acquisition were executed as previously described [117]. The ENV primary antibody was used as described in Li, et al. 2013 [21, 78]. For TDP-43 immunohistochemistry, the primary full length human TDP-43 antibody (Protein Tech, 10782-2-AP) was used at a 1:100 dilution, and the primary pSer409 phosphorylated human TDP-43 antibody (Sigma Aldrich, SAB4200223) was used at a 1:500 dilution separately in conjunction with a 1:200 dilution of an Alexa Fluor 488-conjugated secondary antibody (Thermo Fisher Scientific, A-11070). Repo co-labeling was performed using a 1:200 dilution of primary antibody (Developmental Studies Hybridoma Bank, 8D12) and a 1:200 dilution of a Cy3-conjugated secondary antibody (Molecular Probes, A10521). DAPI co-staining was performed after a brief wash in 1x PBS immediately subsequent to secondary antibody staining using DAPI Dilactate (Thermo Fisher Scientific, D3571) as per manufacturer specifications. All brains co-stained with DAPI were imaged on a Zeiss LSM 780 confocal microscope using a UV laser and the Zeiss ZEN microscope software package.
The gain on the confocal microscope was set using the positive control (Repo > GFP or OK107 > GFP) and kept consistent across all subsequent brains imaged. The GFP signal of the median 10 optical sections of the appropriate structures (either the full brain for Repo or both lobes of the calyx for OK107, respectively) was calculated using ImageJ software, as previously described [118]. These ten values were then averaged, and this number used as a representation for each individual brain. 5–10 brains were analyzed per group.
For TUNEL staining, the In Situ Cell Death Detection Kit, TMR red (Roche, 12156792910) was used. The same dissection, fixation, and penetration and blocking protocol used for antibody staining was followed [117], at which point the brains were transferred to the reaction mix from the kit for 2 hours at 4°C followed by 1 hour at 37°C. Brains were then washed, mounted, and imaged as previously described [117]. For imaging, the gain on the confocal microscope was set using the positive control (Repo > hTDP-43) and kept consistent across all subsequent brains imaged. A projection image was generated using the middle 50 optical slices from the z-stack image of the whole brain. This projection image was then thresholded using the maximum entropy technique (See: [119]) via the Fiji plug-in for ImageJ software, and the subsequent binary image was subjected to puncta quantification using ImageJ software. Puncta quantification was thresholded for puncta greater than 3 pixels to reduce the likelihood of counting background signal. The total number of puncta counted was then used as a representation for the number of TUNEL-positive nuclei for each brain in subsequent statistical analysis. 7–12 brains were analyzed per group. Quantification of gypsy ENV immunoreactive puncta was performed in the same manner.
Flies of the appropriate age and genotype were placed at -70°C for 25 minutes and then kept on ice until immediately prior to imaging. Imaging was performed using a Nikon SMZ1500 stereoscopic microscope, Nikon DS-Vi1 camera and Nikon Digital Sight camera system, and the Nikon NIS-Elements BR3.2 64-bit imaging software package. The experiment was designed such that each group is balanced for the number of mini-white transgenes and heterozygous for genomic white+.
Drosophila heads were removed, the cuticle removed and the brains fixed overnight in 2% paraformaldehyde and 2% glutaraldehyde in 0.1 mol/L PBS. Samples were rinsed in distilled water and post-fixed for one hour in 1% osmium tetroxide in 1.5% potassium ferrocyanide in distilled water. Next, the samples were dehydrated in a graded series of ethanol and the final 100% ethanol was replaced with a solution of absolute dry acetone (Electron Microscopy Sciences, Hatfield PA). The samples were then infiltrated with agitation for one hour in an equal mixture of acetone and Epon-Araldite resin, followed by infiltration with agitation overnight in 100% resin. Samples were transferred to embedding capsules with the posterior head facing towards the bottom of the capsule and the resin was polymerized overnight in a vented 60°C oven. Thin sections were made from the mushroom body region and collected on Butvar coated 2mm x 1mm slot grids (EMS) and the sections were counterstained with lead citrate. Thin sections were imaged with a Hitachi H700 transmission electron microscope and recorded on Kodak 4480 negatives that were scanned with an Epson V750 Pro Scanner at 2400 DPI. 3 individual brains processed for Repo / +, 4 individual brains processed for Repo > hTDP-43; many images collected of each brain.
For qPCR data, the p-values of all data sets with only two groups were calculated using an unpaired t-test. Where an effect of age for more than two time points within one genotype was determined, a one-way ANOVA was performed, and where multiple ages and genotypes are represented a two-way ANOVA was performed; the results are reported in the figure legends. All pairwise comparisons for qPCR reported in the figures were corrected using the Bonferroni method for multiple comparisons. For both the locomotion data and the GFP quantification, p-values were reported using the Sheffé method; ANOVA results are reported in the figure legends. The CAFE assay is similarly reported, with pairwise comparisons between treatments being made using one-way ANOVA with Tukey’s Multiple Comparisons test. Survival analyses for the lifespan curves were performed using the Kaplan-Meier method, and the Gehan-Breslow-Wilcoxon test were used to compare survival curves. All pairwise comparisons for lifespan curves were corrected using the Bonferroni method. Sample sizes were selected based on standard practices in the literature. No randomization was employed in this study.
|
10.1371/journal.pcbi.1005936 | Static length changes of cochlear outer hair cells can tune low-frequency hearing | The cochlea not only transduces sound-induced vibration into neural spikes, it also amplifies weak sound to boost its detection. Actuators of this active process are sensory outer hair cells in the organ of Corti, whereas the inner hair cells transduce the resulting motion into electric signals that propagate via the auditory nerve to the brain. However, how the outer hair cells modulate the stimulus to the inner hair cells remains unclear. Here, we combine theoretical modeling and experimental measurements near the cochlear apex to study the way in which length changes of the outer hair cells deform the organ of Corti. We develop a geometry-based kinematic model of the apical organ of Corti that reproduces salient, yet counter-intuitive features of the organ’s motion. Our analysis further uncovers a mechanism by which a static length change of the outer hair cells can sensitively tune the signal transmitted to the sensory inner hair cells. When the outer hair cells are in an elongated state, stimulation of inner hair cells is largely inhibited, whereas outer hair cell contraction leads to a substantial enhancement of sound-evoked motion near the hair bundles. This novel mechanism for regulating the sensitivity of the hearing organ applies to the low frequencies that are most important for the perception of speech and music. We suggest that the proposed mechanism might underlie frequency discrimination at low auditory frequencies, as well as our ability to selectively attend auditory signals in noisy surroundings.
| Outer hair cells are highly specialized force producers inside the inner ear: they can change length when stimulated electrically. However, how exactly this electromotile effect contributes to the astonishing sensitivity and frequency selectivity of the inner ear has remained unclear. Here we show for the first time that static length changes of outer hair cells can sensitively regulate how much of a sound signal is passed on to the inner hair cells that forward the signal to the brain. Our analysis holds for the apical region of the inner ear that is responsible for detecting the low frequencies that matter most in speech and music. This shows a mechanisms for how frequency-selectivity can be achieved at low frequencies. It also opens a path for the efferent neural system to regulate hearing sensitivity.
| Our ability to hear is due to an intricate mechanotransduction process that takes place inside the inner ear. Sound-evoked waves on the basilar membrane, an elastic structure stretching along the cochlear canal, cause the deflection of mechanosensitive hair bundles of the sensory cells, thus gating ion channels in the cell membrane and producing electrical signals that are ultimately transmitted to the brain [1]. The transfer of basilar-membrane motion to deflection of the hair bundles is shaped by the structurally complex organ of Corti (Fig 1(A)), the outer hair cells of which can provide mechanical force [2]. Changes in transmembrane voltage cause these cells to change length, a phenomenon referred to as electromotility [3, 4]. Furthermore, the hair bundles of outer hair cells can also generate mechanical force [5, 6]. Both mechanisms may contribute to an active modulation of the sound-evoked motion of the organ of Corti [7–9].
The active mechanical feedback by outer hair cells is essential for the extraordinary sensitivity, tuning, and dynamic range of mammalian hearing organs, and damage to the outer hair cells consequently results in hearing loss [10–12]. Although this feedback presumably operates on a cycle-by-cycle basis, outer hair cells can also exhibit quasi-static length changes that occur on much slower timescales. In particular, outer hair cells respond to acoustic stimulation through static contraction and sometimes elongation. Moreover, the static length change is largest when the frequency of the stimulation matches the characteristic frequency of the cochlear location of the outer hair cell [13, 14]. Although discovered almost thirty years ago, the biophysical relevance of such static length changes of outer hair cells for the functioning of the organ of Corti remains unresolved.
Another main uncertainty regarding the feedback by outer hair cells concerns the micromechanics of the organ of Corti in the low-frequency apical region of the cochlea that is responsible for detecting frequencies below a few kHz that are most important for speech and music [15, 16]. Recent in vitro experimental studies have indeed shown that the apical organ of Corti deforms in a complex and unexpected way [16–21]. When stimulated electrically, the outer hair cells contracted and pulled the reticular lamina, in which the hair bundles of outer hair cells are anchored, towards the basilar membrane. Surprisingly, the lateral portion of the organ of Corti composed of the Hensen cells moved in the opposite direction, away from the basilar membrane, at an amplitude larger than that of the reticular lamina [20]. No vibration could be detected from the adjacent portion of the basilar membrane [16]. The mechanisms producing this complex motion of the organ remain unclear.
Here we set out to identify the origin of the complex internal motion of the organ of Corti at the cochlear apex and the influence of static length changes of outer hair cells. We show that a plausible assumption about the apical organ of Corti, namely that each cross-section is incompressible, highly constrains the organ’s internal motion. The deformation of the organ of Corti that results from length changes of the outer hair cells can then be described through a mathematical model that is based on the organ’s geometry. We develop this model and verify it through comparison with existing [16, 20] as well as newly acquired experimental data, where length changes of the outer hair cells were induced by current injections inside scala media. Our results reveal that static length changes of the outer hair cells can sensitively determine how much of the sound-evoked motion is transferred to the reticular lamina, thus providing a novel mechanism for outer hair cells to regulate hearing sensitivity.
Sound elicits a traveling wave on the basilar membrane which triggers the deflection of hair bundles and thus the electromotile response of the outer hair cells. As the outer hair cells contract, the reticular lamina and basilar membrane are pulled towards each other [7] (Fig 1). This can potentially reduce the cross-sectional area of the fluid-filled space of Nuel, causing fluid inside the organ of Corti to be displaced longitudinally, that is, along the cochlear canal. The volume of displaced fluid is proportional to the change in cross-sectional area, multiplied by the longitudinal extent l of the organ that contracts. For a traveling wave, this longitudinal extent l is approximately half the wavelength, and the amplitude of the evoked fluid velocity is proportional to the displaced volume and thus to the wavelength.
Near the cochlear apex, low-frequency sound elicits a wave with a long wavelength of several millimeters [2, 22]. The longitudinal extent over which the organ of Corti deforms similarly thus far exceeds the width and the height of the space of Nuel, which are of the order of 100 μm. Longitudinal fluid flow inside the organ of Corti would thus require velocities much larger than the velocity of the length-changing outer hair cells, and would hence be counteracted by viscous friction. We conclude that such longitudinal flow is suppressed and that the cross-sectional area of the organ of Corti in the apex is accordingly conserved when the outer hair cells change length. The same reasoning holds for in vitro experiments using electrical stimulation, due to the long effective range of the electrodes, but not for deformation of the organ of Corti near the cochlear base where the wavelengths in the peak region of a traveling wave can be much shorter, below one millimeter [2].
The cross-section of the organ of Corti can be divided into two components, a fluid-filled space on the neural side of the organ, and a portion representing the body of Hensen cells on the abneural side (see Fig 1(B)). The cross-sectional area of each component needs to be conserved separately: the fluid space because of the argument above, and the Hensen cells because their cytoplasm cannot escape longitudinally.
The motion of the cochlear partition can be decomposed into a passive component, where all structures follow the sound-evoked displacement of the basilar membrane [23], and an active component that involves internal deformation of the organ of Corti caused by outer hair cell forces. Here we seek to determine the motion of various structures of the organ of Corti—in particular of the Hensen cells, the reticular lamina, and the outer hair cells—relative to the basilar membrane.
We use the constraint of a conserved cross-sectional area to estimate the active deformation of the organ of Corti from its geometry. The length change of outer hair cells is characterized by its relative contraction ϵ, such that the length of an outer hair cell is given by LOHC(ϵ) = (1 − ϵ)LOHC,0, where LOHC,0 is the resting length of the cell (Fig 1(B)). A length change of the outer hair cells can result from electromotility as well as from hair bundle motility that can exert force on the reticular lamina [24–26]. Experiments using electrical stimuli on isolated and unloaded outer hair cells indicate that the magnitude of the outer hair cell contraction is |ϵ| ≲ 0.02 [4]. Other anatomical elements of the organ of Corti are assumed to have constant length, except for the Deiter’s cells and the contour of the Hensen cells. Motion of the reticular lamina can be approximated as pivoting about the top of the pillar cells, which is why we lump the three rows of outer hair cells and Deiter’s cell in a single, effective row, located near the third row of outer hair cells.
Since we consider small deformations only, we assume linear relationships between the length change of the outer hair cells and the length LDC(ϵ) of the Deiter’s cells as well as the length LHC(ϵ) of the contour of the Hensen cells. We can therefore write LDC(ϵ) = (1 + ϵΔ)LDC,0 and LHC(ϵ) = (1 + ϵΓ)LHC,0 with the resting lengths LDC(ϵ = 0) = LDC,0 and LHC(ϵ = 0) = LHC,0. The two parameters Δ and Γ quantify the extent to which the Deiter’s cells and the Hensen cell contour change their length as a result of an outer hair cell length change, respectively. In the following, we will refer to them as extensibilities. Note that we have introduced Δ and Γ as purely geometrical parameters; they do not correspond in a simple way to material properties of the Deiter’s or Hensen cells alone. Rather, they are the result of the complex interplay of the material properties and the geometrical arrangement of all the different elements that comprise the cochlear partition and resist deformation through outer hair cell forces. Their values are therefore a priori unknown. Furthermore, we assume that Δ and Γ are constants and, in particular, independent of the frequency of outer hair cell deformations. We thus consider elastic deformations only and neglect effects from viscosity or inertia in the system. However, at the low frequencies considered here, the latter are indeed relatively less important [27].
We assume that the Deiter’s cells can pivot around their attachment on the basilar membrane and that they do not bend. The arc of Hensen cells is treated as a thin elastic body that deforms around a preferred shape, characterized by its local curvature along its length. Details of the model calculations are given in the Materials and Methods.
We characterize the motion of the deforming organ of Corti by the motion of specific points along the arc of Hensen cells (Fig 2). Our model shows that, depending on the values of the Deiter’s cell extensibility Δ and the extensibility Γ of the Hensen cell contour, different types of motion can occur. Which of these does occur in experiments?
Having constrained both free parameters of our model, we compared the resulting model predictions to additional known features of apical micromechanics. Recent in vitro experiments have shown that outer hair cells essentially pivot around their attachment at the reticular lamina when stimulated electrically [20]. Outer hair cells were first subjected to a negative current, yielding a reference state, and then to a positive current of equal magnitude. The change in current leads to contraction of the outer hair cells which were found to rotate the cell’s base towards the stria vascularis (Fig 5(A)). The angle of this rotation was quantified for different amplitudes of electrical stimulation and was found to increase linearly for small stimulation amplitudes but to saturate at larger ones [20] (Fig 5(B)).
Our model shows that the reticular lamina moves much less than the length change of the outer hair cells which is consistent with the essentially rotational motion of the outer hair cells found experimentally. Direction and amount of the rotation depend on the size of the outer tunnel of Corti as parametrized by the angle φ between the outer hair cells and the arc of the outer tunnel (Figs 1(B) and 5(C)). For simplicity, we here consider the reticular lamina as fixed and regard the organ at the hyporpolarized state of the outer hair cell as the reference position. The amount of contraction considered in Fig 5(C) therefore ranges from ϵ = 0 to approximately ϵ = 0.04, rather than from ϵ = −0.02 to ϵ = 0.02 as before. Our model correctly predicts the direction of outer hair cell rotation when the outer tunnel is large, which agrees with the geometry commonly seen in micrographs: the realistic geometry is arguably the one where the outer tunnel arc lies in almost tangential continuation of the reticular lamina. Furthermore, the amplitude of the applied current can be related to the amplitude of the length change of the outer hair cells if we assume that the saturation observed for high currents corresponds to the maximal contraction of the outer hair cells of about 4% [4, 19]. The predicted rotation angles for a realistic geometry (the bluest curve in Fig 5(C)) are then in good quantitative agreement with the experimental data (Fig 5(B)).
While our model does not explicitly describe displacements of internal points of the Hensen cells, it suggests a motion pattern in which the entire body of Hensen cells is essentially displaced as one by the contracting outer hair cells with little internal deformation. As a consequence, structures at different depths within the organ are expected to show approximately constant vertical displacement, and the displacement decreases only close to the basilar membrane. In particular, the direction of displacement remains the same throughout the entire height of the organ. In contrast, if the cross-section of the organ of Corti were to change, such as through fluid being pressed into the outer tunnel, vertical displacement would vary markedly and change direction as a function of depth.
We interferometrically determined current-evoked displacements from positions at different depths. We found that the direction of Hensen-cell displacement, as well as the displacement amplitude, vary little with depth (Fig 6(A)). While the direction of the displacements with respect to the applied current was consistent in all preparations, the amplitude of the evoked displacements varied considerably between preparations, as well as with time in a given preparation. For this reason, results shown in Fig 6(B) have been normalized to the average displacement at the surface of the Hensen cells. The displacement amplitude exhibited a small but significant decrease with increasing depth (13% on average; a linear mixed model reveals a negative slope of -0.0008/μm in normalized displacement units, p = 0.0014, t = −3.22, d.f. = 532; data from 540 measurements from 7 preparations). This agrees with our model that revealed that the counter-intuitive direction of Hensen-cell motion under electrical stimulation is due to large motion at the bases of outer hair cells.
As the measurement became increasingly noisy with increasing depth inside the tissue, we were not able to determine the location of the basilar membrane. The fact that large displacement amplitudes persist with depth suggests, however, that some basilar membrane motion occurs underneath the Hensen cells. In contrast, such motion was not detectable in the portion of the basilar membrane lateral to the organ of Corti. This is consistent with recent in vivo measurements obtained using optical coherence tomography [16].
Inner hair cells are responsible for detecting the mechanical sound vibrations and transducing them into electrical signals that are then forwarded to the brain. The hair bundles of the inner hair cells are deflected by oscillatory fluid flow between the reticular lamina and the tectorial membrane, whose magnitude is dependent on the vibration amplitude of the reticular lamina, at least for frequencies up to 3 kHz [17]. Therefore, the nonlinear reticular-lamina displacement upon length change of the outer hair cells that is predicted by our model has striking consequences for inner hair cell stimulation (Fig 4(B) and 4(C)).
Sound vibration at a frequency f leads to an oscillating length change of the outer hair cells around some resting position ϵ(0):
ϵ ( t ) = ϵ ( 0 ) + ϵ ( osc ) sin ( 2 π f t ) . (1)
This length change elicits an oscillating reticular-lamina motion DRL(t) at an amplitude D R L ( osc ) around the steady displacement D RL ( 0 ) that is set by the outer hair cell’s steady contraction ϵ0:
D RL ( t ) = D RL ( 0 ) + D RL ( osc ) sin ( 2 π f t ) . (2)
The amplitude of an oscillating length change of an outer hair cell for sound pressures in the hearing range is small [28], |ϵ(osc)| ≪ 0.02. The amplitude of the resulting reticular-lamina vibration D RL ( osc ) can thus be approximated by a linear expansion around the resting amplitude D RL ( 0 ):
D RL ( osc ) = d D RL d ϵ | ϵ ( 0 ) ϵ ( osc ) . (3)
For the Deiter’s cell extensibility Δ = 1.15 that we identified above, the derivative of reticular-lamina displacement with respect to hair-cell contraction, dDRL/dϵ, varies monotonically from approximately zero at a resting length change ϵ(0) = −0.02 of the outer hair cells to a value of approximately -7 at ϵ(0) = 0.02. As a result, an oscillating length change of the outer hair cells around a maximally elongated resting length defined by ϵ(0) = −0.02 produces virtually no oscillation of the reticular lamina. On the other hand, a vibration of the outer hair cell length around a maximally contracted resting length defined by ϵ(0) = 0.02 leads to a seven-fold larger vibration of the reticular lamina (Fig 7). The resting length of the outer hair cells can thus sensitively determine how much vibration of the reticular lamina is elicited by an oscillating length change of the outer hair cells at low frequencies.
We have developed a model for the deformation of the organ of Corti that is based on the organ’s geometry as well as on the plausible assumption that the organ of Corti near the cochlear apex is incompressible. The model involves only two parameters that are not derived from the geometry, namely the extensibility of the Deiter’s cells and of the outer edge of the Hensen cells. Qualitative comparison of model predictions with experimental data highly constrains these parameters, and the resulting model predictions agree excellently with further data on the displacement of the outer hair cells and the vertical vibration at different depths in the organ of Corti.
A limitation of our model stems from the fact that we have considered the organ of Corti as composed only of spring-like elastic elements and neglected viscous and inertial impedances. However, as the deformation frequency is lowered, the relative importance of the latter components decreases. For frequencies up to several hundred Hertz, the organ of Corti’s impedance (albeit with the tectorial membrane removed) is found to be dominated by stiffness, rather than viscosity [27]. This corresponds to the characteristic frequencies found at the apical locations studied here, which is why our results can provide a valid approximation also for the acoustic response. We also note that our assumption of unimportant viscous and inertial impedance regards only the part of the organ of Corti between the basilar membrane and the reticular lamina, but not the subtectorial space. The latter presumably contributes the major viscous damping to the cochlear partition, and this damping may be counteracted by a cycle-by-cycle length change of outer hair cells [29–31].
Our model generically produces the non-intuitive counterphasic motion of the reticular lamina and the Hensen cells that was recently observed experimentally [17, 18, 20, 21]. Importantly, our analysis suggests that this behaviour does not result from perilymph being pressed against the Hensen cells, as hypothesized recently [21]. Instead, our model and our measurements evidence that the entire body of Hensen cells is being pulled upwards by the contracting outer hair cells. Generally, the experimental data are reproduced if the base of the outer hair cell is allowed to move somewhat more than its apex, such that the largest displacements then occur inside the organ of Corti. Intriguingly, this is corroborated by our own recent in vivo measurements using optical coherence tomography [16].
What is the origin of this internal motion? In our model, the vibrational pattern is achieved through Deiter’s cells that are fairly compliant, at least in response to quasi-static or low-frequency forcing by outer hair cells [32, 33]. Alternatively, or in addition, large displacements at the bases of outer hair cells could also occur as a consequence of a locally very compliant basilar membrane [16]. We have not detected basilar-membrane motion lateral of the organ of Corti in response to current stimulation [16, 34]. However, our interferometric measurements from different depths inside the Hensen cells indicate that some basilar-membrane motion may be present in a limited region underneath the organ, while the decrease in amplitude with depth suggests that some stretching occurs as well. Conservation of the cross-sectional area of the organ of Corti may then require counterphasic displacement of the arcuate zone of the basilar membrane, as observed by Nuttall et al. in more basal regions in response to electrical stimulation [35]. We did not include this mode of deformation in our model, as no corresponding data are available for the cochlear apex.
Current theories of cochlear function suggest that the mechanical activity of outer hair cells serves to amplify the motion of the basilar membrane [2] or the reticular lamina [36, 37] in order to render faint sounds more easily detectable for the stereocilia of inner hair cells. In this light, it seems surprising that the largest motion would occur in the interior of the organ. However, our geometrical analysis and experiments suggest that this motion pattern is associated with a nonlinear dependence of the reticular-lamina motion on the resting length of the outer hair cells. In consequence, we find that the resting length of the outer hair cells can control the magnitude of vibration of the reticular lamina that is evoked by an oscillating length change of the outer hair cells. Experimental evidence for this effect comes from the observed nonlinear dependence of sound-evoked motion on an imposed endocochlear potential in vitro [20]. It has been suggested previously that static length changes of the outer hair cells might influence the operating point of hair bundles [38], or of the micromechanics of the organ of Corti as a whole [39], but the details of such a mechanism have remained unclear. Our analysis shows that the incompressibility of the organ of Corti together with a high level of compliance at the base of outer hair cells yields a novel and intriguingly simple mechanism for the outer hair cells to regulate hearing sensitivity through their static length change. While we have thus shown the availability of such a mechanism, further experimental work and improved imaging techniques are needed to verify it in the living cochlea.
Our geometrical model quantifies the internal motion of the organ of Corti. The actual sound-evoked and active motion of the cochlear partition is a linear combination of the internal deformation and an overall net displacement. While internal motion is due to active amplification by outer hair cells, the net displacement of the organ can be caused both by sound stimulation as well as by the mechanical activity of outer hair cells. In a recently proposed ratchet mechanism, or unidirectional amplification, the outer hair cells may cause only internal deformation of the organ of Corti without displacement of the basilar membrane [24], in agreement with some recent experimental observations [16]. Further modeling that integrates the geometric model presented here with an analysis of the different forces produced by outer hair cells and their effects on the overall motion of the organ of Corti, as well as further experimental results on the linear or nonlinear response of the reticular lamina and the basilar membrane to varying sound intensity, are needed to clarify these issues.
Our findings are particularly relevant for two lines of further research. First, our results could shed new light on the role of the static and frequency-dependent motile response of outer hair cells to acoustic stimulation whose biophysical origin and function in the cochlea remain poorly understood [13, 14]. Because our model shows that a sustained length change of outer hair cells can sensitively regulate the reticular lamina’s vibration, the tuned sound-evoked static length changes of outer hair cells can serve as an effective tuning mechanism that can circumvent the poor mechanical tuning of the basilar membrane in the cochlear apex [2]. As set out above, elongated outer hair cells will transfer only little of their oscillating length change to the reticular lamina. The mechanical sound signal elicited by a pure tone may, however, cause outer hair cells at the characteristic position to contract such that their additional oscillatory response to sound is leveraged into a large vibration of the reticular lamina and thus of the hair bundles of the inner hair cells. This effect can thus endow the motion of the reticular lamina with a frequency selectivity that is independent of the mechanical tuning of the basilar membrane which is comparatively poor in the cochlear apex [2].
Second, the discussed principle could present a potential mechanism for efferent medial olivocochlear (MOC) nerve fibers that innervate the outer hair cells to modulate the auditory stimulus [40]. This efferent feedback is thought to play an important role, for instance, in our ability to understand speech in noisy environments. Our results show that efferently-mediated static length changes of the outer hair cells can modify the transfer of outer hair cell activity to reticular-lamina motion. Experiments have indeed observed efferently-induced modifications in the auditory nerve signal that is not found in the mechanics of the basilar membrane, suggesting that inner hair cell stimulation is in part directly due to outer hair cell activity [41]. This effect was present throughout the cochlea, and was particularly prominent in low-frequency regions. A mechanism as the one described here could underlie these observations.
Experiments were performed on guinea pigs. All experimental procedures were approved by the local ethics committee (permit N32/13).
We used detailed morphometric data on the guinea pig’s organ of Corti in the cochlear apex [42–44] in conjunction with high-quality micrographs [45] as a basis for the geometrical model (Fig 1). Relative sizes and orientations of different structures in the organ of Corti show a high level of consistency between the different data sources. The contour of the Hensen cells is represented by a polynomial curve approximating the shape seen in micrographs. Since we assume the reticular lamina to pivot as a stiff rod around its attachment near the inner hair cell [8, 9], we have for simplicity lumped the three rows of outer hair cells and Deiter’s cells into a single one, located at the position of the outermost row.
Young pigmented and albino guinea pigs of both sexes weighing 200 to 400 g were used in the current study. The animals were housed at six animals per cage and a 12-hour light/dark cycle. Using procedures approved by the local ethics committee (permit N32/13), the temporal bones were removed, attached to a custom holder, and the bulla opened to expose the cochlea. The preparation was then immersed in oxygenated tissue culture medium (Minimum Essential Medium, Invitrogen, Carlsbad, CA, USA) and a small opening created over scala vestibuli in the apical turn. This opening provided optical access to the organ of Corti and also allowed the tip of a beveled glass microelectrode to be pushed through the otherwise intact Reissner’s membrane. The electrode was used throughout the experiment to monitor the sound-evoked potentials produced by the sensory cells. Data collection was aborted if these potentials underwent sudden changes, or if their initial amplitude was abnormally low. The electrode was also used for injecting electrical currents into scala media. The currents were generated by an optically isolated constant current stimulator (A395, World Precision Instruments, Sarasota, FL, USA). For our different experiments, we used either linear current ramps or current steps as stimuli, with durations between 50 ms and 100 ms and amplitudes of up to 30 μA. Scala tympani was continuously perfused with oxygenated tissue culture medium at a rate of ∼ 0.6 ml/h, starting within 10 minutes of decapitation, and the perfusion system was also used to introduce the dye RH795 (5 micromolars, Biotium, Howard, CA, USA), which provides fluorescent labeling of the cell membranes of sensory cells and neurons. All experiments were performed at room temperature (21–24°C).
|
10.1371/journal.ppat.1004285 | Cryptococcus gattii VGIII Isolates Causing Infections in HIV/AIDS Patients in Southern California: Identification of the Local Environmental Source as Arboreal | Ongoing Cryptococcus gattii outbreaks in the Western United States and Canada illustrate the impact of environmental reservoirs and both clonal and recombining propagation in driving emergence and expansion of microbial pathogens. C. gattii comprises four distinct molecular types: VGI, VGII, VGIII, and VGIV, with no evidence of nuclear genetic exchange, indicating these represent distinct species. C. gattii VGII isolates are causing the Pacific Northwest outbreak, whereas VGIII isolates frequently infect HIV/AIDS patients in Southern California. VGI, VGII, and VGIII have been isolated from patients and animals in the Western US, suggesting these molecular types occur in the environment. However, only two environmental isolates of C. gattii have ever been reported from California: CBS7750 (VGII) and WM161 (VGIII). The incongruence of frequent clinical presence and uncommon environmental isolation suggests an unknown C. gattii reservoir in California. Here we report frequent isolation of C. gattii VGIII MATα and MATa isolates and infrequent isolation of VGI MATα from environmental sources in Southern California. VGIII isolates were obtained from soil debris associated with tree species not previously reported as hosts from sites near residences of infected patients. These isolates are fertile under laboratory conditions, produce abundant spores, and are part of both locally and more distantly recombining populations. MLST and whole genome sequence analysis provide compelling evidence that these environmental isolates are the source of human infections. Isolates displayed wide-ranging virulence in macrophage and animal models. When clinical and environmental isolates with indistinguishable MLST profiles were compared, environmental isolates were less virulent. Taken together, our studies reveal an environmental source and risk of C. gattii to HIV/AIDS patients with implications for the >1,000,000 cryptococcal infections occurring annually for which the causative isolate is rarely assigned species status. Thus, the C. gattii global health burden could be more substantial than currently appreciated.
| The environmentally-acquired human pathogen C. gattii is responsible for ongoing and expanding outbreaks in the Western United States and Canada. C. gattii comprises four distinct molecular types: VGI, VGII, VGIII, and VGIV. Molecular types VGI, VGII, and VGIII have been isolated from patients and animals throughout the Western US. The Pacific Northwest and Canadian outbreak is primarily caused by C. gattii VGII. VGIII is responsible for ongoing infections in HIV/AIDS patients in Southern California. However, only two environmental C. gattii isolates have ever been identified from the Californian environment: CBS7750 (VGII) and WM161 (VGIII). We sought to collect environmental samples from areas that had confirmed reports of clinical or veterinary infections. Here we report the isolation of C. gattii VGI and VGIII from environmental soil and tree samples. C. gattii isolates were obtained from three novel tree species: Canary Island pine, American sweetgum, and a Pohutukawa tree. Genetic analysis provides robust evidence that these environmental isolates are the source of human infections.
| Outbreaks of infectious diseases caused by all major classes of microbial pathogens occur globally, annually, and these outbreaks pose public health challenges [1]–[4]. Advancing understanding of forces driving outbreaks to enhance our ability to predict, contain, and blunt their impact include: 1) identification of environmental sources and vectors, and 2) defining genetic mechanisms that give rise to infectious microbes with altered virulence or transmissibility.
Over 200 species of fungi are recognized as human/animal pathogens [5]. Fungal pathogens also cause outbreaks of disease. This includes clusters of infections caused by Coccidioides immitis/posadasii, Histoplasma capsulatum, or Apophysomyces trapeziformis following soil perturbations (earthquakes, tornadoes, dust storms, construction, landscaping) [6]–[13]. The Cryptococcus pathogenic species complex includes the globally distributed human pathogens C. neoformans and C. gattii, which cause significant fungal disease burden and increasing public health costs in immunocompromised and immunocompetent individuals worldwide [14]–[19]. Cryptococcus is annually responsible for >1,000,000 infections, >620,000 deaths, and one-third of all AIDS related deaths [17]. Cryptococcus neoformans and C. gattii have been isolated from various environmental sources (soil, trees, bird guano) but reports genetically linking specific environmental reservoirs to individual cases are limited because: 1) infections often take considerable time to diagnose, 2) are not reportable diseases in the USA and abroad, 3) C. neoformans and C. gattii are not often distinguished by species or by molecular type, and 4) retrospective environmental surveys are rare and scattered, may lack molecular analysis and corresponding clinical isolates, and/or follow many years after reported clinical infections [13], [20].
Cryptococcus gattii comprises four distinct molecular types: VGI, VGII, VGIII, and VGIV, without evidence of genetic exchange of nuclear genomes, providing evidence the four molecular types represent distinct, related species [16], [21]–[25]. C. gattii VGII and VGIII are associated with two distinct expanding outbreaks in the Western US [22], [26]. Phylogenetic analysis suggests that C. gattii and C. neoformans diverged ∼40 million years ago and VGII is the ancestral molecular type to the C. gattii clade [23], [24], [27]–[30]. Genetic rearrangement between the VG types could actively suppress recombination, acting as reproductive barriers, and limiting productive recombination. However, analysis of the mitochondrial genomes of C. gattii VGI and VGII lineages indicates a highly clonal mitochondrial genome within each lineage consistent with uniparental mitochondrial inheritance but different genealogies support the hypothesis that examples of mitochondrial genome transmission from VGII into VGI isolates have occurred [31], [32]. Furthermore, transmission of hypervirulence traits within and between different molecular types was recently demonstrated in laboratory crosses [33]. Thus, given evidence linking mitochondrial function to enhanced intracellular proliferation of VGII outbreak isolates in macrophages, and previous studies linking mitochondria to virulence of plant fungal pathogens, these findings illustrate how genetic exchange can impact virulence of pathogenic fungi [33]–[35].
C. gattii molecular type VGII, which is highly virulent and has a predilection for infecting apparently healthy hosts, is causing an outbreak on Vancouver Island that has expanded to the Canadian mainland and also Washington, Oregon, and possibly California [21], [22], [36], [37]. As a result of increased sampling and molecular analysis, three sub-molecular types are now recognized: VGIIa, VGIIb, and VGIIc. The Vancouver, BC and Pacific Northwest (PNW) outbreak is characterized by infections caused predominantly by VGIIa, the major genotype, with a lower frequency of incidence of the VGIIb minor genotype. A novel, distinct, highly virulent molecular type designated as VGIIc is presently restricted to Oregon [26], [38]. Prior to the now recognized outbreak in the Pacific Northwest, C. gattii VGIIa was isolated in 1970 from a human sputum sample (Seattle, Washington, NIH444) and in 1992 a closely related VGII isolate was obtained from an environmental sample from San Francisco, California (CBS7750) [21], [39], [40]. Sporadic isolates possibly related to the PNW VGIIa major genotype at a limited number of MLST loci have been reported outside the USA [25], [41], [42]. Due to the scarcity of these VGIIa related genotypes outside the USA, the environmental reservoir and source has not been established. On the other hand, while isolates related to the VGIIb minor genotype at a limited number of MLST loci have been reported from several geographical regions, VGIIb isolates that are indistinguishable across 30 MLST loci have only been reported from Australia, thus providing robust evidence that this is the likely geographic source of the VGIIb outbreak isolates [21], [25], [27], [42], [43].
In contrast, C. gattii molecular type VGIII is responsible for ongoing infections in immunocompromised HIV/AIDS patients in Southern California and the Southwestern US [18], [22], [37], [44], [45]. Outside the US VGIII has been associated with sporadic infections in Brazil, Colombia, Mexico, India, Germany, and Korea [14], [23], [37], [46]–[51]. Despite the high preponderance of infections caused by C. gattii VGIII in the HIV/AIDS population of California and the Southwestern USA, only one environmental VGIII isolate (WM161) has ever been recovered from California [23], [37], [44], [52]. On the other hand, in Colombia and Argentina VGIII isolates have been isolated from the environment, yet clinical prevalence appears low in those localities [47], [53]–[55].
C. gattii VGIII has been further classified into two groups, VGIIIa and VGIIIb, based on MLST analysis [22]. Analysis of the Californian VGIII clinical population indicated that VGIIIa is more clonal with only one identified MATa isolate in comparison to VGIIIb, in which both MATα and MATa isolates were frequently identified [22], [45]. Unlike other VG molecular types or C. neoformans, MATa isolates are frequently isolated from both clinical and environmental populations of VGIIIb, suggesting that the VGIIIb population may be fertile and actively undergoing a-α sexual recombination in nature [56]–[58]. VGIII environmental isolates have been reported from Tipuana tipu trees in Argentina, Terminalia catappa, Corymbia ficifolia, Eucalyptus sp, and Ficus sp. in Colombia, and Manilkara hexandra in India [47], [49], [53]–[55]. Analysis of 60 VGIII isolates from California resulted in the identification of only four alleles shared between VGIIIa and VGIIIb in four independent isolates, CAP10 and TEF1 appear to be ancestral, while the shared PLB1 allele appears to have been introgressed between VGIIIa and VGIIIb [22]. The MLST analysis, haplotype analysis, paired allele graphs, percentages of compatible loci, and indices of association all support the divergence of the VGIIIa and VGIIIb subtypes, limited recombination within VGIIIa, and more frequent recombination in VGIIIb [22].
Prior to the Pacific Northwest outbreak caused by C. gattii VGII, California was historically recognized as the oldest known region associated with the occurrence of cryptococcosis within the USA. Early studies denoted California as the most prominent site of serotype C infections, later identified as C. gattii VGIII. Byrnes et al. 2011 was the first to establish the prevalence of C. gattii VGIII within the HIV/AIDS clinical population of California but did not obtain nor identify any VGIII isolates directly from the environment [22]. Of note, the MLST profile of the historically identified environmental isolate WM161 did not match the MLST profiles of any of the reported clinical isolates obtained from California; therefore, the infectious environmental reservoir of C. gattii VGIII remained unknown. Previous environmental surveys within the Pacific Northwest, USA have identified the environmental niche of C. gattii VGII [59], [60]. The lack of genetic exchange between the molecular types, genomic rearrangements confined within molecular types, and differences in host predilection strongly suggest that the molecular types recognized within C. gattii comprise cryptic species. We therefore sought to identify the environmental reservoir of C. gattii VGIII in California by sampling plants and soil in areas near residents with confirmed C. gattii infections.
As a result of this study we have identified the local environmental source of C. gattii VGI and VGIII associated with trees and soil debris in the greater Los Angeles area of California. Molecular typing and whole genome analysis of the Californian C. gattii VGIII environmental isolates indicates a high amount of molecular diversity exists in this region, VGIIIa and VGIIIb can colonize similar and overlapping niches, and some environmental isolates are closely related to clinical isolates and the likely source of previously reported human infections, and a potential reservoir for inciting new infections. Furthermore, our results suggest that many of these isolates are sexually competent and capable of intra- and inter-molecular mating which might facilitate dispersal and propagate altered infectious characteristics such as virulence or transmissibility. Further work is needed to fully elucidate the links between Cryptococcus molecular types, environmental distributions, and the propensity to initiate disease in exposed hosts.
The long term and extensive clinical prevalence of C. gattii in California has implicated a local endogenous reservoir but even with increased environmental sampling the environmental source of infections has remained undefined. We therefore sought to collect environmental samples from areas that had confirmed reports of clinical and/or veterinary infections. From soil and swab samples collected at 24 locations throughout the greater Los Angeles area over a two year period (Figure S1), we screened 146 environmental (n = 107) and clinical (n = 39) isolates utilizing canavanine glycine bromothymol blue (CGB) and niger seed (NGS) indicator media. We identified 30 (20.6%) potential C. gattii isolates, 11 clinical and 19 environmental isolates obtained from 4 out of 24 sites (16.7%) that were melanin positive on NGS and produced a blue color reaction on CGB agar (Figure 1 and Table S4). Eleven of 39 clinical (28.2%) and 19 of 107 (17.8%) environmental isolates were also confirmed as C. gattii via differences in ATP6 PCR product length and MLST analysis (Figure 1, Tables S4 and 5) [61]. C. gattii was isolated in association with three novel host tree species: Pinus canariensis (Canary Island pine, isolates LMES-3A, MCP-1A, MCPR1-X, USC-X, USC2-SIC), Liquidamar styraciflua (American sweetgum, isolates 78-1-S3A, BHPP3-X), and Metrosideros excels (Pohutukawa tree, isolate BHPP1-X). All C. gattii isolates were haploid by FACS analysis (Table S4). An additional 100 isolates were further identified as C. neoformans based on positive pigmentation on NGS agar, negative blue color formation/no growth on CGB agar, and IGS1 sequence (data not shown).
To characterize the molecular types prevalent in the Californian isolates, MLST analysis was performed on 11 unlinked loci for the 11 clinical and 19 environmental C. gattii strains obtained from Los Angeles, California USA. The data were compared to previously reported VGIII isolates [22] to determine the presence of novel or shared MLST profiles between environmental and clinical isolates (Figure 1 and Table S4). Isolates primarily clustered into three previously recognized molecular types: VGI, VGIIIa, and VGIIIb. The VGI molecular type found in California shares the most common MLST type observed worldwide (Table S5). Both VGIIIb MATα (5/14, 36%) and MATa (9/14, 64%) mating types were obtained from clinical and environmental isolates, suggesting the potential for an actively sexually recombining population in Southern California within this molecular type subgroup. In contrast, all of the VGIIIa isolates characterized were found to be MATα, suggesting that opposite sex mating may be rare among isolates of this subtype. Overall, nine isolates typed as VGIIIb MATa (2 clinical and 7 environmental), five as VGIIIb MATα (4 clinical and 1 environmental), 12 as VGIIIa MATα (4 clinical and 8 environmental), and three isolates typed as VGI MATα (Tables S4 and S5). These results reveal a, widespread reservoir of C. gattii in the greater Los Angeles area as the likely source of frequent cryptococcal infections in HIV/AIDS patients in Southern California.
Sixteen isolates, 14 environmental (7 VGIIIa and 7 VGIIIb) and 2 clinical (1 VGIIIa and 1 VGIIIb), shared four different MLST profiles (designated groups G1 to G4, represented with brackets) with previously reported clinical isolates (Figure 1), indicating shared descent between recently identified environmental isolates and previously reported clinical isolates spanning an isolation period of up to 12 years. Six new MLST sequence types were identified: five VGIIIb (4 clinical and 1 environmental) and one uncategorized VGIII type (7685027) (Figure 1). Additional analysis of the CAP59, TOR1, SOD1, and URA5 loci did not improve resolution of isolates that shared an indistinguishable 8-locus MLST profile. Thus, there was a high level of diversity in these environmental isolates reflected by the recovery of a multitude of distinct MLST profiles (Figure 1 and Table S4). Furthermore, we obtained both VGIIIa and VGIIIb molecular sub-types from the same sample (BHPP1 designation) or environmental location (MCP/MCPR or BHPP1/BHPP3 designations) suggesting that both molecular types can colonize the same ecological niche in California fostering the potential for hybridization. Thus, their lack of frequent genetic exchange suggests that the two molecular types could represent distinct species in which mating is limited by genetic and/or temporal barriers limiting productive introgressions between the molecular types.
One clinical isolate (7685027) did not align with either of the two previously recognized VGIIIa or VGIIIb molecular types and either represents a novel VGIII subgroup that has been undersampled or a novel hypermutator isolate [62]. This isolate contains many completely novel MLST alleles (URA5, TOR1, PLB1, MPD1, LAC1, CAP10, GPD1, and IGS1) not observed previously (Figures 1 and 2), consistent with a possible origin via action of a hypermutator phenotype strain resulting in the rapid emergence of novel MLST alleles [62]. On the other hand, other MLST alleles (SOD1, CAP59, and TEF1) are shared with two additional isolates (IHEM14941S and IHEM14941W) obtained from a Mexican immigrant diagnosed in Spain and thus they may represent a novel VGIIIc sub-molecular type.
In addition to MLST analysis of nuclear loci, we expanded our analysis to explore the mitochondrial genome because the mitochondria are primarily uniparentally inherited from the MATa parent, uniparental/biparental mitochondrial inheritance rates may differ between molecular types, and their genome evolves independently from the nuclear genome [32], [33], [63]–[66].
Phylogenetic analysis indicated that the MtLrRNA locus was indistinguishable between VGIIIa and VGIIIb isolates, suggesting closely related shared ancestry (data not shown). Utilizing ATP6-Byrnes primers, we observed distinct PCR size polymorphisms between VGI (300 bp), VGII (200 bp), and VGIII (≥600 bp) [61]. Additional PCR and MLST analysis of all VGIII isolates indicated a distinct size polymorphism between VGIIIa (600 bp) and VGIIIb (700 bp) except for a small subset of C. gattii VGIIIb isolates that exhibited a PCR product equivalent to the VGIIIa (600 bp) isolates. Large differences between the ATP6-Byrnes [33], [61] PCR product size limited informative phylogenetic sequence analysis between C. gattii molecular types so we utilized additional ATP6-Bovers primers previously used for mitochondrial analysis in Cryptococcus species [32]. Phylogenetic analysis of ATP6-Bovers primer sequences indicated two clades mostly segregating VGIIIa from VGIIIb, and MATa from MATα isolates as expected with the exception of a small group of clinical VGIIIb isolates (71805076, DUMC140.97, and HM-1) that share mitochondrial sequences with VGIIIa (Figure S2a). DUMC140.97 also shares nuclear sequence of CAP10 (allele 2) with VGIIIb but no evidence of shared nuclear sequence is evident in the MLST analysis of HM-1 and 7180567. Haplotype network analysis indicates the MATa VGIIIb allele as ancestral and the shared MATα allele as derived suggesting further evidence of a rare introgression event with mitochondrial recombination occurring between VGIIIa and VGIIIb isolates (Figure S2b).
We further examined the presence and distribution of shared MLST alleles between VGIIIa and VGIIIb by paired allele and haplotype network analysis, determining the evolutionary history of each allele and testing for evidence of recombination within or between molecular types in California. Two clinical VGIIIb isolates (6797194 and 7618666) share CAP10 allele 2 with known VGIIIa isolates. Paired allele diagrams between the MAT locus and LAC1, or LAC1 and TEF1, illustrate recombination occurring in the Californian VGIII population (Figure 3). Evidence of recombination is present within VGIIIb and the more clonal VGIIIa population (Figure S3). Especially notable is the evidence for recombination between the MAT locus and LAC1 in VGIIIb within the Californian isolates, suggesting that a-α mating is ongoing between isolates harboring SXI2a allele 4 and SXI1α allele 48.
Haplotype network analysis indicates that for two loci, PLB1 and CAP10, the alleles shared between VGIIIa and VGIIIb isolates represent recent introgression events because the shared allele occupies a position other than the ancestral haplotype network allele (represented within the rectangle, Figure 4). In contrast haplotype network analysis of the TEF1 share allele 21 represents the origin of the allele as ancestral. Additional haplotype network analysis of the evolutionary history of 9 additional MLST loci not shared between VGIIIa and VGIIIb robustly supports genetic isolation and provides evidence for cryptic speciation although evidence of additional shared alleles may emerge in the VGIII population with increased sampling and analysis of a broader set of nuclear loci (Figure S4).
We next determined that the newly isolated environmental strains are extremely closely related to the previously reported clinical strains based on whole genome sequencing. We sequenced matched clinical and environmental isolates representing 3 of 4 matched MLST groups, VGIIIa (group 1, CA1308, MCP-1A, 78-1-S3A and group 2, CA1053 and BHPP3-S1A) and VGIIIb (group 3, CA1508 and MCPR1-S1B) to determine whether environmental isolates could represent the source population for previously reported clinical infections. Whole genome sequence analysis of MLST matched clinical and environmental strains indicates that the genomes of grouped isolates are remarkably similar and very closely related (Figure 5 and Table S6, S7, and S8). The genomic sequences for these strains were aligned to the closest reference genome available, the VGII outbreak strain R265. SNP calling indicated a very high level of diversity in the total set of sequenced strains with a total of 773,900 polymorphic sites relative to the VGII isolate R265 [67]. This corresponds to approximately 4.4% of the genome. Within the VGIII sequenced set, there was ∼10-fold less diversity with 87,219 polymorphic sites corresponding to 0.49% of the ∼20 MB genome, which is still substantial. However, in spite of this diversity, the MLST matched isolates were remarkably similar to each other, ranging from a minimum of 46 differentiating sites (0.00026%) between VGIIIa isolates CA1053 (clinical) and BHPP3-S1A (environmental) to a maximum of 183 (0.0010%) between VGIIIb isolates CA1508 (clinical) and MCPR1-S1B (environmental, Figure 5). The SNPs identified within each matched set were characterized based on potential impact and the findings are summarized in Tables S6, S7, and S8. Variants altered three splice sites and introduced a moderate number of nonsynonymous SNPs. These nonsynonymous sites were also characterized based on predicted function, and a number of candidate pathogenesis factors were identified, including some with putative roles in heat tolerance, oxidative stress resistance, and stress response. Whole genome sequencing analysis indicates that the three identified environmental C. gattii populations are the likely environmental reservoirs for recent clinical infections and could serve as the source of additional infections.
In addition to MLST analysis and population analysis, mating assays were conducted that established these environmental and clinical VGIII isolates are fertile with C. neoformans (VNI, VNII) and C. gattii (VGII, VGIII) isolates (Table S9a and S9b). Both MATa and MATα isolates produced positive mating reactions as assessed by light and electron microscopy with lateral hypha emanating from the colony (Figure 6), fused clamp connections (Figure S5a and S5b), basidia, and basidiospores (Figure 6 and Figure S5). All strains tested with the exception of two clinical strains (7618666 and 7795476) were fertile with at least one tester strain, although this did not always include mating with a VGIII strain (Table S7b). Both VGIIIb MATα and MATa isolates were fertile, and all VGIIIa MATα environmental isolates were fertile with VNI, VNII, or VGIII tester strains (Table S9b). Environmental isolates were generally more fertile than clinical isolates; however, spore viability was not assessed to determine if these matings produced viable progeny. Clinical isolates are in general less fertile and may have undergone genetic or epigenetic modifications in the host that reduce their fertility. These results suggest that 1) VGIIIb isolates may undergo a-α opposite-sex mating in the environment, 2) many VGIIIa isolates are fertile and could participate in opposite-sex mating when a fertile MATa isolate is present, and 3) interactions between molecular types could play a role in mating in the environment.
Virulence attributes of clinical and environmental C. gattii isolates were ascertained based on intracellular proliferation rates (IPR) and percent mitochondrial tubularization within J774 BALB/c macrophage cell lines. An IPR≥2 indicates a high replication rate within macrophages and is frequently correlated with higher virulence. An IPR score <1 indicates low proliferation levels were observed in macrophages, which have been correlated with a lower potential for virulence in whole animal models in previous studies. An IPR between 1 and 2 indicates moderate proliferation in macrophages and moderate virulence attributes. We observed moderate to low IPR for selected VGIIIb and VGIIIa isolates in direct comparison to the highly virulent VGIIa C. gattii major Pacific Northwest outbreak isolate R265 (Figure 7a and 7b). Percent yeast cells observed with tubularized mitochondria ranged from 9 to 37%, which suggests mitochondrial tubularization is decoupled from IPR in macrophages for C. gattii VGIII isolates compared to the VGIIa isolate R265 [21], [34] (Figure 7b). The coefficient of determination (R2 = 0.2593 is <1, p = 0.0155) indicates a weak positive linear correlation exists between IPR and the presence of tubular mitochondria in all VGIII isolates. A positive but weakly linear correlation is observed for VGIIIa (R2 = 0.3692 is <1, p = 0.0624) but not for VGIIIb (R2 = 0.001 is <1, p = 0.9226) suggesting differences in IPR and mitochondrial regulation may be associated with different molecular sub-types. Our results demonstrate that survival and replication in macrophages (IPR) is similar between environmental and clinical VGIIIa and VGIIIb isolates of both mating types. Notably some VGIII strains have high rates of tubularization prior to a known encounter with the macrophage niche (environmental isolates) while others do not display increased tubularization after encountering the macrophage niche (clinical isolates). The presence of tubular mitochondria did not appear to be strongly correlated with IPR in VGIII isolates, in contrast with previous studies on VGII Pacific northwest outbreak isolates [21], [34]. These data may reflect that VGIII isolates commonly infect HIV/AIDS patients versus VGII isolates that rarely occur in HIV/AIDS patients, and hence VGIII isolates may be in general less virulent than VGII isolates.
The virulence attributes of VGIII C. gattii isolates were further addressed in vivo utilizing the murine intranasal model. The first two environmental isolates, MCP-1A and 78-1-S3C (both identified in late 2011), were assessed for virulence in the BALB/c mouse model and the average length of survival ranged from 56 to more than 250 days post infection (Figure S6a and S6b). Mean survival ranked from shortest to longest is as follows: CA1232 (84 days), CA1053 (90.5 days), CA1308 (96 days), CA1508 (136 days), BHPP1-S2B (149 days), NIH191 (173.5 days), BHPP3-S1A (197 days) and all others were undetermined because some mice were still surviving at the termination of the experiment. MCP-1A and 78-1-S3A were less virulent than the VGIII clinical isolates RR and VV (p<0.002, Figure S6b). Environmental isolate 78-1-S3A was similar in virulence to clinical isolate CA1089 (p = 0.6409) and somewhat more virulent than the historical NIH312 isolate (p = 0.034) (Figure S6b). Our results indicate that environmental isolate MCP-1A is less virulent than environmental isolate 78-1-S3A (p = 0.0083) with which it shares an indistinguishable MLST profile, IPR, and differs by less than 90 SNPs at the whole genome level (Figure 5, Table S8), suggesting that relatively few genetic differences could be responsible for differences in virulence between two very closely related isolates. At a minimum, no more than 46 SNPs seemed to be necessary to confer a phenotypic difference (Figure 5, Table S7). Additionally, differences in epigenetic or transcriptional modifications could contribute to phenotypic differences observed between matched MLST isolates. C. gattii VGIII environmental isolates exhibited attenuated virulence in comparison to VGIII clinical isolates in the intranasal murine mouse model. Virulence observed in the mouse model correlated with previously determined moderate IPR and was decoupled from mitochondrial phenotype suggesting roles of intracellular proliferation and mitochondrial program in the divergence in virulence of C. gattii molecular types.
Virulence attributes of closely related environmental versus clinical genotypes were further examined in A/JCr mice, which are typically more permissive to cryptococcal infection than BALB/c mice. Four groups of MLST matched strains representing VGIIIa (group 2, CA1308, CA1574, and MCP-1A; group 3, CA1053 and BHPP3-S1A) and VGIIIb (group 1, CA1232 and BHPP1-S3A; group 4, CA1508, MCPR1-S1B, NIH191/ATCC32608) were compared in the A/JCr murine model (Figure 8a). Clinical isolates exhibited reduced survival time and higher virulence in comparison to MLST matched environmental isolates. More virulent clinical isolates were associated with more virulent environmental isolates (Figure 8a). MCP-1A and BHPP1-S3A were severely attenuated for virulence in comparison to other virulent strains with 9 of 10 infected animals surviving to the conclusion of the study. Two isolates, one environmental (MCPR1-S1B) and one clinical (CA1574), were avirulent for the duration of study (250+ days post infection) in A/JCr mice (all 10 of 10 animals infected surviving). Tissue burdens ranged from 107 to 108 cells/gram tissue in the lungs, 102 to 106 cells/gram tissue in the spleen, and 102 to 106 cells/gram tissue in the brain and varied greatly between isolates. Consistently, and in agreement with previously published studies, higher organ loads were observed in the lung tissue in comparison to the brain tissues and may suggest that pulmonary cryptococcosis could contribute to the observed mortality [22], [68]. Granulomas were observed macroscopically in the lungs post necropsy and microscopically in histopathological sections for both VGIIIa and VGIIIb isolates (data not shown). Although we observed differences in mean survival time between VGIII strains, high tissue burdens were observed for all strains (of both high and low virulence) post mortem (Figure 8b). This suggests that determination of virulence is not just a static predetermined trait but also dependent on duration of exposure and host status, which could play a more prominent role in disease development in longer-lived human and animal populations.
Cryptococcus is an environmentally acquired fungal pathogen of humans and infections caused by differing molecular types may result in differences in clinical presentations, treatment, and therapeutic outcomes. In this study, the antifungal susceptibilities to amphotericin B, fluconazole, flucytosine, and ketoconazole were assessed for C. gattii VGIII isolates to examine if differences in antifungal susceptibility might be associated with molecular type and/or origin of the isolates (environmental or clinical). Our results provide evidence that the MIC of VGIIIa isolates can be higher than VGIIIb for recommended drug therapies, amphotericin B (p = 0.0178, Figure 9a) and flucytosine (p<0.0001, Figure 9b), but not for the maintenance antifungal drug fluconazole (p = 0.1059, Figure 9c), or the second line drug ketoconazole (p = 0.0685, Figure 9d). No significant differences were observed between clinical and environmental isolates within VGIIIa or VGIIIb isolates suggesting the genetic propensity for antifungal resistance may exist in environmental populations (Figure 9e–h, p>0.05). However, resistance was observed in response to exposure to fluconazole and/or flucytosine in VGIIIb clinical (HM-1, 7180567, 7618666, CA1508) and environmental isolates (MCPR1-S1B, MCPR1-S2B, MCPR1-S2A, BHPP1-S2A, BHPP1-S3A) as well as VGIIIa clinical (KB-1, CA1308) and environmental (MCP-1A, BHPP1-S2B, BHPP3-S2A) isolates suggesting some isolates may have a greater propensity for the microevolution of drug resistance within the host, possibly influencing long term treatment outcomes. In summary, we observed the emergence of resistance (fluconazole and flucytosine) and higher MICs (amphotericin B and flucytosine) among different subsets of the C. gattii VGIII population for the recommended first line treatment drugs, amphotericin B and flucytosine, suggesting molecular type and the genetic propensity for resistance could possibly impact treatment and long-term outcomes.
The clinical and veterinary prevalence of cryptococcosis and the identification of C. gattii as the causative agent in the PNW outbreak stimulated increased environmental sampling that resulted in the identification of the ecological niche of C. gattii in the environment in association with native trees on Vancouver Island and in the Pacific Northwest [59]. Increased molecular typing has resulted in the identification of both C. gattii VGIIa and VGIIb molecular types with shared MLST profiles between clinical and environmental isolates, indicating the endogenous environmental reservoir for the ongoing C. gattii PNW outbreak [22], [26], [60]. The outbreak is caused by three largely clonal lineages: VGIIa/major, VGIIb/minor, and VGIIc/novel which are members of a global, fertile, recombining population that gave rise to the hypervirulent outbreak strains [21].
Previous studies have demonstrated the presence and molecular diversity of C. gattii in the clinical population of Southern California but only one historical environmental VGIII isolate (WM161) had been reported from this region, even with increased sampling [22], [45], [69]–[71]. The historical VGIII environmental isolate (WM161) differed in MLST profile (due to unique IGS and GPD1 alleles) compared to all known VGIII clinical isolates from California. Ongoing epidemic incidences of C. gattii that have emerged in the Western USA are cause for concern due to clinical infections in both immunocompetent individuals and populations with immunocompromised status, including HIV/AIDS. C. gattii is an environmentally acquired pathogen and thus the paucity of environmental isolates is incongruent with its known presence in the clinical population in California. The identification and environmental isolation reported herein of C. gattii VGI and VGIII from four independent sampling sites in the greater Los Angeles area mimics the patchy and often sporadic isolation reported for other environmentally acquired fungal pathogens (Blastomyces, Coccidioides) associated with periodic outbreaks in the USA [6], [72]–[74].
In this study we report the environmental isolation of 19 C. gattii VGIIIa, VGIIIb, and VGI isolates from Los Angeles, California from samples obtained from three novel tree species Pinus canariensis (Canary Island pine), Liquidamar styraciflua (American sweetgum), and Metrosideros excels (Pohutukawa tree). These environmental isolates were obtained from areas in close proximity to the residences of patients. We have identified an environmental source of the most common C. gattii VGI molecular type, which constitutes a smaller proportion of C. gattii infections in California and elsewhere in the US [22], [45], [75], [76]. Of the 19 environmental isolates, 16 share indistinguishable MLST profiles with previously reported clinical isolates from the Southern California region identifying endogenous environmental reservoirs for both VGIIIa and VGIIIb infections in the Los Angeles, metropolitan area, USA [22]. Therefore, we have the rare opportunity to identify and genetically link known environmental reservoirs to individual cases in Southern California.
In this study we report the presence of fertile MATa and MATα environmental isolates and demonstrate evidence of recombination within the VGIIIb population in the Southern California region. This is notable, at least in part, because a-α mating is relatively rare in the Cryptococcus species complex, with a few notable exceptions, because of the paucity of MATa isolates [56], [58], [77], [78]. In the case of C. neoformans var. grubii, the majority of MATa isolates are part of a specific population restricted to Botswana that is also hypothesized to be the origin of the species [79]. The population of VGIIIb MATa isolates could similarly represent a locally restricted founder population in California. It is also possible that the paucity of MATa isolates within the VGIIIa population, may be indicative of a high frequency of α-α mating in the population. As in VGIIIb, there is evidence for recombination within the VGIIIa population, but in this case may be occurring without the involvement of MATa isolates. While we cannot rule out the existence of MATa isolates that have remained unsampled, our current evidence suggests this population may reassort without bisexual mating. Further sampling will increase the power of this analysis Sex within C. gattii populations is important for two reasons. First, the sexual cycle produces spores that can be aerosolized as potential infectious propagules [80]. Second, sex can aid in the evolution of virulence. Same-sex mating has been proposed as a factor in the development of the VGIIa major outbreak strain in the Pacific Northwest, while a-α sexual reproduction aids in the transmission of hypervirulence [25], [33]. Through both routes, sex can contribute to the development and persistence of an outbreak. In addition, sexual reproduction led to the emergence of the modern pathogenic lineage for the common human parasite Toxoplasma gondii and selfing has recently been shown to drive outbreaks [81], [82].
C. gattii VGIII clinical and environmental isolates share similar in vitro IPR and mitochondrial phenotypes. Moderate ranges of IPR appear unrelated to the extent of mitochondrial tubularization in contrast to C. gattii VGII isolates, suggesting these VGIII isolates have reduced virulence in comparison to the VGII outbreak strains [21], [33]. We observed reduced virulence of VGIII in comparison to VGII in the murine model, consistent with the observed predilection of VGIII to infect immunocompromised individuals with HIV/AIDS compared to VGII that commonly infects otherwise healthy patients. An alternative explanation is that C. gattii VGIII causes dormant infections that can be reactivated at the time of immunosuppression whereas VGII causes primary infections. However, this alternative explanation seems less likely because VGII has also been associated with the potential to cause dormant infections [14]. We tested paired clinical and environmental isolates that shared indistinguishable MLST profiles, and although WGS indicated relatively few genetic differences, all clinical isolates were more virulent in comparison to the environmentally isolated paired MLST strains except for VGIIIa CA1574. Furthermore, in the paired MLST matched isolates that were analyzed in murine virulence assays we found that all VGIIIa strains were more virulent than VGIIIb strains, in agreement with previous findings [22]. Tissue burden and histopathological analyses indicated significantly higher fungal burdens in the lungs in comparison to the spleen or brain, consistent with previous reports of a predilection of VGIII to initiate prolonged pulmonary infections [22], [32]. Thus, although clinical and environmental isolates shared indistinguishable MLST profiles, IPR, and >99.9% sequence identity, their virulence potential differed in the murine model.
Whole genome sequencing of three different MLST matched groups representing both VGIIIa and VGIIIb populations indicated that previously reported clinical isolates (circa 2000–2005) and newly identified environmental isolates (2011–2012) are extremely similar, differing at only 46 SNPs between VGIIIa clinical isolate CA1053 and environmental isolate BHPP3-S1A, and 183 SNPs between the VGIIIb clinical isolate CA1508 and the environmental isolate MCPR1-S1B. This is all the more poignant given that the clinical and environmental isolates were ascertained over an isolation period spanning 5 to 12 years, further forging a link between the environmental isolates and clinical infections in Southern California. This can be compared to H99 lab passaged isolates obtained over a ten year period, which are distinguished by 11 SNPs and 11 indels [83]. Furthermore, a recent report on Staphylococcus aureus (genome size 2.4 Mb) used a cutoff of 40 single nucleotide variants (SNVs) to decide whether two strains were indistinguishable or different [84]. We aligned our sequences to R265 with a total genome size of 17.5 Mb, which would allow for 6.2-fold more variants and up to 249 SNVs. Thus, by even this strict comparison the VGIIIa/b clinical and environmental isolates are closely related than bacterial isolates concluded to be causally related in chains of transmission.
Whole genome sequencing allows new approaches to pathogenesis. However, comparative genomic approaches have limitations. When comparing highly pathogenic and less pathogenic isolates, finding a genetic change important for pathogenesis is challenging because of unrelated incidental genetic diversity. One method to avoid this is use strains that have a common, recent origin. This reduces genomic differences that could obscure causative genotypic changes and increases the probability of linking genotype to phenotype. In this study, environmental isolates that we propose directly gave rise to the infectious isolate were chosen via MLST matching; however, similar approaches could be used for longitudinal studies of individuals with recurrent fungal infections or for other environmental infections where a source population is readily available. While beyond the scope of this study, resource populations established through MLST matching of isolates with disparate phenotypes will provide an avenue for linkage between genotype and phenotype. Substantial differences in virulence were observed in our studies between two very similar isolates. At a minimum, no more than 46 SNPs seems necessary to confer a phenotypic difference between the clinical isolate CA1053 and the environmental isolate BHPP3-S1A; only 15 of these SNPs result in nonsynonymous changes. We sequenced three closely related VGIIIa isolates from group 2, representing two independent environmental isolates obtained from different locations and a previously identified clinical isolate. The three isolates differ by only 90 SNPs but exhibit differences in pathogenicity and antifungal sensitivity. Furthermore this genotype is noteworthy because it is represented in the current clinical population by isolate 7489719. This suggests that adaptation from an environmental lifestyle to a pathogenic lifestyle may be a very rapid event, involving epigenetic modification, changes in transcriptional regulation, or requiring few genetic changes. Passage of Cryptococcus through amoeba, nematodes, slime molds, plants, and animals yields passaged strains with enhanced virulence [68], [85]–[89]. Prolonged lab culture without host exposure can attenuate virulence [90]. Passage of strain H99 resulted in a lineage with enhanced or attenuated virulence with only 4 associated SNPs, two of which could contribute to attenuated virulence [83]. Also notable is that none of the SNPs from our three matched sets was shared between sets or in common genes. This suggests that either there are multiple independent routes of adaptation to produce a more virulent organism, or that transcriptional and epigenetic changes may also contribute.
C. gattii is an environmentally acquired opportunistic fungal pathogen that can cause acute or latent infections in both immunocompromised and immunocompetent hosts and increased recognition of ongoing outbreaks in humans, pets, and wildlife is emerging as a serious public health concern and financial burden. We report the isolation of fertile C. gattii VGIIIa, and VGIIIb isolates from environmental soil and swab samples from non-Eucalyptus host trees in the greater Los Angeles area of Southern California. MLST and WGS analysis coupled with mating and virulence studies demonstrate that environmental C. gattii VGIII isolates are actively recombining and the likely source of human acquired infections in this region and can further contribute to ongoing infections in Southern California. There is now a substantial body of accumulated data that suggest that C. gattii is associated in the environment with many tree species (native and non-native) worldwide and these hosts function as environmental reservoirs for disease outbreaks. Future environmental and epidemiological studies can further incorporate molecular and whole genome analysis to connect known clinical/veterinary incidences of cryptococcosis with newly identified ecological niches to further address the link between genotype, risk of exposure, and propensity for initiating disease.
Swabs of individual tree trunks and soil samples from around the base of trees were collected during the summers of 2011 and 2012 utilizing BD BBL Single application CultureSwabs with Liquid Amies (VWR#90001-036). In summary, over two years 24 sites were sampled to obtain 109 tree swabs of over 30 tree species and 58 soil samples from collected from the greater Los Angeles area. In 2011 samples were obtained from 9 locations, 64 trees (30 different species, including 10 Eucalyptus trees from four independent sites.), and 25 soil samples in the greater Los Angeles area. In 2012 the two sites that had previously yielded C. gattii were resampled and 15 additional sites were sampled focusing on Pinus canariensis and Liquidambar styraciflua. From these trees 45 trees were swabbed and 33 soil samples were collected. The swabs were streaked onto Niger seed (NGS) agar containing chloramphenicol (0.5 g/L, Sigma C0378-100G) Two grams of soil were suspended in 10 mL of sterile ddH20 and allowed to settle. Three 100 µl aliquots were plated on NGS agar. Plates were incubated at 30°C for 1 to 3 days. Yeast colonies producing brown pigmentation were selected and colony purified. All clinical and environmental isolates were streaked onto NGS and canavanine-glycine bromothymol blue (CGB) agar and incubated for 1 to 3 days to identify C. gattii isolates. Clinical isolates were obtained from 48 patients who were treated at the University of Southern California, Harbor-UCLA Medical Center, or Kaiser Permanente Downey Hospital between February 2008 and January 2013. Of these patients, 32 were considered to be immunocompromised due to HIV/AIDS or other causes. The clinical isolates were de-identified and linked to major cross streets near the patients' homes. Stocks were maintained at −80°C in 25% glycerol. All strains utilized in this study are listed in Table S1. Genomic DNA was prepared by the CTAB method for all isolates. Potential C. gattii isolates were screened for molecular type by size differences of the ATP6 PCR product [33], [61].
Multilocus sequence typing was performed on 12 loci (SXI1α or SXI2a, IGS, TEF1, GPD1, LAC1, CAP10, PLB1, MPD1, CAP59, TOR1, SOD1, and URA5) [15], [21], [25]. For each isolate, genomic regions were PCR amplified, purified (ExoSAP-IT, Qiagen), sequenced, and both forward and reverse strands were assembled with complete double-strand coverage. All primers utilized in this study are listed in Table S2. MLST sequences were viewed and edited by Sequencher; alignment and phylogenetic analyses were conducted using MEGA version 5 [91]. Each allele was assigned a corresponding number, or given a new number if the sequence was not already assigned an allele number in the GenBank database (NCBI), International Society for Human and Animal Mycoses MLST Database (ISHAM), or previously published reports [15], [22], [27], [45]. GenBank accession numbers with corresponding allele numbers are listed in Table S3. Haplotype network analysis was performed using TCS software (version 1.21) [92].
Ploidy was determined by fluorescence activated cell sorting (FACS) as previously described [93]. Briefly, cells were grown, passaged twice in 25 ml YPD broth at 25°C, collected by centrifugation, and washed and resuspended in 5 mL of 1x PBS. 1 mL of cells was collected by centrifugation and fixed in 70% ETOH overnight with gentle shaking at 4°C. Cells were collected by centrifugation, washed, and resuspended in 1 mL of 1x NS buffer. Cells were centrifuged and resuspended 200 µl (180 µl NS buffer, 14 µl RNase A, 6 µl propidium iodide and incubated overnight at room temperature. 50 µl of each strain (10,000 cells) was mixed with 500 µL of Tris-PI mix (482 µl 1M Tris pH 7.5+18 µl propidium iodide) and analyzed using the FL1 channel (slow laser scan) on a Becton-Dickinson FACScan. C. gattii NIH444 and C. neoformans XL143 were utilized as haploid and diploid reference controls.
Whole Genome Sequencing (WGS) was done by Illumina paired end reads. Genomes of strains CA1053, CA1308, CA1508, BHPP3-S1A, MCP-1A, 78-1-S3A, MCPR1-S1B, and 97/433 were sequenced using HiSeq2500 using paired end reads with an insert size of approximately 300 bp and a read length of 100 bp. Initial processing was performed using the Illumina Pipeline (v.1.8.2) [94]. The genome of B4546 was previously sequenced [33]. Reads were aligned to the published VGII reference genome supercontigs using BWA-sampe [95] and SNPs were called using the Genome Analysis Toolkit (GATK V2.4-9) Unified Genotyper with the haploid ploidy setting. The resulting calls were then filtered to remove calls with a quality score below 30, and individual depths below 5 reads. SNPs common to each pair were removed and the remaining SNPs differentiating the subset were manually examined using the BAM files to remove erroneous calls based on poor mapping or repetitive sequences. The resulting variants were analyzed using SnpEff to determine impact of SNPs [96]. Assignment of gene function was carried out using a combination of BLAST to the S. cerevisiae genome and CD search through NCBI.
Whole genome trees were generated from a 100 kb region aligning to supercontig 5 of the VGII reference genome. SNPs were manually filtered to eliminate erroneous calls. The resulting SNPs were used to generate alternate reference sequences using GATK's FastaAlternateReferenceMaker that were then aligned using Kalign [97]. The resulting alignment was used to produce a maximum likelihood tree using MEGA5 with 500 bootstraps [91].
Allele information derived from MLST analysis was used to construct paired allele networks, where alleles present in combination were connected with a solid line. The presence of all four combinations of two alleles was represented as an hourglass shape, with the lines darkened to indicate evidence for recombination [22]. This analysis relies on the assumption that homoplasy is a rare event.
MATa and MATα isolates were grown on YPD agar at room temperature for two days. Cells were scraped off YPD agar, washed, and suspended in autoclaved ddH2O. Cells were spotted alone or in combination (MATα and MATa) on V8 pH = 5 and V8 pH = 7 mating agar and incubated in the dark at room temperature. Plates were checked once a week for 24 weeks for hyphae and/or the formation of basidia and basidiospores. To examine mating reactions for morphological features associated with mating, scanning electron microscopy studies were conducted. One centimeter square blocks were excised from the agar plates and fixed in 2% glutaraldehyde (Electron Microscopy Sciences, EMS, Hatfield, PA, USA) with 0.05% malachite green oxalate (EMS) in 0.1 M sodium cacodylate buffer and incubated at 4°C until further processing. The fixation buffer was removed, blocks were dehydrated by ethanol series, critical point dried (Pelco CPD2, Ted Pella, Inc., Redding, California, USA), sputter coated, and imaged with the FEI XL30 SEM-FEG (FEI Company, Hillsboro, Oregon, USA) at the electron microscopy facility at North Carolina State University.
Intracellular proliferation rate (IPR) was determined as previously described utilizing J774 macrophages [21], [33], [34]. Macrophages were co-incubated for 2 hours with opsonized cryptococcal cells (18B7 antibody) as described previously [21], [34]. Wells were washed with phosphate-buffered saline (PBS) four times to remove excess extracellular yeast cells and 1 ml of fresh serum-free DMEM was then added. For the control time point T = 0, the DMEM was discarded and 200 µl of sterile ddH2O was added to lyse the macrophages. After 30 minutes, lysed macrophages released intracellular yeast cells, which were then collected in an additional 200 µl ddH2O. For the additional time points T = 18 hrs and 24 hrs, intracellular cryptococcal cells were collected and independently counted with a hemocytometer.
The IPR assay was replicated at least three times for each strain tested using independently propagated batches of macrophages. The IPR value was calculated by dividing the maximum intracellular yeast number by the initial intracellular yeast number at T = 0. We confirmed that Trypan Blue stains 100% of the cryptococcal cells in a heat-killed culture, but only approximately 5% of cells from a standard overnight culture. Compared to a conventional colony counting method, this method was shown to be more sensitive in detecting the clustered yeast population or yeast cells undergoing budding.
Tubularized mitochondria were determined as previously described [33], [34]. C. gattii cells were grown overnight at 37°C in DMEM, harvested, washed twice in PBS, and re-suspended in PBS plus Mito-Tracker Red CMXRos (Invitrogen) at a final concentration of 20 nM for 15 min at 37°C. Cells were washed three times and re-suspended in PBS. 100 yeast cells in three replicates per strain were randomly chosen, imaged using a Zeiss Axiovert 135 TV microscope with a 100X oil immersion Plan-Neofluor objective or a Nikon Eclipse Ti Plan Apo VC 60X oil immersion objective, images were collected, and percent tubularized mitochondria were calculated. All images were processed identically in ImageJ and mitochondrial morphologies were analyzed and counted blindly. IPR and tubularization data were analyzed for statistically significant differences using one-way ANOVA analysis with multiple comparisons by Tukey's Honestly Significant Difference (HSD) posthoc test. A p-value of <0.05 after controlling for multiplicity was considered to be statistically significant. IPR and percent yeast cells with tubular mitochondria were plotted and linear regression, residual plot, and R2 analysis was completed with GraphPad Prism version 6.03 (Windows, GraphPad Software, La Jolla California USA, www.graphpad.com.).
Six-week-old female A/JCr mice (Cat. No. 01A24, NCI-Frederick) or male BALB/c mice (Cat. No. 01B05, NCI-Frederick) were used. Mice were acclimatized in the facility for one week prior to infection by intranasal instillation and were housed in cages at 21°C and 50% humidity with a 12 hr light/12 hr dark cycle. Cells were grown in YPD broth with two successive passages and collected by centrifugation, washed with autoclaved ddH2O, and suspended in autoclaved ddH2O. Mice were inoculated intranasally with 106 cells in 40 µl. At the first signs of poor health or discomfort, mice were euthanized with CO2. Kaplan-Meier survival curves were constructed by GraphPad Prism version 6.03 (Windows, GraphPad Software, La Jolla California USA, www.graphpad.com.). Additional data on colonization was obtained from lung, brain, and spleen tissues postmortem. Tissues were aseptically harvested postmortem. Tissue homogenates were serially diluted and plated on YPD agar, incubated at 30°C for 2 to 3 days, and colony forming units (CFUs) were determined. Two mice were chosen at random from each treatment group, and lung, spleen, or brain tissues were fixed in 10% buffered formalin (lung and spleen tissues) or Bouin's fixative (brain tissues), processed into paraffin blocks, sectioned, and stained with hematoxylin and eosin (H & E) and Mayer's mucicarmine for histopathological examination.
All animal studies were conducted in the Division of Laboratory Animal Resources (DLAR) facilities at Duke University Medical Center (DUMC) and animals were handled according to the guidelines defined by the United States Animal Welfare Act and in full compliance with the DUMC Institutional Animal Care Use Committee (IACUC). Animal models were reviewed and approved by DUMC IACUC under IACUC protocol # A217-11-08.
C. gattii isolates were grown and passaged two times in YPD broth at 30°C. Cells were collected by centrifugation and washed and suspended in 0.85% NaCl to an OD600 = 2 (equivalent to 1 McFarland turbidity) following BioMerieux recommended protocol (http://www.biomerieux-diagnostics.com). Two RPMI agar plates (RPMI 1640+MOPS+2% Glucose+1.5% agar) per strain were swabbed in three contrasting directions and allowed to dry at room temperature. Amphotericin B, fluconazole, flucytosine, and ketoconazole BioMerieux Etest strips were place on plates in pairs, and incubated 48–72 hours at 35°C with 5% CO2 and MIC values were reported as indicated in the Etest instructions.
|
10.1371/journal.ppat.0030120 | Serum Amyloid P Aids Complement-Mediated Immunity to Streptococcus pneumoniae | The physiological functions of the acute phase protein serum amyloid P (SAP) component are not well defined, although they are likely to be important, as no natural state of SAP deficiency has been reported. We have investigated the role of SAP for innate immunity to the important human pathogen Streptococcus pneumoniae. Using flow cytometry assays, we show that SAP binds to S. pneumoniae, increases classical pathway–dependent deposition of complement on the bacteria, and improves the efficiency of phagocytosis. As a consequence, in mouse models of infection, mice genetically engineered to be SAP-deficient had an impaired early inflammatory response to S. pneumoniae pneumonia and were unable to control bacterial replication, leading to the rapid development of fatal infection. Complement deposition, phagocytosis, and control of S. pneumoniae pneumonia were all improved by complementation with human SAP. These results demonstrate a novel and physiologically significant role for SAP for complement-mediated immunity against an important bacterial pathogen, and provide further evidence for the importance of the classical complement pathway for innate immunity.
| Serum amyloid P (SAP) is a protein that is found in high concentrations in the blood, the exact function(s) of which are not clear. However, no known natural state of SAP deficiency has been identified, which suggests that SAP does have a vital role in human health. SAP can bind to molecular patterns found on the surface of bacteria, and it has been proposed that this may mark bacteria for attack by the immune system. We have investigated whether SAP helps protect against an important bacterial pathogen, Streptococcus pneumoniae. We show that SAP binds to different strains of S. pneumoniae, and that this leads to activation of an important component of the immune response called the complement system. Complement is particularly important for defence against S. pneumoniae infections, and using animal models of infection, we demonstrate that loss of SAP makes mice more susceptible to S. pneumoniae pneumonia. These results suggest that SAP helps the immune system to recognise invasion by bacteria and describe a new mechanism required for control of S. pneumoniae infections. This study may help the design of new therapeutic strategies to prevent or treat important bacterial diseases.
| The pentraxin serum amyloid P (SAP) is a glycoprotein that is a major constituent of human serum, present in concentrations of about 30–50 μg/ml. Pentraxin proteins are distinguished by their pentameric assembly and calcium-dependent ligand binding, and include another important serum protein, C reactive protein (CRP). Pentraxins are components of the acute phase response, with serum levels of SAP increasing markedly in mice during sepsis [1,2]. No known natural state of SAP deficiency has been identified, suggesting that SAP has a vital role in human health, but the exact function(s) of SAP are ill-defined. SAP binds to DNA, chromatin, and apoptotic cells, and is thought to aid their clearance [3,4]. However, by binding to amyloid fibrils and stabilising their structure, SAP also promotes amyloid persistence and is therefore an important component of the disease amyloidosis [5–7]. Interactions of SAP with the immune system have also been described, the physiological relevance of which is not clear. These include binding to the complement factor C1q and preventing the inhibitory function of the complement regulatory component C4 binding protein (C4BP), both of which can lead to activation of the classical pathway of complement [8–11], and improving Fcγ receptor–mediated phagocytosis of zymosan and apoptotic cells [12–15]. In addition, SAP can bind to structures found on microbial surfaces, including lipopolysaccharide (LPS), phosphorylcholine (PC), and terminal mannose or galactose glycan residues [16–20]. As a consequence, SAP can bind to a range of microbes, including Gram-positive and Gram-negative bacterial pathogens and human influenza A virus [21,22].
These data suggest that SAP might act as a pathogen recognition receptor and assist innate immunity to microbial pathogens, which is analogous to the known role of CRP, to which SAP has 51% homology at the amino acid level [19,23]. Several studies have investigated the potential role of SAP for host immunity, but have produced conflicting results. SAP enhances killing of Listeria monocytogenes by macrophages without affecting phagocytosis [24], inhibits the growth of intra-erythrocyte malaria parasites [25], reduces uptake of Mycobacterium tuberculosis by macrophages [26,27], and prevents influenza A infection of cell cultures [22], phenotypes which could improve immunity to these pathogens. However, data from experimental infections in SAP-deficient mice have shown that SAP has little effect on immunity to influenza A [28]. Furthermore, SAP prevents classical pathway complement deposition and phagocytosis of rough strains of Escherichia coli, and SAP-deficient mice are protected against infection with rough strains of E. coli and Streptococcus pyogenes [29,30]. In contrast, SAP does not bind to or affect phagocytosis of a smooth strain of E. coli, and SAP-deficient mice had increased susceptibility to this E. coli strain by unknown mechanisms [29,30]. At present, whether SAP aids immunity to a common human pathogen has not been clearly demonstrated.
One human pathogen of major importance worldwide is Streptococcus pneumoniae. S. pneumoniae is the second most common cause of death due to bacterial infection, is responsible for the majority of cases of pneumonia, and is a significant cause of septicaemia and meningitis in both infants and adults [31–33]. The high mortality of severe S. pneumoniae infections (over 20% even if treated with appropriate antibiotics) and the spread of antibiotic resistance amongst clinical strains underline the importance of understanding host immunity to S. pneumoniae. Experimental and human data have convincingly demonstrated the essential role of the complement system for preventing S. pneumoniae infections and for controlling replication of S. pneumoniae within the lungs and the systemic circulation [34–36]. We have previously reported that, in contrast to S. pyogenes, even in the absence of specific acquired antibody the classical pathway is the most important complement pathway for innate immunity to S. pneumoniae [34]. The mechanisms by which the classical pathway is activated by S. pneumoniae infection include recognition of PC on the bacterial surface by natural IgM and CRP [34,37–40], and binding of bacterial capsular polysaccharide to the lectin SIGN-R1 expressed on marginal zone macrophages within the spleen [41]. As SAP also binds to PC and is known to interact with the classical pathway, we hypothesised that SAP could be an additional mediator of classical pathway activity against S. pneumoniae and contribute towards innate immunity to this important pathogen. In this study, we investigated the role of SAP for immunity to S. pneumoniae, in particular its role during complement activation, phagocytosis, and infection in mouse models of disease using mice genetically engineered to be deficient in SAP.
Whole-cell ELISAs were used to investigate whether human SAP (hSAP) binds to three S. pneumoniae strains representing capsular serotypes (STs) 2, 23F, and 4 using the smooth E. coli O111:B4 strain, which is known not to bind to SAP [30], as a negative control. Dose-dependent binding to purified hSAP was demonstrated to all three S. pneumoniae strains (Figure 1A). To confirm these results for live bacteria in a physiological medium, hSAP binding to S. pneumoniae incubated in human serum was assessed by a flow cytometry assay using the S. pyogenes H372 strain and the E. coli O111:B4 as positive and negative controls, respectively. For all three S. pneumoniae strains and S. pyogenes, a significant proportion of bacteria were positive for hSAP (ST2 38% standard deviation [SD] 10, ST4 32% SD 8.3, ST23F 39% SD 12, H372 50% SD 2.2), whereas the E. coli O111:B4 strain showed no significant binding to hSAP (6.7% SD 4.0) (Figure 1B). Addition of EDTA reduced the proportion of the ST2 S. pneumoniae strain positive for SAP (Figure 1C), demonstrating that SAP binding was calcium dependent. Furthermore, the presence of PC inhibited binding of both SAP and CRP (used as a positive control, as CRP is known to bind to S. pneumoniae cell wall PC) to a similar degree, suggesting that PC is a major ligand for SAP binding to S. pneumoniae (Figure 1D and 1E). Significant levels of SAP binding still occurred in the presence of PC or high concentrations of EDTA, and this might reflect some non-specific binding of SAP to S. pneumoniae. These data show that SAP can bind to a range of S. pneumoniae strains, and therefore could potentially act as a pathogen recognition receptor and mediate innate immune responses to this pathogen.
To determine whether SAP affects complement activation by S. pneumoniae, C3b deposition on S. pneumoniae incubated in serum from wild-type or mice genetically engineered to be deficient in SAP (Apcs−/− mice) was analysed using a flow cytometry assay. Starting 1 min after incubation in serum, C3b deposition on the ST2 S. pneumoniae strain was strongly impaired in serum from Apcs−/− mice at all time points compared to the results for serum from wild-type mice (Figure 2A and 2D). The proportion of bacteria positive for C3b after incubation in serum from Apcs−/− mice was also reduced for the ST4 and ST23F strains (Figure 2B, 2C, and 2E). The effect of SAP deficiency on C3b deposition varied between the three S. pneumoniae strains investigated, with a very marked effect for the ST23F strain, a relatively weak effect for the ST4 strain, and an intermediate effect on the ST2 strain. Backcrossing of SAP-deficient Apcs−/− animals onto the C57BL/6 genetic background results in the translocation of surrounding chromosomal DNA from the mouse 129 strain into the C57BL/6 strain, and this combination has been shown to explain some of the phenotypes associated with Apcs−/− mice [12]. However, C3b deposition on the ST2 strain in serum from congenic C57BL/6 mice engineered to carry a similar 129 fragment but no deletion of the SAP gene (C57BL/6.129(D1Mit105–223)) [42] was identical to C3b deposition in serum from wild-type C57BL/6 mice (unpublished data), demonstrating that this genetic combination was not responsible for the impaired C3b deposition detected in serum from Apcs−/− mice. Furthermore, C3 deposition on the ST2 strain in Apcs−/− serum was partially restored by addition of serum from wild-type mice and by the addition of exogenous hSAP (Figure 2F and 2G). These results confirm that the reduced complement deposition on S. pneumoniae in serum from Apcs−/− mice is due to loss of SAP, and suggest that hSAP has a similar functional effect on complement activation by S. pneumoniae as mouse SAP.
As SAP has been shown to bind to C1q [9], the first component of the classical pathway, one mechanism by which the SAP may aid C3b deposition on S. pneumoniae could involve increasing C1q binding to the bacteria and thereby activating the classical pathway. To investigate this possibility, the three serotypes of S. pneumoniae were incubated with physiological concentrations of purified human C1q protein and hSAP protein, and the deposition of C1q on the bacteria analysed using a flow cytometry assay. For all three strains, the presence of hSAP increased the deposition of hC1q on the S. pneumoniae surface (Figure 3A–3D). Furthermore, addition of exogenous hSAP to human sera increased C1q deposition on all three S. pneumoniae serotypes (Figure 3E and 3F). To confirm that the effects of SAP on C3b deposition are mainly mediated by increased C1q deposition activating the classical pathway, C3b deposition assays were repeated using serum from mice with deficiencies of both SAP and C1q (Apcs−/−.C1qa−/−) with and without addition of exogenous hSAP. In contrast to the results for Apcs−/− serum, addition of hSAP to Apcs−/−.C1qa−/− serum had no effect on C3b deposition on the ST2 S. pneumoniae strain (Figure 2H). These data suggest that SAP-mediated C3b deposition on S. pneumoniae is C1q and therefore classical pathway dependent.
To study the functional consequences of reduced complement deposition on S. pneumoniae in Apcs−/− serum, we analysed phagocytosis of S. pneumoniae in serum from Apcs−/− mice using flow cytometry to assess the proportion of fluorescent bacteria associated with HL60 cells, a human neutrophil cell line. In this assay, the association of fluorescent bacteria with phagocytes is mainly due to phagocytosis of the bacteria rather than simple binding to the cell surface [43]. Phagocytosis of all three S. pneumoniae strains was mainly serum dependent, with only low levels of uptake after incubation in Hank's Balanced Salt Solution (HBSS) medium alone (Figure 4). Phagocytosis was consistently reduced when the bacteria were incubated with serum from Apcs−/− mice compared to serum from wild-type mice for the three S. pneumoniae strains investigated (Figure 4). The level of phagocytosis was partially restored in serum from Apcs−/− mice when mixed with serum from wild-type mice or by addition of exogenous hSAP (Figure 4A and 4E). However, addition of hSAP to HBSS alone or to Apcs−/−.C1qa−/− serum did not stimulate phagocytosis of S. pneumoniae, suggesting that the effects of hSAP are dependent on the classical pathway (Figure 4E and 4F). These results are consistent with the results of the C3b deposition assays, and indicate that reduced opsonisation of S. pneumoniae with C3b in Apcs−/− serum is associated with a reduced efficiency of phagocytosis.
Clearance of S. pneumoniae from the circulation is thought to be dependent on complement and on phagocytosis by the reticuloendothelial system [44]. To test whether the effects of SAP deficiency on complement deposition and phagocytosis of S. pneumoniae result in an impaired ability to clear bacteria from the blood, wild-type and Apcs−/− mice were inoculated by i.v. injection with 2 × 105 cfu of the ST2 strain D39, and bacterial cfu in the blood calculated by serial dilutions at 2 and 4 h post-inoculation and in spleen homogenates 4 h post-inoculation (Figure 5). Apcs−/− mice had between 2 and 3 logs greater cfu in both the blood and spleen compared to wild-type mice, demonstrating that Apcs−/− mice have a marked impairment in their ability to clear S. pneumoniae from the systemic circulation consistent with the impaired phagocytosis of S. pneumoniae found in SAP-deficient serum.
A mouse model of pneumonia was used to confirm a biological role of SAP for innate immunity to S. pneumoniae. C57BL/6 mice are partially susceptible to the ST2 S. pneumoniae strain D39 after intranasal (i.n.) inoculation and therefore provide a sensitive model for identifying immunological defects that result in increased susceptibility [34]. Groups of wild-type and Apcs−/− C57BL/6 mice were inoculated i.n. (to mimic the natural route of infection) with 1 × 106 cfu of D39 of S. pneumoniae and the development of lethal infection monitored. Lethal disease developed faster in Apcs−/− mice, with a median time to fatal infection of 64 h (interquartile range [IQR] 52 to 68 h) compared to 86 h for wild-type mice (IQR 77 to 92 h), and all the Apcs−/− mice developed lethal infection whilst 33% of wild-type mice survived (Figure 6A). Plating of aliquots of blood obtained from the tail veins of mice 48 h after i.n. inoculation demonstrated that all the Apcs−/− mice had large numbers of bacteria in the blood, whereas only 44% of wild-type mice had detectable septicaemia (Figure 6B). Hence, Apcs−/− mice are more susceptible to S. pneumoniae pneumonia, and this increased susceptibility is associated with increased levels of bacteria infection in the blood.
To characterise the role of SAP during innate immunity in more detail, groups of wild-type and Apcs−/− mice were culled 4 and 24 h after i.n. inoculation with 1 × 106 cfu of D39 and the number of bacteria present in target organs calculated by plating serial dilutions of bronchoalveolar fluid (BALF), lung and spleen homogenates, and the blood. After 4 h of infection, there were slightly higher levels of S. pneumoniae cfu in BALF and lung from Apcs−/− mice compared to wild-type mice (Table 1). By 24 h after inoculation, Apcs−/− mice had over 1 log greater bacterial cfu in the BALF, 2 logs in lung homogenates, and 4 logs in the blood compared to wild-type mice (Table 1), demonstrating that Apcs−/− mice were unable to control bacterial replication within the lung and systemic circulation. To ensure that chromosomal translocation of 129 DNA surrounding the SAP gene into the C57BL/6 mouse background was not responsible for the increased susceptibility of the Apcs−/− mice to S. pneumoniae infection, experiments were repeated using the C57BL/6.129(D1Mit105–223) congenic mice and C57BL/6 animals. No differences in bacterial cfu were identified between these strains 24 h after inoculation of ST2 S. pneumoniae, indicating that SAP deficiency is likely to be responsible for the phenotype seen in Apcs−/− mice (unpublished data). To further link the observed increased susceptibility of Apcs−/− mice to S. pneumoniae pneumonia to deficiency in SAP, Apcs−/− mice were supplemented by tail vein injection with 5 mg/kg of hSAP 1 h prior to inoculation with D39, and the bacterial cfu in target organs obtained at 24 h. Although there was wide variation in the numbers of bacteria recovered between mice, for all target organs the median S. pneumoniae cfu recovered from Apcs−/− mice complemented with hSAP were similar to those for wild-type mice and 1 to 4 log fewer than the median cfu recovered from Apcs−/− mice given PBS alone (Table 1).
As well as opsonising bacteria, activation of the complement system stimulates pro-inflammatory responses to infection. To investigate the effect of SAP deficiency on the inflammatory response to S. pneumoniae pneumonia, the levels of pro-inflammatory cytokines were measured in BALF from wild-type and Apcs−/− mice 4 h and 24 h after i.n. inoculation with 1 × 106 cfu of the S. pneumoniae D39 strain. At 4 h after inoculation, despite the slightly higher numbers of bacterial cfu in BALF and lungs of Apcs−/− mice (Table 1), the levels of TNF-α and IL-6 were lower in Apcs−/− mice than in wild-type mice (Figure 7A), suggesting that at this early stage of infection Apcs−/− mice had an impaired inflammatory cytokine response to S. pneumoniae pneumonia. Levels of IL-12, IL-10, MCP-1, or IFN-γ in BALF at 4 h were very low or undetectable (unpublished data). By 24 h, at which time point Apcs−/− mice had considerably greater numbers of S. pneumoniae within target organs (Table 1), the levels of IL-6, IL-12, TNF-α, and MCP-1 in BALF were raised in Apcs−/− mice compared to those of wild-type mice (Figure 7B). The levels of IL-10 and IFN-γ in BALF at 24 h remained very low (unpublished data). The consequences of differences in inflammatory cytokines between Apcs−/− and wild-type mice were assessed by scoring the level of inflammation in lung sections. There were no significant differences in the score for the degree of histological inflammation of the lungs 4 h after inoculation (a median score of 15 for both Apcs−/− and wild-type mice), and although there was an increase in the inflammation score 24 h after infection in the lungs from Apcs−/− mice, with a median score of 70 (IQR 60–80) compared to a median score of 30 (IQR 19–68) for wild-type mice, this did not reach statistical significance (p = 0.48). Overall, these results suggest that during the early stages of S. pneumoniae pneumonia, there is a more pronounced pro-inflammatory response in wild-type mice compared to Apcs−/− mice. However, at later stages of infection, when there is considerably greater bacterial cfu in the target organs of Apcs−/− mice (Table 1), Apcs−/− mice have more pronounced levels of inflammation than wild-type mice.
The pentraxin SAP is an abundant plasma protein in both humans and mice, but its physiological role is not fully understood. By analogy to the related proteins CRP, which is known to mediate complement-dependent immunity [38,39,45], and Pentraxin3 [46], a role for SAP in innate immunity has been suggested. This possibility is supported by data demonstrating that SAP can bind to pathogen-associated structures such as PC and LPS [16,17,20]. Furthermore, SAP may interact with the classical pathway component C1q through its collagen binding site, and possibly stimulates phagocytosis through Fcγ receptors [9,10,13,14]. However, the physiological relevance of these observations is unclear, and although in vitro phenotypes associated with hSAP suggest it may protect against a variety of pathogens, including tuberculosis, malaria, or influenza A [22,25–27], other authors have shown that in mice SAP actually aids the virulence of S. pyogenes and E. coli, possibly by preventing classical pathway–mediated complement activity and phagocytosis [29,30].
Using SAP-deficient mice, we have investigated the biological role of SAP during infection by the Gram-positive pathogen S. pneumoniae. We have previously shown that the classical pathway is vital for innate immunity to S. pneumoniae, partially through recognition of S. pneumoniae by natural IgM [34]. However, natural IgM-deficient mice were markedly less susceptible to S. pneumoniae infection than C1q-deficient mice [34], suggesting there are other mediators of classical pathway activity against S. pneumoniae. The lectin SIGN-R1 has recently been shown to be one such mediator, with binding of the S. pneumoniae capsule to SIGN-R1 resulting in activation of the classical pathway [41]. However, although SIGN-R1-deficient mice have an increased susceptibility to S. pneumoniae infection, like natural IgM mice they are more resistant than C1q-deficient mice [41], indicating that additional molecules may contribute to complement activation. CRP is also thought to bind to S. pneumoniae and activate the classical pathway [40,47], but is present only in low levels in mice and therefore probably does not contribute strongly to innate immunity in the mouse models of S. pneumoniae infection. As SAP from different mammalian species can bind both to PC and C1q [9,10,16,17], we hypothesised that SAP could also contribute to the activation of the classical pathway by S. pneumoniae, and therefore aid both innate and acquired immunity to this important pathogen. This hypothesis is supported by our data showing binding of hSAP to three different capsular serotypes of S. pneumoniae, and impaired C3b deposition on these S. pneumoniae strains in Apcs−/− serum compared to serum from wild-type mice. Furthermore, we have shown that the binding of human C1q to S. pneumoniae is increased by the presence of hSAP and that the effects of SAP on complement are dependent on an intact classical pathway. The reduced complement activity in SAP-deficient serum versus S. pneumoniae results in an impaired ability of SAP-deficient mice to control S. pneumoniae replication within both the lungs and the bloodstream, leading to uncontrolled infection in a mouse model of pneumonia. This phenotype is very similar to that seen in mice deficient in natural IgM [34], and the data support the hypothesis that SAP and natural IgM both contribute, along with SIGN-R1 and CRP, towards classical pathway–mediated immunity to S. pneumoniae. Previous reports that SAP does not aid immunity to S. pneumoniae after intravenous (i.v.) inoculation [48] used wild-type mice infected with bacteria that had been incubated with SAP rather than SAP-deficient animals, and this model was therefore probably too insensitive to identify the effects we have shown for SAP.
Reduced complement activity results in increased susceptibility to infection by impairing C3b-mediated clearance of bacteria by phagocytes and/or by decreasing complement-mediated inflammatory responses to infection [44,49,50]. Our results suggest that both mechanisms could affect the susceptibility of Apcs−/− mice to S. pneumoniae. Uptake of the three different S. pneumoniae strains by a neutrophil-like cell line was impaired in Apcs−/− serum, and clearance of S. pneumoniae from the systemic circulation after i.v. inoculation, which is mainly dependent on phagocytosis by the reticuloendothelial system [51], was markedly reduced in Apcs−/− mice. These data suggest that by reducing opsonisation of S. pneumoniae with C3b, SAP deficiency results in impaired phagocytosis. The effects of SAP on phagocytosis of S. pneumoniae were serum- and classical pathway–dependent, with no stimulation of phagocytosis when SAP was added to medium alone or serum deficient in C1q, indicating that SAP assisted phagocytosis through classical pathway activity and not by direct binding to Fcγ receptors. In addition, we found that in BALF from Apcs−/− mice obtained at an early stage of infection there were lower levels of the pro-inflammatory cytokines TNF-α and IL-6 despite containing slightly greater numbers of cfu than wild-type mice in BALF and lung homogenates at this stage. Hence, the early pro-inflammatory response to S. pneumoniae pneumonia was impaired in Apcs−/− mice, and this may contribute to the increased susceptibility of these mice. Whether the reduced inflammatory responses in Apcs−/− mice are due to loss of direct effects of SAP on modulating the inflammatory response, or is secondary to reduced complement activation and phagocytosis, requires further evaluation. The increased inflammation in Apcs−/− mice compared to wild-type mice at 24 h probably reflects the overwhelming infection present in Apcs−/− mice at this stage rather than direct effects of SAP deficiency.
Although human and murine SAP have a high degree of homology at the amino acid level and both bind to PC, they do have some differences in their structure and interactions with other proteins [19]. Furthermore, CRP is the major component of the acute phase response in humans and SAP, although present in high concentrations in human sera, is the major acute phase response protein in mice [2]. Hence, to identify the possible human relevance of results obtained with Apcs−/− mice, we have complemented our assays using hSAP. Complementation of Apcs−/− serum with hSAP restored C3b deposition on bacteria and phagocytosis close to the levels seen with wild-type serum, and administration of hSAP before infection with S. pneumoniae increased the resistance of Apcs−/− mice to infection. Furthermore, the evidence for SAP-dependent C1q deposition on S. pneumoniae was obtained with human reagents. These data suggest that SAP is also important for classical pathway–mediated host immunity to S. pneumoniae in humans as well as mice. As S. pneumoniae is one of the commonest causes of infant mortality in the developing world [32], a role for SAP in preventing serious S. pneumoniae infections helps explain why there is no natural state of SAP deficiency.
In contrast to our results with S. pneumoniae pneumonia, Apcs−/− mice are protected against infection with S. pyogenes inoculated intraperitoneally despite a high level of SAP binding to the bacteria [30]. A possible explanation for the differences in the effect of SAP on immunity between these related pathogens could be differences in their interaction with the complement system. We have previously shown that the classical pathway is the dominant pathway for innate immunity to S. pneumoniae [34], whereas the alternative pathway is more important for innate immunity to S. pyogenes [52]. As the classical pathway does not contribute strongly towards complement activation by most strains of S. pyogenes [52], SAP may not be able to aid innate immunity to this pathogen. In addition, S. pyogenes does not express PC on its surface, and to which bacterial surface structure SAP binds may influence its functional role and interactions with complement factors. For example, de Haas et al. have reported that in direct contrast to the results presented in this manuscript, binding of SAP to the LPS expressed by some E. coli strains inhibits classical pathway–mediated complement activity, perhaps by preventing direct binding of C1q to LPS [29]. Further research is required to identify whether SAP mediates complement-dependent immunity to other important pathogens, and to determine why SAP has contrasting effects on susceptibility to closely related pathogens such as S. pneumoniae and S. pyogenes.
In summary, we have demonstrated that SAP aids complement activity against S. pneumoniae and is an important component of the innate immune response to this pathogen. To our knowledge, this is the first report demonstrating a positive role for SAP in complement-mediated immunity to a microbial pathogen. The data make a significant contribution to our understanding of the biological role of SAP and to our knowledge of the complex mechanisms leading to activation of complement by S. pneumoniae.
S. pneumoniae strains belonging to capsular STs 2 (D39), 4 (JSB4, previously M313), and 23F (JSB23F, previously Io11697) were used for the majority of the studies [53]. The E. coli strain O111:B4 and the S. pyogenes strain H372 were used for SAP binding assays [30,52]. S. pneumoniae and S. pyogenes strains were cultured in Todd-Hewitt broth supplemented with 0.5% yeast extract (Oxoid, http://www.oxoid.com/) while E. coli O111:B4 was cultured in LB broth (Oxoid). All strains were grown to an optical density (OD580) of 0.4 (corresponding to about 108 cfu/ml) and stored at −70 °C in 10% glycerol as single-use aliquots.
SAP binding to S. pneumoniae was analysed by whole-cell ELISA as previously described [54]. Briefly, bacterial cultures from late log phase were resuspended in PBS to an OD550 of 1.0, 200 μl of this suspension added to each well of 96-well plates (Nunc MaxiSorp, http://www.nuncbrand.com/), air dried at room temperature, and blocked with 200 μl of PBS-0.5 % BSA-NaN3 for 1 h before 50 μl of different concentrations of hSAP (Calbiochem, http://www.emdbiosciences.com/html/CBC/home.html) were added to each well. After incubation overnight at 4 °C, the plates were incubated with 50 μl of rabbit anti-human SAP (Calbiochem) diluted 1/2000 for 5 h at 4 °C, incubated overnight with 50 μl of goat anti-rabbit AP (Sigma, http://www.sigmaaldrich.com/) diluted 1/1000, and developed using FAST p-nitrophenyl phosphate (Sigma) for 30 min before determining the OD405 using a microtiter plate reader (Multiskam ACC/340; Titertek, http://www.titertek.com/).
C3b deposition on S. pneumoniae was measured using a flow cytometry assay as previously described [34,52]. Briefly, 107 cfu of S. pneumoniae were incubated with 10 μl of serum from wild-type or Apcs−/− mice and bacteria coated with C3b identified using a FITC-goat anti mouse C3 antibody and flow cytometry. CRP, SAP, and C1q binding assays were performed by a similar assay using either rabbit anti-human SAP or CRP (Calbiochem, with an appropriate FITC labelled secondary antibody) or FITC sheep anti-human C1q antibody (Serotec, http://www.ab-direct.com/), and incubating at 37 °C S. pneumoniae for 1 h with serum with or without addition of EDTA or PC (Sigma), 10 or 50 μg/ml hSAP (Calbiochem), and/or human C1q protein (Calbiochem). C3 levels (measured by ELISA) in the serum of Apcs−/− and wild-type mice were similar at 332 mg/l SD 101.7 (n = 40) for Apcs−/− mice, and 357 mg/l SD 123.5 (n = 20) for C57B/6 mice. Serum deficient in both C1q and SAP was obtained from Apcs−/−.C1qa−/− mice created by interbreeding the previously described Apcs−/− and C1qa−/− mouse strains [5,55]. C1q binding assays were also performed in human serum depleted in C1q (Calbiochem) with or without addition of 50 μg/ml of hSAP protein, using a serum that had been treated similarly by depletion of a terminal complement pathway component (C9, Calbiochem) to represent normal serum.
Phagocytosis of S. pneumoniae in serum from Apcs−/−, Apcs−/−.C1qa−/−, and wild-type mice was investigated using a previously described flow cytometry assay and the human tissue culture cell line HL-60 (promyelocytic leukemia cells; CCL240; American Type Culture Collection, http://www.atcc.org/) differentiated into granulocytes [53,56]. S. pneumoniae were fluorescently labelled with 5,6-carboxyfluorescein succinimidyl ester (FAM-SE; Molecular Probes, http://probes.invitrogen.com/) as described and stored at −70 °C in 10% glycerol as single-use aliquots (7 × 108 cfu/ml). FAM-SE-labelled bacteria (106 cfu) were opsonised with 10 μl of dilutions of serum obtained from wild-type or Apcs−/− mice in a 96-well plate for 20 min at 37 °C with horizontal shaking (150 rpm). HL-60 cells (105) were added to each well and incubated for 30 min at 37 °C, fixed with 3% PFA, and analysed using a FACScalibur flow cytometer (minimum of 6,000 cells per sample) to identify the proportion of cells associated with fluorescent bacteria as a marker of phagocytosis [43].
Wild-type C57BL/6 mice were purchased from commercial breeders, and Apcs−/− and C57BL/6.129(D1Mit105–223) congenic mice were bred in-house by one of the authors (MB) [5,42]. All mice used were 8–16 wk old, and within each experiment groups of mice were matched for age and sex. Studies were performed according to UK Home Office and university guidelines for animal use and care. Mice were inoculated i.n. (under halothane anesthesia, 1 × 106 cfu/mouse) or intravenously (1 × 106 cfu/mouse) with the S. pneumoniae D39 strain appropriately diluted in PBS. For survival studies, mice were killed when they exhibited signs of severe disease from which recovery was unlikely [57]. For experiments to test the number of cfu in different target organs or to perform immunological analysis, target organs were recovered 4 and 24 h after inoculation as previously described [34]. Bacterial counts were calculated by plating serial dilutions of the homogenised organs suspensions, blood, and BALF onto blood agar and incubated at 37 °C in 5% CO2.
The levels of inflammatory cytokines and chemokines (IL-6, IL-10, MCP-1, IFN-γ, TNF-α, and IL-12p70) in BALF were analysed by flow cytometry in 50 μl of pooled BALF from wild-type or Apcs−/− mice using the Mouse Inflammation Cytometric Bead Array kit (Becton Dickinson, http://www.bdbiosciences.com/) according to manufacturer protocols and using the BD CBA software [52]. For the histological analysis of inflammation, in a proportion of infection experiments the left lung was fixed in 4% neutral buffered formalin, processed to paraffin wax, and stained with haematoxylin and eosin. Inflammation was assessed using a simplified score based on a previously described scoring system for inflammation during S. pneumoniae pneumonia [58]. The extent of lung involvement was estimated by examining lung cross sections at ×10 magnification. The degree of inflammation for six fields was scored at ×200 magnification as 1 (no visible inflammatory change), 2 (minimal swelling of alveolar walls with slight change in architecture), 3 (increased swelling with presence of erythrocytes and inflammatory cells and an increase in type II pneumocytes), and 4 (considerable haemorrhage with inflammatory cell influx, widespread alveolar disorganisation with interstitial swelling and pneumocyte proliferation). A total score for each mouse was obtained by multiplying the percentage of involved lung by the mean score for the areas analysed, and data presented as medians with IQRs.
The complement factor binding and opsonophagocytosis data presented are representative of results obtained from several independent experiments. The data for mouse infection experiments are representative of duplicate experiments that gave similar results. The results of C3b deposition, C1q, CRP and SAP binding experiments, cytokine levels, and phagocytosis assays were analysed using 2-tailed t tests. Bacterial cfu recovered from target organs and histology scoring were analysed using the Mann–Whitney U test for non-parametric data. Differences in survival curves between mouse strains were compared using the log rank method. All error bars given on the figures represent standard deviations. |
10.1371/journal.pcbi.0030092 | Predictive Modeling of Signaling Crosstalk during C. elegans Vulval Development | Caenorhabditis elegans vulval development provides an important paradigm for studying the process of cell fate determination and pattern formation during animal development. Although many genes controlling vulval cell fate specification have been identified, how they orchestrate themselves to generate a robust and invariant pattern of cell fates is not yet completely understood. Here, we have developed a dynamic computational model incorporating the current mechanistic understanding of gene interactions during this patterning process. A key feature of our model is the inclusion of multiple modes of crosstalk between the epidermal growth factor receptor (EGFR) and LIN-12/Notch signaling pathways, which together determine the fates of the six vulval precursor cells (VPCs). Computational analysis, using the model-checking technique, provides new biological insights into the regulatory network governing VPC fate specification and predicts novel negative feedback loops. In addition, our analysis shows that most mutations affecting vulval development lead to stable fate patterns in spite of variations in synchronicity between VPCs. Computational searches for the basis of this robustness show that a sequential activation of the EGFR-mediated inductive signaling and LIN-12 / Notch-mediated lateral signaling pathways is key to achieve a stable cell fate pattern. We demonstrate experimentally a time-delay between the activation of the inductive and lateral signaling pathways in wild-type animals and the loss of sequential signaling in mutants showing unstable fate patterns; thus, validating two key predictions provided by our modeling work. The insights gained by our modeling study further substantiate the usefulness of executing and analyzing mechanistic models to investigate complex biological behaviors.
| Systems biology aims to gain a system-level understanding of living systems. To achieve such an understanding, we need to establish the methodologies and techniques to understand biological systems in their full complexity. One such attempt is to use methods designed for the construction and analysis of complex computerized systems to model biological systems. Describing mechanistic models in biology in a dynamic and executable language offers great advantages for representing time and parallelism, which are important features of biological behavior. In addition, automatic analysis methods can be used to ensure the consistency of computational models with biological data on which they are based. We have developed a dynamic computational model describing the current mechanistic understanding of cell fate determination during C. elegans vulval development, which provides an important paradigm for studying animal development. Our model is realistic, reproduces up-to-date experimental observations, allows in silico experimentation, and is analyzable by automatic tools. Analysis of our model provides new insights into the temporal aspects of the cell fate patterning process and predicts new modes of interaction between the signaling pathways involved. These biological insights, which were also validated experimentally, further substantiate the usefulness of dynamic computational models to investigate complex biological behaviors.
| Describing mechanistic models in biology in a formal language, especially one that is dynamic and executable by computer, has recently been shown to have various advantages (see review [1]). A formal language comes with a rigorous semantics that goes beyond the simple positive and negative interaction symbols typically used in biological diagrammatic models. If the language used to formalize the model is intended for describing dynamic processes, the semantics, by its very nature, provides the means for tracing the dynamics of system behavior, which is the ability to run, or execute, the models described therein.
Dynamic models can represent phenomena of importance to biological behaviors that static diagrammatic models cannot represent, such as time and concurrency. In addition, formal verification methods can be used to ensure the consistency of such computational models with the biological data on which they are based [2,3]. It was previously suggested that by formalizing both the experimental observations obtained from a biological system and the mechanisms underlying the system's behaviors, one can formally verify that the mechanistic model reproduces the system's known behavior [3].
Formal models are used in a variety of situations to predict the behavior of real systems and have the advantage that they can be executed by computers; often at a fraction of the cost, time, or resource consumption that the observation of the real system would require. In addition, formal models have the advantage that they can be analyzed by computers. For example, it may be possible to predict, by analyzing a model, that all possible executions will reach a stable state, independent of environment behavior. The result of such an analysis would not be obtainable by executing the real system, no matter for how long or how many times, as there are often infinitely many possible environment behaviors. This process of computational model analysis, in the case of state-based models, is called model checking [4].
Here we follow the idea that model execution and model checking can be used to test a biological hypothesis: if the execution and analysis of the model conform to the experimental observations of the biological system, then the model may correctly represent the mechanism that underlies the system behavior; otherwise, the model needs modification or refinement. Thus, the model can be seen as a “hypothesis,” i.e., an explanation for a biological mechanism and experiments can confirm or falsify the hypothesis.
As part of an ongoing effort to model C. elegans vulval development [3,5], we have previously created a formal dynamic model of vulval cell fate specification based solely on the proposed diagrammatic model of Sternberg and Horvitz from 1989 [6]. Our previous work [3] has demonstrated that state-based mechanistic models are particularly well-suited for capturing the level of understanding obtained using the tools and approaches common in the field of developmental genetics, and that creating such executable biological models is indeed beneficial. Since the original model proposed by Sternberg and Horvitz (1989), our understanding of the molecular pathways governing vulval fate specification has advanced significantly. In particular, several modes of crosstalk and lateral inhibition between the major signaling pathways specifying the vulval cell fates have been discovered. Here, we report on a dynamic computational model of the more sophisticated understanding of vulval cell fate specification that we have today. In addition, we use model checking to test the consistency of the current conceptual model for vulval precursor cells (VPC) fate specification with a large set of observed behaviors and experimental perturbations of the vulval system.
The C. elegans vulva is formed by the descendants of three VPCs that are members of a group of six equivalent VPCs named P3.p–P8.p (Figure 1). Each of the six VPCs is capable of adopting one of three cell fates (termed 1°, 2°, or 3°) [6–8]. The actual fate a VPC adopts depends upon the integration of two opposing signals that each VPC receives: an inductive signal emanating from the gonadal anchor cell (AC) in the form of the LIN-3 epidermal growth factor activates the epidermal growth factor receptor (EGFR)/LET-23 in the VPCs. The inductive signal is transduced downstream of the EGFR/LET-23 by the conserved RAS/MAPK signaling cascade to specify the 1° cell fate. In response to the inductive signal, the VPCs produce a lateral signal that counteracts the inductive AC signal in the neighboring VPCs (lateral inhibition) by inducing the expression of a set of inhibitors of the EGFR/RAS/MAPK pathway collectively termed lst genes (for lateral signal targets) [9–11]. In a second step, the lateral signal induces the 2° cell fate (lateral specification) [12]. The lateral signal is encoded by three functionally redundant members of the Delta/Serrate protein family (dsl-1, apx-1, and lag-2) and transduced by the LIN-12 / Notch receptor [13,14]. Moreover, two functionally redundant inhibitory pathways defined by the synthetic Multivulva (synMuv) genes prevent the surrounding hypodermal syncytium (hyp7) from producing the inductive LIN-3 signal, thus allowing the AC to establish a gradient of inductive LIN-3 signal [15]. The fate of the VPCs is influenced by their relative distance from the AC and the lateral signals between the VPCs. The cell closest to the AC (P6.p) receives most of the inductive signal and adopts the 1° fate. The neighboring cells P5.p and P7.p receive the stronger lateral signal from P6.p and hence adopt the 2° fate. The remaining distal VPCs P3.p, P4.p, and P8.p adopt the 3° fate as they do not receive enough inductive signal or lateral signal; and LIN-3 expression in the hyp7 is repressed by the synMuv genes [8,15–17].
One remarkable feature of vulval fate specification is its absolute precision. Despite the ability of each cell to adopt any of the three cell fates, the pattern of fates adopted by P3.p–P8.p in wild-type animals is always 3°-3°-2°-1°-2°-3°, respectively. This precision is thought to be achieved by multiple modes of crosstalk between the inductive and lateral signaling pathways discovered in recent years [6,9–11,18,19].
Here we use computer simulations and formal verification to investigate whether the known gene interactions are sufficient to produce such patterning precision and to gain insights into the system dynamics. Using the language of Reactive Modules (RM) [20] and the Mocha tool [21], we have constructed a discrete, dynamic, state-based mechanistic model consisting of the key components of the inductive and lateral signaling pathways with their interconnections. By looking analytically at all possible behaviors of a model, we find previously unnoticed dependencies that are present in the data and explained by the model. Specifically, the analysis of our model predicts additional genetic interactions necessary for efficient lateral inhibition and, through the analysis of the behavior of lin-15 mutants, gives new insights into the temporal order of events necessary to achieve a stable pattern of cell fates, which were also validated experimentally.
Models in the language of RM are constructed by defining the objects of the system (these are the modules), and their variables representing semi-independent components of an object. The state of the system is determined by the states of its objects, which in turn are determined by the values of all their variables. Changes in the value of a variable depend on the previous values of the variable and possibly on other variables. A behavior of the system is a sequence of states that the system goes through during execution.
Our model consists of a worm module that comprises an AC module and six identical copies of a VPC module (Figure 2). Additional modules handle the synchronization between VPCs (i.e., the scheduler in Figure 2, which is setting the order of interaction between the VPCs for a particular execution) and manage the initialization of simulations (i.e., the organizer in Figure 2, which is setting the initial conditions for a particular execution). Each VPC module runs its own copy of the same program simultaneously, based on the inputs it receives from its neighboring cells (AC, hyp7, and the adjacent VPCs) (Figure 3). All VPCs begin with the same conditions determined by the genetic background but may receive different levels of inductive signal depending on their distance from the AC.
The AC module contains variables that indicate if the AC is ablated or formed and determine the level of inductive signal sensed by the VPCs according to their distance from the AC. If the AC is ablated, the inductive signal variables in all VPCs are set to the OFF level. If the AC is not ablated, the VPC closest to the AC (P6.p) senses HIGH inductive signal, the next closest (P5.p and P7.p) sense MEDIUM inductive signal, and the farthest (P3.p, P4.p, and P8.p) sense LOW inductive signal (Figure 3).
The VPC module contains variables that represent the behavior of the EGFR/RAS/MAPK pathway, the LIN-12 Notch-mediated lateral signaling pathway, and the lin-15–mediated inhibition of LIN-3 EGF in hyp7 (Figure 4). In addition, there is a variable for each VPC that follows the temporal progress toward fate acquisition. Each of these variables is now described briefly:
The lateral signal variable (LS) can be either ON or OFF. The variable starts as OFF and is turned ON upon activation by the EGFR/RAS/MAPK pathway. Once the lateral signal is ON, it is sensed by the immediate neighbors of the respective VPC.
The lin-12 variable represents the level of lin-12 / Notch activity. If lin-12 activity is specified as wild-type, its activity level starts as MEDIUM in all VPCs. If lin-12 activity is eliminated [lin-12(0) mutation], then the lin-12 variable is set to OFF (when using the word set we mean that its value cannot change). By contrast, increasing lin-12 activity [lin-12(d) mutation] leads the variable to start as HIGH. Upon activation by the lateral signal, lin-12 activity increases from MEDIUM to HIGH. Upon inhibition of lin-12 activity by the EGFR/RAS/MAPK pathway, lin-12 activity decreases from MEDIUM or HIGH to LOW.
The lst genes variable, which can be either ON or OFF, collectively represents the activation state of the lst genes lip-1, ark-1, dpy-23, and lst-1 through lst-4 [9–11]. If lst genes are mutated to an inactive state, the variable is set to OFF. If all lst genes are wild-type, the variable starts as OFF and switches to ON upon activation by lin-12. We consider all the lst genes either as wild-type or as mutated to an inactive state.
The lin-15 variable collectively represents the state of lin-15 and other synMuv genes in hyp7, which can be either ON or OFF. If lin-15 is wild-type, this variable is set to ON and constitutively inhibits EGFR activation by LIN-3 from hyp7 (Figure 3). On the other hand, if lin-15 is mutated to an inactive state, the lin-15 variable is set to OFF, and the EGFR/RAS/MAPK pathway is constitutively and uniformly activated in all VPCs.
The inductive EGFR/RAS/MAPK pathway is represented by variables describing the status of the following four core components: let-23 (the EGF receptor), sem-5 (the Grb2-like adaptor), let-60 (the RAS GTP-binding protein), and mpk-1 (the MAP kinase). We consider either the wild-type behavior or mutations that completely inactivate a component (e.g., let-23(0) mutation causing the rest of the pathway never to be activated). Before the inductive signal is produced, let-23 egfr is OFF in all VPCs due to the presence of lin-15 and other synMuv genes, which prevent ectopic activation of LET-23 by repressing lin-3 egf expression in hyp7 [15]. Upon receiving the inductive signal from the AC, the variables simulate the activation of let-23, then sem-5, then let-60, and then mpk-1. The activation by a MEDIUM inductive signal is slower than the activation by a HIGH inductive signal. The EGFR/RAS/MAPK pathway can be counteracted by the lst variable described above during every stage of this activation sequence (see Figure 3, middle VPC).
A simulation starts by setting the type of mutation(s) that we would like to examine, and then following an execution of the model by choosing how to schedule the different VPCs. Once the cells assume their fates, we compare the fate assumption versus the desired experimental results.
To capture the diverse behavior often observed in biological systems, such as cases where the same genotype leads to different fate patterns, we allow complete freedom in the order of reactions between the different VPCs modules, but restrict the amount of progress each cell makes before its neighbors. The resulting model is highly nondeterministic, allowing many choices of execution without giving priorities or quantities to each choice. Each VPC is treated as a separate process. By adding a mechanism that decides which VPC to advance and for how long, we could get various patterns of VPC fates in different executions. Consequently, the model has approximately 4,000 different possible ways to complete one round in which all cells move. Subsequently, there are about 1036 possible executions of the model. In addition, the model has 48 initial states, corresponding to 48 different experimental conditions, and about 92,000 different reachable states (possible assignments to all the variables), each corresponding to a snapshot of the system.
As the number of possible executions of the model is astronomical, we use formal verification to ensure that all possible runs of the model emanating from a given mutation produce results that match the experimental results. To do that, we have formalized the experimental observations that led to formulate the mechanistic model underlying VPC pattern formation (e.g., if the model starts in the wild-type state, the VPCs assume fates according to the following pattern: 3°-3°-2°-1°-2°-3°), and used them to formally check whether the mechanistic model reproduces the reported experimental observations. Once we have established a model that reproduces all the experimental data, we can also use simulations to predict the outcome of new experiments that have not been performed yet.
In nondeterministic models, simple simulations (i.e., testing) are not sufficient to verify the model's consistency with the experimental data. The reason is that in nondeterministic models the number of possible behaviors resulting form the same initial condition could be enormous. Therefore, to test a nondeterministic model, one would have to run many simulations (one for each scenario). Another way to test nondeterministic models is to use model checking [4], which allows us to formally check all the different executions of the system against a formal specification. By exploring all the possible states and transitions of a system, we can determine whether some property holds true for the system. In the case that the property does not hold, the model-checking algorithm supplies a “counterexample,” which is an execution of the system that does not satisfy the given requirement, in the current case an experimentally observed pattern of vulval cell fates.
Here, we have used model checking for two purposes. First, to ascertain that our mechanistic model reproduces the biological behavior observed in different mutant backgrounds. For that, we have formalized the experimental results described in a set of papers (for references see Table 1) and verified that all possible executions satisfy these behaviors. That is, regardless of the order of interactions from a given set of initial conditions, different executions always reproduced the experimental observations. Second, we used model checking to query the behavior of the model. By phrasing queries such as which mutations may lead to a stable or an unstable fate pattern, we analyze the behavior of the model. Once an unstable mutation was found, we determined what part of the execution allows this kind of mutation by disallowing different behavioral features of the model and checking when the instability disappears.
We have tested the behavior of our model for a set of 48 perturbations corresponding to 24 mutant combinations, which were analyzed in the presence and absence of the AC (Table 1). For some of the combinations, the outcome has been tested experimentally as indicated in Table 1 by the respective references, but many other combinations have not been tested experimentally, as some of the double, triple, or quadruple mutants might be technically very difficult to generate. For example, complete loss-of-function mutations in most components of the inductive signaling pathway cause early larval lethality, or homozygous lin-12(gf) mutants lack an AC.
Forty-four of the 48 conditions tested yielded a stable fate pattern, as all possible executions gave the same result. All four conditions leading to an unstable pattern included the lin-15 knockout mutation. The cause of the unstable pattern in these four cases is discussed in the next section. Twenty-two of the conditions have been tested experimentally and the observed results are reported in the literature (see references in Table 1). Our model faithfully reproduces the predominant cell fate patterns that had been reported except for the phenotype of lin-12(d); lin-15(0) double mutants. While in lin-12(d); lin-15(0) animals, the distal VPCs (P3.p, P4.p, and P8.p) adopt either a 1° or a 2° cell fate [6], our model predicted that the six VPCs would always adopt a 2° fate. This discrepancy was traced to the fact that the high activity of LIN-12 simultaneously induced lst expression in all VPCs, which immediately repressed the transduction of the EGFR signal that was activated in all VPCs due to the lin-15(0) mutation (Figure 5A). The high levels of lst gene expression thus prevented the cells from engaging the mechanisms reducing LIN-12 activity, which is necessary for a 1° fate specification. In spite of having represented LIN-12 downregulation by EGFR signaling [18] and lst-mediated lateral inhibition on EGFR signaling [9,10], the model could not reproduce the experimental observations. Therefore, we postulated that some additional regulation is needed to allow primary fates while avoiding adjacent primary fates [22]. One possibility is that EGFR signaling downregulates one or several lst genes in addition to inducing lin-12 degradation (Figure 5B). If this happens before the activation of the lst genes completely blocks EGFR/RAS/MAPK signaling, then at least some VPCs are allowed to adopt a 1° fate. We further suggest that in order to avoid adjacent primary cells in these lin-12(d); lin-15(0) mutants, the lateral signal can still override the EGFR signal and lst activity prevails.
These insights led to a revised model with at least one additional negative feedback loop indicated by the red line in Figure 3. This refined model reproduces all the experimentally observed cell fate patterns including the lin-12(d); lin-15(0) double mutants (Table 1, rows 21 and 45). Of particular interest are those cases where two signaling pathways specifying different cell fates are simultaneously perturbed. For example, if the lateral signal is constitutively activated and at the same time the transduction of the inductive signal is blocked, then all VPCs are predicted to adopt the 2° cell fate irrespective of the presence or absence of the AC (Table 1, rows 19 and 43). Indeed, in lin-12(n137gf) mutants that carry a dominant-negative or strong reduction-of-function mutation in let-60/Ras, all VPCs were found to adopt the 2° cell fate [23]. In another condition we examined the interaction between the inhibitory lin-15 pathway and the lst genes. If both components are inactivated at the same time, all the VPCs are predicted to adopt a 1° cell fate in the presence as well as in the absence of the AC (Table 1, rows 6 and 30). As predicted by modeling, in the majority of lin-15(n309); lip-1(zh15) double mutants, adjacent VPCs adopt the 1° cell fate indicated by the expression of the 1° fate marker egl-17::gfp and by morphological criteria ([9] and unpublished data). An example for a condition that could not be tested experimentally is shown in Table 1, row 38. If all three signals, the inductive and lateral signals as well as the lin-15–mediated inhibition of hyp7, are inactive and the lst genes are mutated, all six VPCs are predicted to adopt the 1° cell fate as long as the EGFR/RAS/MAPK pathway is functional. This suggests that the default fate in the vulval equivalence group is 1°. Conversely, if the inductive and lateral signaling pathways are both constitutively activated (“lin15kolin12d” in Table 1, row 21), then the VPCs may adopt a 1° or 2° fate depending on the activity state of the lst genes (Table 1, rows 21, 22, 45, and 46).
In summary, by using model checking to compare our executable model with existing experimental data, we can predict novel interactions in the regulatory network governing vulval fate specification. In addition, analysis of the model allows us to predict the outcome of perturbations that are difficult to test experimentally.
Using model checking, we found that 44 out of 48 perturbations affecting vulval development lead to a stable fate pattern, despite the vast number of possible executions of our model. The only four mutations leading to unstable patterns are lin-15(0), lin-15(0); lin-12(d), lin-15(0); ac- and lin-15(0); lin-12(d) and ac- (Table 1, rows 15, 21, 29, and 45). To determine whether variations in the exact timing of the lateral signaling are the cause of this instability, we asked, using model checking, whether it is possible to get an unstable fate pattern without allowing variations in the timing of the lateral signal and found this not to be the case. We discovered that in order to adopt two different cell fates in two different executions, a VPC has to send the lateral signal before its neighbors in one execution, and after its neighbors in another execution. In the first case, the VPC will adopt a 1° fate and force its neighbors to adopt a 2° fate, while in the second case it is forced by one of its neighbors to adopt a 2° fate before it can adopt the 1° fate. Specifically, we found that in the four cases containing the lin-15(0) mutation, perturbation of the intricate timing dependency between the activation of the lateral signal and the inhibition of LIN-12 activity by the EGFR/RAS/MAPK pathway allows VPCs to adopt different fates in different executions of the model. Figure 6 distinguishes between stable and unstable fate patterns according to the ordering of events derived from the analysis of our model. In a stable pattern, the response to the inductive signal is temporally graded in a way that allows one VPC (e.g., VPC1 in Figure 6A) to send the lateral signal always before its neighbors reduce their level of LIN-12. In unstable patterns, on the other hand, the activation of the EGFR/RAS/MAPK pathway occurs more or less simultaneously in all VPCs, and small, stochastic timing differences result in variable patterns among genetically identical animals (Figure 6B). We note that this instability comes into effect only in AC-ablated animals or in VPCs that are too distant from the AC, suggesting that the AC organizes not only the spatial but also the temporal order of events.
To test the predictions made by our model, we examined the expression of cell fate-specific transcriptional reporters in developing animals. Using a strain carrying both the egl-17::cfp and lip-1::yfp transgenes as reporters for the 1° and 2° cell fate, respectively, we could simultaneously observe the activation of the inductive and lateral signaling pathways in the VPCs of individual animals. We first performed a time-course analysis in a wild-type background and quantified the strength of the fluorescent signals of the 1° and 2° fate-specific markers during the critical phase from the mid L2 stage on (22 h after starvation-induced L1 arrest) until the end of the L2 stage just before the VPCs have adopted their fates and start dividing (28 h after starvation-induced L1 arrest). In all animals except for one case at the 25-h time point, an increase in the expression of the 1° fate marker egl-17::cfp was observed in P5.p, P6.p, and P7.p before a significant upregulation of the 2° fate marker lip-1::yfp occurred in P5.p and P7.p (Figure 7A, 7B, and 7D). Thus, the inductive signaling pathway is activated already in mid-L2 larvae (+22 h), while lateral signaling is effective only toward the end of the L2 stage (+28 h). These experimental data provide, for the first time to our knowledge, direct evidence for a sequential activation of the inductive and lateral signaling pathways during vulval induction, as predicted by our model in the case of stable fate patterns. It should be noted that mosaic analysis of let-23 egfr had already suggested a sequential model for vulval fate specification [24], though the relative timing of the inductive versus lateral signaling events has to date not been investigated.
Next, we tested if in lin-15(n309) mutants that exhibit an unstable fate pattern in the distal VPCs the sequential activation of signaling pathways may be disrupted. Since larval development in lin-15(n309) animals is significantly delayed (unpublished data), it was not possible to perform the same time-course analysis as shown above for wild-type animals. We therefore staged lin-15(n309) animals carrying the egl-17::cfp and lip-1::yfp reporters based on the length of their posterior gonad arms and the shape of the VPCs to identify late L2 larvae corresponding approximately to the +28 h time point in wild-type larvae (see Materials and Methods). In 12 out of 22 late L2 lin-15(n309) larvae, the 1° and 2° fate markers were simultaneously expressed in at least one of the distal VPCs, P3.p, P4.p, or P8.p, which is consistent with the unstable fate pattern predicted for the distal VPCs in lin-15 mutants (Figure 7C and 7F). Moreover, in 21 out of 22 lin-15(n309) animals, P5.p and/or P7.p expressed both 1° and 2° fate markers (Figure 7F). In wild-type animals, on the other hand, co-expression of the 1° and 2° fate markers was never observed in the distal VPCs, but 12 out of 21 animals showed weak 1° fate marker expression in P5.p and/or P7.p in addition to the strong 2° marker expression (Figure 7E).
Thus, we could experimentally confirm two key predictions provided by our modeling work (Figure 6); the temporal gradient in the activation of the inductive and lateral signaling pathways in wild-type animals and the loss of sequential signaling in lin-15 mutants leading to an unstable fate pattern.
Formal executable models have become valuable tools to enhance our understanding of complex biological systems [1,3,25–30]. Here, we present an up-to-date comprehensive model of C. elegans vulval fate specification and experimental validation of two key predictions made by the model. Our model represents the current understanding of the regulatory signaling network and includes multiple modes of crosstalk between the EGFR/RAS/MAPK and NOTCH signaling pathways such as the LIN-12 / Notch-mediated lateral inhibition [9,10,22]. Since the model is dynamic and nondeterministic, it allows a very large number of different executions for a given starting condition. By using model checking, which permits us to investigate all possible executions of the model, we identify gaps in the conceptual understanding of the events leading to a stable pattern of vulval cell fates. The insights gained through model checking can then be used to refine an initial model until it fits all the experimental data. There could be several different ways to refine a model, and every conjecture made in the refinement process should then be validated experimentally. For example, our model suggests that the EGFR/RAS/MAPK pathway not only represses lin-12/Notch signaling [18] but also negatively regulates lst gene expression in 1° cells. In the 2° cells, on the other hand, lateral signaling overrides this postulated negative loop and lst activity prevents 1° cell fate specification. Although such a molecular mechanism has not yet been elucidated, our modeling study makes explicit the importance of this putative negative feedback loop. Interestingly, it was previously reported that some lst genes are not only positively regulated by lateral signaling but are also negatively regulated by inductive signaling [10]. In addition, the recently discovered homolog of the mammalian tumor suppressor dep-1 gene might also be part of this postulated negative feedback loop [31]. DEP-1 dephosphorylates the EGFR and thereby inhibits inductive signaling in the 2° cell lineage in parallel with the lst genes, while inductive signaling simultaneously downregulates DEP-1 and LIN-12/Notch expression in the 1° cell lineage, allowing full activation of the EGFR in these cells [18,31]. Thus, the reciprocal activation of EGFR/RAS/MAPK signaling and lateral inhibitors in 1° and 2° VPCs, respectively, might in part be mediated by a novel negative feedback loop downstream of the MAP kinase.
In mammals, the negative crosstalk between EGFR and Notch signaling may be important to control the balance between stem cell proliferation and differentiation [32], and alterations in the connections between these two signaling pathways may lead to cancer in humans [11]. Thus, future studies investigating the molecular details of negative feedback loops between EGFR signaling and the lst genes may help elucidate conserved mechanisms underlying EGFR function as an oncogene.
Our computational model allows flexibility in the order between different reactions, which resembles variations in the rate of biochemical reactions. This is akin to the robustness of simple biochemical networks that are resistant to variations in their biochemical parameters [33–35]. Despite this variability, we have found by model checking that in a wild-type situation all possible executions reach a stable state, independently of the order of reactions between the VPCs. This behavior of the model closely resembles the remarkable robustness of vulval development observed under various experimental conditions in the laboratory as well as in free living Nematodes. Furthermore, we observed that for most perturbations (i.e., mutations in the inductive or lateral signaling pathways), all the different executions lead to stable fate patterns. This suggests that the mechanism underlying VPC specification is relatively resistant to genetic variability and might therefore represent a process subject to high selective pressure.
A notable exception is the behavior of lin-15(lf) mutants, which—both experimentally and by modeling—exhibit an unstable fate pattern as long as the inductive signaling pathway is functional. We could trace down the cause of this instability to the fact that lin-15 mutations abrogate the temporal order in the activation of the inductive versus lateral signaling pathway among individual VPCs. Interestingly, recent experiments have demonstrated that lin-15(lf) mutations result in the ectopic expression of the inductive LIN-3 EGF signal in the hypodermal syncytium hyp7 [15]. Since all VPCs are in direct contact with hyp7, it seems reasonable to assume that in lin-15 mutants the EGFR is simultaneously activated in all VPCs, which likely disrupts the relative timing of inductive versus lateral signaling among adjacent VPCs. Our analysis of the behavior of lin-15 mutants thus illustrates how computational modeling can provide a plausible mechanistic explanation for phenotypic instability observed in real life.
Through model checking an executable model representing the crosstalk between EGFR and LIN-12 / Notch signaling during C. elegans vulval development, we have gained new insights into the usage of these conserved signaling pathways that control many diverse processes in all animals. While many modeling efforts use simulations that allow us to investigate only a few possible executions, our work emphasizes the power of analyzing all possible executions using model checking. Previous attempts to use model checking in biological modeling have concentrated on adapting model checking to formalisms such as differential equations and probabilistic modeling [36–38]. Our work demonstrates how biological processes can be described and analyzed with the use of formal methods, which enhances our comprehension of complex biological systems. We suggest that combining model-checking analysis with high-level modeling, similar to the level of abstraction used by biologists in describing mechanistic models, can help in many areas of biology to obtain more accurate, formal, and executable models, eventually leading to better understanding of biological processes.
RM is a modeling language for reactive systems [20]. RM is designed to describe systems which are discrete, deadlock-free, and nondeterministic. The elementary particles in RM are variables. We describe the behavior of variables in atoms and combine atoms into modules. Modules can be combined to create more complicated modules (including combinations of several copies of the same module). Each variable ranges over a finite set of possible values. An atom describes the possible updates on variables. An atom can be synchronous, meaning that it updates the variables it controls in every step of the system, or asynchronous, meaning that it updates the variables it controls from time to time. An update of a variable may depend on the value of itself as well as the values of other variables. There can also be dependencies between the mutual update of several variables in the same step. RM enables nondeterminism by allowing multiple overlapping update options. The current RM model does not include probabilities.
Mocha is a software tool for the design and analysis of RM [21]. Mocha can simulate a model by following step-by-step evolution of the variables in the model. Simulations show the sequence of values assumed by variables during the simulation. In simulation of nondeterministic models, the user is expected to choose the next step between different nondeterministic options. The simulation engine can highlight the variable values that lead to the assignment of a certain value. Mocha supports invariant model checking directly (to check that all reachable states satisfy some property that relates to the values of variables in the state), as well as model checking of safety properties using monitors (to check that all executions satisfy some property). We use both enumerative and symbolic (using Boolean Decision Diagrams, BDDs) model checking; the difference between the two has to do with their performance in practice. Counterexamples are presented as sequences of variable values.
Parallelism is an important property of biological systems. In computer science, this is referred to as concurrency (processes running in parallel and sharing common resources). We usually distinguish between two forms of concurrency: synchronous and asynchronous. In synchronous systems, all components move together. That is, there is some basic work unit that all components share. All components do one work unit in parallel simultaneously. Then they all move to the next unit. In asynchronous systems, every component moves separately. Usually, in asynchronous systems, we do not allow components to move together and we cannot guarantee the relative speed of different components. Biological systems, while highly concurrent, are neither completely synchronous nor completely asynchronous. Different molecules, or cells, do not progress in perfect lockstep, and neither does any molecule or cell rest for arbitrary amounts of time. For this reason, we have introduced a new notion of bounded-asynchrony into our computational model. In bounded-asynchrony, the scheduler, which chooses the next component to move, is not completely free in its choices. No component can be neglected more than a bounded number of times. This captures the phenomenon that the components of a biological system (say, molecules or cells, depending on the level of modeling granularity) progress neither in lockstep nor completely independently, but that they are loosely coupled and proceed approximately along the same timeline. We find the notion of bounded asynchrony a pragmatic way to model cell–cell interactions in an abstract discrete framework. Further studies are needed to identify the appropriate model for concurrency in different biological contexts.
Standard methods were used for maintaining and manipulating C. elegans [39]. The C. elegans Bristol strain, variety N2, was used as the wild-type reference strain in all experiments. Mutations used: lin-15(n309) [22]; integrated transgene arrays used: arIs92[egl-17::cfp,tax-3::gfp] [10], mfIs42[lip-1::yfp] (gift of M. A. Félix).
Synchronized populations of L1 larvae were obtained by isolating embryos from gravid adults using sodium hypochlorite treatment and arresting the newly hatched larvae by food starvation. The arrested L1 larvae were then placed on standard NGM growth plates containing E. coli OP50 and collected for microscopic observation at the indicated time points. For observation under Nomarski optics, animals of the indicated stages were mounted on 4% agarose pads with M9 buffer containing 10 mM sodium azide. Fluorescent images were acquired on a Leica DMRA wide-field microscope equipped with a cooled CCD camera (Hamamatsu ORCA-ER, http://www.hamamatsu.com/) controlled by the Openlab 3.0 software package (Improvision, http://www.improvision.com/). For quantification of YFP and CFP intensity in the VPCs, all images were acquired with the same microscopy, camera, and software settings using YFP- and CFP-specific filter sets. The mean intensity of CFP and YFP expression in the nuclei of the VPCs was measured using the measurement tool in the Openlab 3.0 software package (Improvision), and each measurement was standardized to the background intensity in the same picture. For each time point, between ten and 12 animals were quantified. Late L2 lin-15(n309) animals were identified by selecting larvae in which the VPCs had adopted an oval shape and the distal tip of the posterior gonad arm had migrated past P7.p (see Figure 7B and 7C). |
10.1371/journal.pcbi.1000826 | Mammalian Sleep Dynamics: How Diverse Features Arise from a Common Physiological Framework | Mammalian sleep varies widely, ranging from frequent napping in rodents to consolidated blocks in primates and unihemispheric sleep in cetaceans. In humans, rats, mice and cats, sleep patterns are orchestrated by homeostatic and circadian drives to the sleep–wake switch, but it is not known whether this system is ubiquitous among mammals. Here, changes of just two parameters in a recent quantitative model of this switch are shown to reproduce typical sleep patterns for 17 species across 7 orders. Furthermore, the parameter variations are found to be consistent with the assumptions that homeostatic production and clearance scale as brain volume and surface area, respectively. Modeling an additional inhibitory connection between sleep-active neuronal populations on opposite sides of the brain generates unihemispheric sleep, providing a testable hypothetical mechanism for this poorly understood phenomenon. Neuromodulation of this connection alone is shown to account for the ability of fur seals to transition between bihemispheric sleep on land and unihemispheric sleep in water. Determining what aspects of mammalian sleep patterns can be explained within a single framework, and are thus universal, is essential to understanding the evolution and function of mammalian sleep. This is the first demonstration of a single model reproducing sleep patterns for multiple different species. These wide-ranging findings suggest that the core physiological mechanisms controlling sleep are common to many mammalian orders, with slight evolutionary modifications accounting for interspecies differences.
| The field of sleep physiology has made huge strides in recent years, uncovering the neurological structures which are critical to sleep regulation. However, given the small number of species studied in such detail in the laboratory, it remains to be seen how universal these mechanisms are across the whole mammalian order. Mammalian sleep is extremely diverse, and the unihemispheric sleep of dolphins is nothing like the rapidly cycling sleep of rodents, or the single daily block of humans. Here, we use a mathematical model to demonstrate that the established sleep physiology can indeed account for the sleep of a wide range of mammals. Furthermore, the model gives insight into why the sleep patterns of different species are so distinct: smaller animals burn energy more rapidly, resulting in more rapid sleep–wake cycling. We also show that mammals that sleep unihemispherically may have a single additional neuronal pathway which prevents sleep-promoting neurons on opposite sides of the hypothalamus from activating simultaneously. These findings suggest that the basic physiology controlling sleep evolved before mammals, and illustrate the functional flexibility of this simple system.
| The diversity of mammalian sleep poses a great challenge to those studying the nature and function of sleep. Typical daily sleep durations range from 3 h in horses to 19 h in bats [1], [2], which has led to recent speculation that sleep has no universal function beyond timing environmental interactions, with its character defined purely by ecological adaptations on a species-by-species basis [3]. Consolidated (monophasic) sleep, has only been reported in primates [2], whereas the vast majority of mammals sleep polyphasically, with sleep fragmented into a series of daily episodes, ranging in average length from just 6 min in rats to 2 h in elephants [1]. Some aquatic mammals (such as dolphins and seals) engage in unihemispheric sleep, whereby they sleep with only one brain hemisphere at a time [4]–[6]. This behavior appears to serve several functions, including improved environmental surveillance and sensory processing, and respiratory maintenance [7], although the physiological mechanism is unknown [8], [9]. Determining which aspects of mammalian sleep patterns can be explained within a single framework therefore has important implications in terms of both the evolution and function of sleep. As we show here, although mammalian sleep is remarkably diverse in expression, it is very likely universal in origin.
Recent advances in neurophysiology have revealed the basic mechanisms that control the mammalian sleep cycle [10], [11]. Monoaminergic (MA) brainstem nuclei diffusely project to the cerebrum, promoting wake when they are active [12]. Mutually inhibitory connections between the MA and the sleep-active ventrolateral preoptic area of the hypothalamus (VLPO) result in each group reinforcing its own activity by inhibiting the other and thereby indirectly disinhibiting itself. This forms the basis of the sleep-wake switch, with active MA and suppressed VLPO in wake, and vice versa in sleep [10]. State transitions are effected by circadian and homeostatic drives, which are afferent to the VLPO [13]. The approximately 24 h periodic circadian drive is entrained by light, and projects from the suprachiasmatic nucleus (SCN) to the VLPO via the dorsomedial hypothalamus (DMH) [14]. The homeostatic drive is a drive to sleep that increases during wake due to accumulation of somnogens, accounting for the observed sleep rebound following sleep deprivation [15]. During sleep, somnogen clearance exceeds production and the homeostatic drive decreases. The exact physiological pathway has yet to be fully elaborated, but some important somnogenic factors have been identified, including adenosine (a metabolic by-product of ATP hydrolysis) [16] and immunomodulatory cytokines [17]. The present work uses a model that does not depend on the precise identity of the somnogen (or somnogens), but may help to elucidate its characteristics.
Whether the above system can account for the wide variety of mammalian sleep patterns is unknown. Is the sleep-wake switch a universal physiological structure among mammals? Or are the qualitative differences in sleep-wake patterns between species such as rats and dolphins due to fundamentally different mechanisms? To answer these questions we apply a recent quantitative physiologically-based model [18], [19]; this approach allows the underlying physiological structure to be related to the observed dynamics. As shown in Fig. 1, the model includes the MA and VLPO groups, circadian and homeostatic drives to the VLPO, and cholinergic and orexinergic input to the MA (for mathematical details, see Methods). The model is based on physiological and behavioral studies of a small number of species, including rats, mice, cats, and humans, and has been calibrated previously to reproduce normal human sleep and recovery from sleep deprivation [18], [19]. But as we will show, the model is also capable of reproducing the typical sleeping patterns for a wide range of mammalian species, including both terrestrial and aquatic mammals.
With nominal parameter values (given in Methods), the model has previously been shown to reproduce normal human sleep patterns, with approximately 8 h of consolidated sleep, and relatively rapid (approximately 10 min) transitions between wake and sleep [18], as shown in Fig. 1. We found that by varying just two of the model parameters, the model could be made to reproduce the bihemispheric sleep patterns of a wide variety of mammals, including many in which the neuronal circuitry controlling sleep rhythms has not been examined. These parameters were: (i) the homeostatic time constant, determining the rate of somnogen accumulation and clearance, and (ii) the mean drive to the VLPO, provided by the SCN, DMH and other neuronal populations. The homeostatic time constant was found previously to be approximately 45 h for humans, based on the rate of recovery from total sleep deprivation [19], but we found here that reducing it below 16 h resulted in polyphasic sleep, as seen in most other mammals. This is because a shorter time constant causes somnogens to accumulate more quickly during wake, and dissipate more quickly during sleep, resulting in more rapid cycling between wake and sleep. Increasing the mean inhibitory drive to the VLPO was found to decrease daily sleep duration with little effect on the other dynamics.
Fitting the model to experimental data for 17 species in which both average daily sleep duration and average sleep episode length have been reliably reported yielded the map in Fig. 2, showing which regions of parameter space correspond to the typical sleep patterns of each species. (Note that at least some quantitative sleep data is available for over 60 species, but these two measures have not both been reliably reported in most cases.) This map enables classification of mammals based on sleep patterns, and can be further populated in future when more data becomes available. The regions corresponding to the human, rhesus monkey, and slow loris lie in the monophasic zone, but with different mean VLPO drives. In each case, the lower bound for the homeostatic time constant was determined by the boundary of the monophasic zone. For humans, the upper bound of 72 h was previously determined using sleep deprivation experiments [19]. In the absence of experiments detailing recovery from total sleep deprivation in non-human primates, we used the same upper bound for both the rhesus monkey and the slow loris; more data is required to rigorously constrain the homeostatic time constant for these species.
Animals that sleep relatively little, such as the elephant, were inferred to have high values of mean drive to VLPO, while animals that sleep a lot, such as the opossum and armadillo, were inferred to have low values of mean drive to VLPO. Those that cycle rapidly between wake and sleep, such as rodents, were inferred to have short homeostatic time constants (around 10 min to 1 h), while those with fewer sleep episodes per day, such as the jaguar and elephant were inferred to have longer time constants (around 5 h to 10 h), thus lying closer to the boundary between polyphasic and monophasic sleep. The extreme cases of no wake and no sleep may correspond to brainstem lesions, such as those documented clinically [31], and possibly other states of reduced arousal (e.g., hibernation, torpor, coma), although we did not pursue them here.
Using parameter values from the appropriate regions in Fig. 2, we generated sample time series for various species. Comparisons to experimental data for the human, elephant and opossum are shown in Fig. 3. In each case, the model reproduced the salient features of the sleep/wake pattern. For the opossum, the circadian signal was shifted in phase by 12 h to reproduce the nocturnal distribution. This is justified by physiological evidence suggesting that temporal niche is determined by how SCN output is modulated by the DMH relay system [11].
Plotting the homeostatic time constants inferred for each species versus body mass in Fig. 4 revealed a positive correlation. Fitting a power-law relationship yielded an exponent of 0.29±0.10 for non-primates. Additional data are required to accurately constrain homeostatic time constants in non-human primates, but using the human-derived upper bound of 72 h yielded an exponent of 0.01±0.26 for primates, and 0.28±0.12 for all species.
Power-law relationships are ubiquitous in biology, although their quantification remains controversial. For mammals it has been found that brain mass scales as approximately , where is total body mass, and metabolic power per unit volume scales as for brain tissue [33]. Without knowing the precise mechanism by which the homeostatic drive is regulated, we nonetheless tested general assumptions that are equally applicable to a wide range of candidate mechanisms. We assumed that somnogen production is proportional to the total power output of the brain (as would plausibly be the case for adenosine), meaning production per unit volume would scale as , with different production rates in wake and sleep. Furthermore, we made the generic assumption that somnogen clearance rate is proportional to working surface area, where this surface area may be glial, vascular, or otherwise, depending on the exact physiological pathway. The total clearance rate then scaled as , where , depending on the geometry: corresponds to surface area scaling as the square of the brain's linear dimension (i.e., as for simple solids), and to scaling as its cube (e.g., as for solids with highly convoluted or fractal surfaces). By assuming clearance rate was also proportional to somnogen concentration, the homeostatic time constant was found to be proportional to (see Methods for a full derivation). For , this yielded a power law exponent of 0.23, consistent with that found for non-primates. The smaller exponent found for primates was consistent to within uncertainties with that found for non-primates; more primate data are required to determine whether is closer to 1 in primates, or whether both groups follow the same scaling law but with different normalization constants.
We next turned to modeling unihemispheric sleep by extending the above model to permit distinct dynamics for the two halves of the brain. As shown in Fig. 1, this was achieved by coupling together two identical versions of the original model, each representing one hemisphere. This division in the model was justified by the fact that all nuclei in the VLPO and MA groups are bilaterally paired [12], [34], with the exception of the dorsal raphé nucleus, which lies on the brainstem midline [12]. Separate homeostatic drives were included for each brain hemisphere, based on experimental evidence for localized homeostatic effects in humans, rats and dolphins [35]–[38]. Aquatic mammals that have been observed to sleep unihemispherically spend little or no time in bihemispheric sleep while in water [8] (although fur seals switch to exclusively bihemispheric sleep when on land [39]). Hence, we postulated the existence of a mutually inhibitory connection between the two VLPO groups in aquatic mammals to prevent both activating at once (just as the mutually inhibitory VLPO-MA connection prevents both those groups activating simultaneously), thereby preventing bihemispheric sleep. This connection is presumably absent or very weak in other mammals.
For VLPO-VLPO connection strengths weaker than a threshold value sleep was purely bihemispheric, and above this value at least some unihemispheric sleep episodes occurred. For connection strengths stronger than a higher threshold the model exhibited purely unihemispheric sleep, typical of cetaceans. Differing homeostatic pressures between the two hemispheres drove alternating episodes of left and right unihemispheric sleep, with episode length controlled by homeostatic time constant, in a way similar to polyphasic bihemispheric sleep as described above. In Fig. 5, increasing the VLPO-VLPO connection strength was shown to cause a transition from polyphasic bihemispheric sleep to unihemispheric sleep, as for fur seals moving from land to water [6], [39]. Since no other parameter changes were required, we hypothesized that fur seals achieve this readjustment by dynamically neuromodulating the VLPO-VLPO connection strength in response to environmental stimuli. The required strengthening by a factor of somewhat more than 2.4 is reasonable given the magnitudes of typical neuromodulator effects.
We have provided the first demonstration that the neuronal circuitry found in a small number of species in the laboratory, including rats, mice and cats, can account for the sleep patterns of a wide range of mammals. Furthermore, this was achieved by varying only two model parameters, with all others taking fixed values determined previously. The implications of this are far-reaching: universality of this fundamental physiological structure across diverse orders would suggest that its evolution predates mammals. This is consistent with findings that show the monoaminergic system is phylogenetically pre-mammalian [40], and that simple organisms such as the zebrafish share homologous neuronal and genetic control of sleep and wake [41], [42]. Our results also demonstrate the inherent functional flexibility of the sleep-wake switch, which plausibly accounts for its evolutionary success in the face of diverse evolutionary pressures on the sleep-wake cycle. Physiological commonality is also of immense importance when using animals in pharmaceutical development, and for inferring the consequences for humans of animal sleep experiments and genetics.
Our findings suggest that the rate of cycling between wake and sleep is largely determined by the homeostatic time constant, which is inferred to have a positive correlation with body mass. Deviations from this relationship are likely due to selective pressures such as predation, food availability, and latitude. Consistent with this, a previous study found a scaling law of exponent 0.20±0.03 between the characteristic timescale of sleep episode durations (which followed an exponential distribution) and body mass [43]. Mean drive to the VLPO determined sleep duration, and no clear correlation was found between this parameter and body size. Experimental evidence suggests that sleep duration is dictated by interplay between physiological and ecological pressures [44].
The primary advantage conferred by using a physiologically-based model to analyze and interpret data is the ability to relate such behavioral measures to physiology, giving new insights into how interspecies differences in sleep patterns arise. Due to the relative paucity of appropriate data, in this study we made use of all data we could find. This meant combining results of behavioral studies with EEG studies, despite the fact that these methods likely produce slightly different estimates of sleep duration and sleep bout length. While this should not affect our main conclusions, it could fractionally shift the zones in Fig. 2. We thus emphasize the importance of experimentalists continuing to study a wide variety of mammalian species, and encourage them to report metrics such as sleep bout length, total daily sleep duration, and transition frequencies.
While the exact physiological mechanism underlying the homeostatic sleep drive is unknown, some pieces of the puzzle have been identified. Growing evidence points to the role of adenosine accumulation at specific brain sites in promoting sleep. In the rat, basal forebrain adenosine concentration has been found to gradually rise and fall during wake and sleep, respectively, with heightened levels following sleep deprivation [16]. Artificial infusion of adenosine reduces vigilance [45], and the wake-promoting effects of caffeine (which is a competitive antagonist of adenosine) provide additional indirect evidence for adenosine's role in homeostatic sleep regulation. However, the pathway by which adenosine induces sleep is not altogether clear. Adenosine inhibits wake-promoting cholinergic neurons in the basal forebrain, and disinhibits the VLPO via another basal forebrain population [13], [46], yet adenosine agonists continue to promote sleep even after cholinergic neurons are lesioned [47]. Immune signaling molecules such as interleukin-1 (IL-1) and tumor necrosis factor (TNF) have also been linked to homeostatic sleep regulation [17]. Levels of TNF and IL-1 alternate with the sleep/wake cycle, and their exogenous administration induces sleepiness [48]. Furthermore, increased cytokine production during bacterial infection increases sleep duration [48], unless the IL-1 system is antagonized [49]. However, the pathway by which cytokines regulate sleep has yet to be fully elaborated. More critically, no physiological process has been demonstrated to account for the homeostatic drive's timescale, which can be up to a week in the case of chronic sleep deprivation in humans [50]. Adenosine's half life in the blood is only seconds [51], suggesting that clearance and production may be rate-limited further upstream.
In this paper, we assumed that somnogen production and clearance rates are proportional to brain volume and surface area, respectively. The utility of this approach is that it does not require precise knowledge of the physiology underlying the homeostatic drive, because these assumptions are equally valid for a wide range of candidate mechanisms. Using them, we were able to relate scaling laws for metabolism and brain mass to the observed interspecies differences in sleep patterns. Additional data is required to ascertain whether primates follow a different scaling law from non-primates, and if so whether this is due to greater cortical folding, cortical thickness, and neuronal density than most other mammals [52], which could feasibly account for geometrical differences in vascular surface area for instance. Furthermore, additional data is required to determine whether the positive correlation between body mass and homeostatic time constant conforms to a power law. In a similar vein, a theoretical study by Savage and West [53] was able to predict an observed power law relationship between body mass and the ratio of sleep to wake duration, based on the assumption that sleep's primary function is brain maintenance and repair, but the present derivation is the first from a dynamical sleep model.
While sleep/wake patterns are controlled at a fundamental level by systems in the brainstem and hypothalamus, it is worth remembering that sleep is a multi-scale phenomenon, regulated at many levels. For example, synaptic homeostasis may contribute to the local regulation of slow wave activity in the cortex during sleep, and could even play a role in generating the homeostatic drive to the sleep-wake switch [54], [55].
The proposed interhemispheric inhibitory connection in unihemispheric sleepers awaits experimental testing. To date, VLPO afferents have only been studied in animals that sleep bihemispherically, with the great majority of these being ipsilateral [34]. It remains to be seen whether aquatic mammals have a stronger contralateral connection. A question that naturally arises is whether an analogous connection might also be present to some degree in animals that sleep bihemispherically, and whether unihemispheric sleep could be induced by decoupling the hemispheres by other means. Acallosal humans have decreased EEG coherence between hemispheres during sleep, but do not display unihemispheric sleep [56], suggesting that hemispheric synchrony is achieved subcortically. Consistent with this, bisection of the brainstem in cats has been shown to result in all four behavioral states: bihemispheric wake, bihemispheric sleep, and unihemispheric sleep in each hemisphere [57]. This suggests that in bihemispheric sleepers, contralateral excitatory connections between wake-promoting brainstem nuclei and/or the VLPO nuclei may be important to maintaining synchrony. However, bisection of the brainstem in monkeys did not induce unihemispheric sleep [58]. The existence of several other commissures between the hemispheres, including the corpus callosum, may help to explain these results, with one able to compensate for the lack of another in some species. Animals that sleep unihemispherically appear to have evolved multiple physiological changes in parallel to enable this mode of sleep, including a narrow or absent corpus callosum in dolphins and birds, respectively, to reduce interhemispheric coupling [59].
In future, our model could be applied to the sleep of species from other classes, including unihemispheric sleep in reptiles and birds [8]. Furthermore, we could consider explicitly modeling the DMH pathway to explore how temporal niche (diurnal vs. nocturnal vs. crepuscular) is determined. Extending the model to differentiate between REM and NREM sleep could provide additional insights. Using such approaches in parallel with physiological investigations could then help to elucidate the evolutionary development of the sleep-wake switch and its specializations.
We begin by reviewing the sleep-wake switch model developed previously; for more details see references [18] and [19]. The model includes the MA and VLPO neuronal populations, and the parameters of the model have been rigorously calibrated by comparison to physiological and experimental data for normal human sleep and recovery from sleep deprivation [18], [19]. Nominal human parameter values are given in Table 1. Each neuronal population has a mean cell-body potential relative to resting and a mean firing rate , where for MA and VLPO, respectively, with(1)where is the maximum possible firing rate, is the mean firing threshold relative to resting, and is its standard deviation. Neuronal dynamics are represented by(2)(3)where the weight the input to population j from k, is the decay time for the neuromodulator expressed by group j. The orexinergic/cholinergic input to the MA group is held at a constant average level to smooth out ultradian REM/NREM dynamics [18]. The drive to the VLPO,(4)includes homeostatic and circadian components, where and are constants determining the strengths of the homeostatic and circadian drives, respectively. The parameter is positive, so that the homeostatic drive promotes sleep; this is consistent with disinhibition of the VLPO by basal forebrain adenosine [13]. The parameter is negative, consistent with the fact that SCN activity promotes wake in diurnal animals [60]. Differences in temporal niche appear to be due in part to an inversion of this signal [60], but as noted in the Discussion, we do not attempt to model this here. The circadian drive is here assumed to be well entrained and so is approximated by a sinusoid with 24 h period,(5)where h−1, is the mean drive to the VLPO, and is the initial phase. The homeostatic sleep drive is represented by somnogen concentration , with its dynamics governed by(6)where is the homeostatic time constant, and is a constant which determines the rate of homeostatic production. Previously, has been considered a model for adenosine concentration in the basal forebrain [18], but this general form is equally applicable to many other candidate somnogens.
As shown in earlier work [18], during normal functioning of the model, is high (∼5 s−1) in wake, is low (∼0 s−1) and is increasing, while is low in sleep, is high and is decreasing. For the purposes of comparing to data, we define the model to be in wake if s−1, based on comparison with experimental data for MA firing rates [61]. The model differentiates wake vs. sleep states, and we make no attempt to reproduce different sleep intensities or intra-sleep architectures between species.
The parameters and are varied to reproduce mammalian sleep patterns using total daily sleep duration and average sleep episode length as metrics to calibrate against. They have previously been estimated to take the values and h for humans. These parameters were selected as best able to account for differences in both total daily sleep duration and sleep bout length based on preliminary investigations and previous sensitivity analysis [18]. Data for calibration were derived from an extensive search of the literature to find studies that reported ranges for both metrics, yielding the 17 species used here. Parameter ranges that satisfied these metrics were plotted as the regions shown in Fig. 2. All of the available data were used, with one exception: additional data for non-human primates that sleep monophasically were omitted since we are unable to derive an upper bound for the homeostatic time constant without obtaining data detailing the dynamics of recovery from total sleep deprivation for these species. Those included in the study (the slow loris and the rhesus monkey) are shown for illustrative purposes using the human-derived upper bound of 72 h.
To produce Fig. 3, we add noise terms with to the right hand sides of Eqs (2) and (3), respectively, so as to make the sleep patterns less regular. The noise is taken from a normal distribution of mean 0 and standard deviation 1, and mV h1/2/(ΔT)1/2, where ΔT is the size of the time step used in the numerical integration. Values of parameters are taken from within the appropriate regions in Fig. 2. For the human, we use , h; for the elephant, we use , h; for the opossum we use , h.
For modeling unihemispheric sleep, the above model, defined by Eqs. (1)–(6) is used identically to model the dynamics of each half of the brain, with the following modification to the VLPO differential equation:(7)where is the firing rate of the VLPO population in the other half of the brain, and represents the strength of the contralateral inhibitory connection.
Mammalian brain mass has been found to follow an approximate scaling law(8)where is body mass [33]. Furthermore, the power output of the brain follows,(9)If the total rate of somnogen production in the brain is assumed to be proportional to the total power output of the brain , then the rate of somnogen production per unit volume, denoted by , is(10)
We assume that the total clearance rate is proportional to the working surface area, which may be glial, vascular, or otherwise. The working surface area will thus scale as the brain's mass, , where depending on the brain's geometry. Therefore, the rate of somnogen clearance per unit volume, denoted by , is(11)Now, if is produced at a rate where is a factor that depends on the state of arousal (i.e., production is expected to be higher in wake than in sleep), and is cleared at a rate , where is constant, then(12)which can be rewritten as(13)where the homeostatic time constant is , and . For , this yields and , justifying the approximation of holding constant while varying throughout this study.
|
10.1371/journal.pntd.0002729 | Use of a Recombinant Cysteine Proteinase from Leishmania (Leishmania) infantum chagasi for the Immunotherapy of Canine Visceral Leishmaniasis | A recombinant cysteine proteinase from Leishmania (Leishmania) infantum chagasi (rLdccys1) was previously shown to induce protective immune responses against murine and canine visceral leishmaniasis. These findings encouraged us to use rLdccys1 in the immunotherapy of naturally infected dogs from Teresina, Piauí, a region of high incidence of visceral leishmaniasis in Brazil.
Thirty naturally infected mongrel dogs displaying clinical signs of visceral leishmaniasis were randomly divided in three groups: one group received three doses of rLdccys1 in combination with the adjuvant Propionibacterium acnes at one month interval between each dose; a second group received three doses of P. acnes alone; a third group received saline. The main findings were: 1) dogs that received rLdccys1 with P. acnes did not display increase of the following clinical signs: weight loss, alopecia, onychogryphosis, cachexia, anorexia, apathy, skin lesions, hyperkeratosis, ocular secretion, and enlarged lymph nodes; they also exhibited a significant reduction in the spleen parasite load in comparison to the control dogs; 2) rLdccys1-treated dogs exhibited a significant delayed type cutaneous hypersensitivity elicited by the recombinant antigen, as well as high IgG2 serum titers and low IgG1 serum titers; sera from rLdccys1-treated dogs also contained high IFN-γ and low IL-10 concentrations; 3) control dogs exhibited all of the clinical signs of visceral leishmaniasis and had low serum IgG2 and IFN-γ levels and high concentrations of IgG1 and IL-10; 4) all of the dogs treated with rLdccys1 were alive 12 months after treatment, whereas dogs which received either saline or P. acnes alone died within 3 to 7 months.
These findings illustrate the potential use of rLdccys1 as an additional tool for the immunotherapy of canine visceral leishmaniasis and support further studies designed to improve the efficacy of this recombinant antigen for the treatment of this neglected disease.
| Visceral leishmaniasis (VL) is an important public health problem and dogs are the main domestic reservoirs of zoonotic VL which has resulted in an annual incidence of 40,100–75,500 new human cases. Because canine VL chemotherapy is limited by the low efficacy of drugs currently used for human VL treatment, immunotherapy may provide a viable alternative. We used a recombinant cysteine proteinase from L. (L.) infantum chagasi, rLdccys1, in combination with the adjuvant P. acnes for the treatment of naturally infected mongrel dogs from Teresina, Pauí a state in Brazil that has a high incidence of VL. Dogs treated with rLdccys1 showed a significant delayed type hypersensitivity reaction against the recombinant antigen and displayed high serum concentrations of IgG2 and IFN-γ and low concentrations of IgG1 and IL-10. Immunotherapy with rLdccys1 resulted in no increase of the clinical signs of canine VL and an extensive reduction of spleen parasite load. Furthermore, all of the dogs treated with rLdccys1 survived for at least 12 months after treatment, whereas those that received either saline or P. acnes alone died within 3 to 7 months. These findings support the potential of rLdccys1 immunotherapy as an additional option for the treatment of canine VL.
| Zoonotic visceral leishmaniasis (VL) is caused by Leishmania (Leishmania) infantum chagasi in Mediterranean, Middle-East, Asian countries, and Latin America and dogs are the main domestic reservoirs of this zoonosis which has resulted in an annual incidence of 40,100–75,000 new human cases [1], [2]. A high human VL incidence has been reported in Brazil mainly due to disease urbanization as a consequence of human migration from rural areas and ineffective vector and reservoir control [3]–[6]. Canine VL control is based on either treatment or euthanasia of infected animals. However, treatment of canine leishmaniasis with drugs successfully used for human VL shows low efficacy and induces the development of parasitic resistance to these drugs [7]–[10]. The WHO thus strongly recommends that the same drugs should not be used for treatment of dogs and humans in a same area [2]. On the other hand, euthanasia of infected dogs is often unacceptable for ethical and social reasons. Furthermore, the elimination of infected dogs has shown controversial results in Brazil [11], [12]. These issues led to the search of immunotherapy as a treatment alternative for canine VL. The administration of L. (L.) infantum chagasi extracts associated with the conventional chemotherapy of naturally infected dogs resulted in a significant reduction in infectivity [13]. Similar results were observed in dogs infected with L. (L.) infantum chagasi that displayed a significant parasite burden reduction after treatment with autoclaved L. (L.) major antigens and heat killed Mycobacterium vaccae administered in conjunction with Glucantime [14]. The healing efficacy of some vaccine candidates has also been tested. Treatment of infected dogs with purified L. (L.) infantum chagasi LiF2 antigen in combination with Glucantime led to the disappearance of clinical signs and a 100% cure rate [15]. Dogs naturally infected with L. (L.) infantum chagasi and treated with the recombinant vaccine Leish-110f formulated with the adjuvant MPL-SE associated with Glucantime showed clinical improvement, parasitological cure and increased survival [16]. Recent data supported the effectiveness of this recombinant vaccine for the treatment of mild cases of canine VL [17]. The immunotherapeutic potential of the Leishmune vaccine alone or in association with chemotherapy for canine VL treatment has also been demonstrated [18]–[20].
A recombinant cysteine proteinase from L. (L.) infantum chagasi, rLdccys1 was previously shown to be an useful immunological marker for different VL stages in humans and dogs, and to offer an appropriate diagnostic tool for human and canine VL [21]–[23]. Furthermore, immunization with either rLdccys1 or the gene Ldccys1, which encodes the cysteine proteinase, induced significant protection against L. (L.) infantum chagasi infection mediated by a predominant Th1 response in a murine model of VL [24]. In that study rLdccys1 was administered with P. acnes, a Gram-positive bacillus, known to induce a prevalent Th1 immune response in mice [25], [26]. In earlier studies we used P. acnes as an adjuvant to immunize BALB/c mice with native Ldccys1; this resulted in a predominant Th1 response and a significant protection against L. (L.) infantum chagasi challenge [27]. These results encouraged us to evaluate the immunotherapeutic potential of rLdccys1 plus P. acnes for naturally infected dogs from Teresina, Piauí, a state in Brazil with a high incidence of VL [28].
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the Brazilian National Council of Animal Experimentation (http://www.cobea.org.br). The protocol was approved by the Committee on the Ethics of Animal Experiments of the Institutional Animal Care and Use Committee at the Federal University of São Paulo (Id # CEP 1540/11).
A total of thirty dogs of different breeds, ages and sex were provided by the Zoonosis Control Center in Teresina, Piauí, in Brazil, a VL endemic area. The animals were kept in a thin-screened kennel in the Faculdade NOVAFAPI within Teresina, Piauí state and fed commercially balanced animal food (Cherokee, PRT, Brazil). Drinking water was provided ad libitum. Diagnostic procedures were performed in the Department of Parasitology and Microbiology at the Federal University of Piauí. All dogs had been pre-treated with anti-tick, anti-scabies and anti-helminthic drugs and had been vaccinated against Parvovirus, Adenovirus type II, Distemper, rabies (Defensor, Pfizer, EUA), Parainfluenza and Corona viruses and Leptospirosis (Vanguard Plus, Pfizer, EUA). Animals diagnosed with distemper, ehrlichiosis, and babesiosis were excluded from the study. Animals with severe renal failure (creatinine and urea values higher than 2.0 mg/dl and 35 mg/dl, respectively) and pancytopenia (total leucocyte number lower than 6,000 and number per mm3 of erythrocytes and platelets lower than 5.5×103 and 200, respectively) were not included in the study. All animals selected for immunotreatment tested positive for leishmaniasis in parasitological, serological and biochemical assays, had typical clinical signs of visceral leishmaniasis and were never treated for canine leishmaniasis. Diagnosis was grounded on positive bone marrow aspirate cultures, and ELISA assays using an L. (L.) chagasi amastigote extract and rLdccys1 as antigens. Additional laboratory studies included: complete blood cell counts, biochemical assays (creatinine, urea, alkaline phosphatase, aspartate aminotransferase, alanine aminotransferase, albumin, total protein, glucose, and direct and indirect bilirubin). Clinical parameters evaluated were: alopecia, anorexia, apathy, cachexia, hyperkeratosis, size of lymph nodes, ocular secretion, onychogryphosis, skin lesions and weight loss.
Microtiter plates (high binding, Costar, Corning Incorporated, Corning, New York, USA) were coated with either 100 ng/well of L. (L.) infantum chagasi lysates from amastigotes isolated from spleens of L. (L.) infantum chagasi-infected hamsters as described elsewhere [29] or with 200 ng/well of rLdccys1 in coating buffer (0.05 M Na2CO3/NaHCO3, pH 9.6). The plates were incubated overnight at 4°C and then blocked with 5% powdered skim milk in PBS for 1 h. Canine sera were diluted 1∶500, added to the plate and incubated for 2 h at room temperature. After three washes with 0.05% Tween 20 in PBS, peroxidase labeled antibodies specific to canine IgG diluted 1∶500 were added to the plate for 1 h at room temperature. The plates were then washed three times with 0.05% Tween 20 in PBS, and the reaction developed with 0.5 mg/ml o-phenylenediamine in 0.05 M sodium citrate, pH 4.5, containing 0.03% H2O2. The reaction was stopped by adding 4 N H2SO4, and the absorbance was measured at 492 nm in a Multiskan MS Plate Reader (Labsystems Oy, Helsinki, Finland). The cut-off values were calculated by adding two S.D. values to the mean absorbance of 20 dog normal sera.
The PCR product corresponding to the ORF of the Ldccys1 gene previously obtained [21] was subcloned into the Bam HI and Eco RI restriction sites of the pHis-parallel 3 expression vector in frame with an amino-terminal six histidine tag [30]. The recombinant plasmids were used to transform E. coli BL21 (DE3). Protein expression was carried out by inoculating 500 ml of LB medium containing 100 µg/ml ampicillin with a 25 ml overnight bacterial culture. The suspension was kept in a rotatory shaker at 37°C until reaching log phase (Abs 600 nm = 0.6), and protein expression was then induced with 0.2 mM IPTG for a further 3 h at 37°C. After growth, the recombinant bacteria were pelleted at 4,000× g for 10 min, and the recombinant antigen was then purified from the insoluble inclusion bodies by affinity chromatography using a Ni-NTA Superflow agarose matrix (QIAGEN), according to Skeiky et al. [31]. Purified protein was analyzed by SDS-PAGE and Western blotting using a previously described monoclonal antibody directed against a cysteine proteinase of 30 kDa from L. (L.) amazonensis (MoAb 2E5D3) that cross reacts with an antigen of 30 kDa from L. (L.) infantum chagasi amastigotes [27], [32].
P. acnes was obtained from Instituto Adolfo Lutz, São Paulo, S.P., Brazil and cultured as previously described [33]. Briefly, the bacteria were grown in anaerobic medium (Hemobac, Probac, São Paulo, S.P., Brazil) for 3 days at 37°C and washed by centrifugation. The resulting pellets were suspended in 0.9% saline and subjected to continuous water vapor for 20 min at 120°C. The protein concentration of the suspension was determined by the Bradford method [34].
The treatment protocol was performed with 30 mongrel dogs selected by the parameters described in the ‘Animals’ section. All selected dogs were positive for leishmaniasis in parasitological, serological and biochemical assays and displayed clinical signs of VL. After selection they were randomly divided into three groups of 10 dogs each: one group received three subcutaneous doses of 500 µg rLdccys1 plus 500 µg P. acnes as an adjuvant on the back at a one month interval; the second group received three doses of P. acnes alone; and the third group received PBS. Half of the dogs from each group were followed by monthly clinical examinations until natural death, while the other half were euthanized three months after the end of treatment.
The development of the clinical signs of VL was described by a score that quantifies the number of signs and discriminates mild from severe signs. The scoring system for clinical signs was based on the diameters of small (score = 1), and large lymph nodes (score = 2). For mild and severe weight loss were given scores of 3 and 4, respectively. For alopecia, onychogryphosis, cachexia, apathy, anorexia, hyperkeratosis, ocular secretion and skin lesions were attributed a score of 3. The dogs were evaluated at time zero, when they received the first dose, and monthly thereafter until two months after the end of treatment (month 4). Clinical evaluation was performed by a veterinarian which was blinded to the treatment groups evaluated. Animals were looked for side effects of the recombinant antigen such as local pain, local swelling, vomit and diarrhoea. These signs were followed for 10 days after each antigen injection.
Parasite burden was evaluated in the spleens of all dogs enrolled in the study immediately after animal death using the limiting dilution assay, as previously described [35], and parasite numbers were determined from the highest dilution at which promastigotes could be grown. Briefly, spleens were aseptically excised, weighed and approximately 1–1.5 g of spleen tissue was collected from the mid area and minced into small pieces with sterile scissors within a sterile Petri dish. The tissue was homogenized in 1 ml of PBS and further diluted in 199 medium (Gibco) containing 4.2 mM NaHCO3, 40 mM HEPES, 1.0 mM adenine, 5 µg/ml hemin, 15% fetal bovine serum and 2% male human urine to obtain a final concentration of 1 mg/ml. Serial dilutions ranging from 1 to 1×10−6 were prepared in the same medium under sterile conditions in 96 wells micro plates (Costar plates; Corning, Inc., Corning, NY). After incubation for 10 days at 26°C, plates were examined using an optical microscope at 3-day intervals. The reciprocal of the highest dilution that was promastigote positive was considered to be the concentration of parasites in the spleen tissue processed and the total parasite load was calculated by multiplying this value by the total spleen weight.
Delayed-type hypersensitivity assays (DTH) were performed by intradermal injection of rLdccys1 (10 µg) into the inner surface of the right thigh. As a negative control, each animal received an injection of PBS into the inner surface of the left thigh. The induration diameter was measured by use of a caliper after 24, 48 and 72 h, and each time the values of the saline control were subtracted from the reaction due to the rLdccys1 antigen. Skin reactions with diameter equal or larger than 5 mm were considered positive. The DTH assays were carried out at time zero and one month after animals received the third dose of rLdccys1.
Specific anti-rLdccys1 antibodies IgG, IgG1 and IgG2 isotypes were evaluated by ELISA at time zero and one month after administration of each dose of rLdccys1 in microtiter plates coated with 200 ng/well of rLdccys1 according to the protocol described in “ELISA assays for dog screening” section except that canine sera were diluted 1∶200. Peroxidase labeled antibodies specific to canine IgG were diluted 1∶500 and to IgG1 or IgG2 isotypes (Bethyl Laboratories, Inc., Montgomery, TX, USA) were diluted 1∶2000 and added to the plate for 1 h at room temperature. The plates were then washed and the reaction developed as described above.
Lymphokine concentrations were measured in dog sera at time zero and one month after administration of each dose of rLdccys1 using a double-sandwich ELISA assay (Quantikine Canine IFNγ and IL-10) (R&D Systems). Microtiter plates (high-binding Costar plates; Corning, Inc., Corning, NY) were coated overnight at 4°C with specific mAb, which was directed to each lymphokine tested and used at 100 ng/well. After washing with 0.05% Tween 20/PBS (PBS/T) and blocking with PBS/T containing 5% skim milk for 2 h at 37°C, 100 µl of dog sera diluted 1∶2 were added to wells. Standard curves were generated using recombinant canine IFNγ and IL-10. After incubation overnight at 37°C, plates were washed with PBS/T, and a second antibody specific to each lymphokine was added (biotinylated antibody diluted 1∶250). After 60 min at 37°C, the plates were washed three times in 0.05% Tween 20 in PBS, and the reaction was developed with 0.5 mg/ml of o-phenylenediamine in 0.05 M sodium citrate, pH 4.5, containing 0.03% H2O2. The reaction was stopped by adding 4N H2SO4, and the absorbance was measured at 492 nm in a Multiskan Plate Reader (Labsystems Oy, Helsinki, Finland). Serum concentrations higher than the minimal values obtained from the respective standards were considered to be positive.
One-way ANOVA and Student's t-test were used to determine the significant differences between groups by use of GraphPad Prisma (version 5.0) and P values smaller than 0.05 (P<0.05) were considered significant. The Pearson correlation coefficient was also calculated by use of GraphPad Prisma (version 5.0).
Figure 1A shows that at time zero, selected dogs from each of the groups exhibited low DTH values, whereas one month after the end of treatment the DTH values were significantly higher in dogs that received rLdccys1 compared to controls. The production of total IgG at time zero and one month after administration of each dose of rLdccys1 is shown in Figure 1B, while dosages of IgG1 and IgG2 are illustrated in Figure 1C. Starting from the first dose, there was a significant increase of IgG production in the sera of dogs treated with the recombinant antigen. Starting from the second dose, there was also an increase of IgG production in sera from control dogs, although this increase was lower than that observed in animals treated with rLdccys1. The characterization of IgG subclasses showed that there was a significant increase of IgG2 in animals treated with rLdccys1, whereas IgG1 production was significantly reduced after the second and third doses. In contrast, there was constant production of IgG1 in controls during all treatment time and a significant reduction of IgG2 starting from the first dose.
IFN-γ and IL-10 concentrations were measured by ELISA in dog sera. Figure 2 shows that one month after the first dose there was a low but significant IFN-γ secretion in animals treated with the recombinant antigen. Furthermore, an increased concentration of this lymphokine was observed after the second and third doses of rLdccys1, while a low IL-10 concentration was detected in these animals. In contrast, the control dogs exhibited a low IFN-γ concentration in all periods analyzed, as well as a significant IL-10 increase after they received the third dose of either P. acnes alone or saline. These data indicate that the treatment with the recombinant antigen led to the activation of Th1 responses.
The evolution of clinical signs of VL is shown in Figure 3 and Table S1. At time zero, animals from the three groups exhibited a similar clinical score average (saline, 6.2; P. acnes, 4.9; rLdccys1 plus P. acnes, 6.3). Control animals exhibited a significant increase of clinical signs that indicate disease progression two months after the end of treatment. In contrast, there was no increase of clinical signs in dogs treated with rLdccys1, indicating that they were able to control the disease development. It is also important to emphasize that clinical signs implicated in the severity of canine VL, such as weight loss, cachexia and anorexia, were observed only in control dogs (Table S1).
The safety of the recombinant antigen was also evaluated. The number of dogs showing pain (25%) increased significantly from the first to the third dose (P<0.001). The pain after each treatment dose lasted for 48 h. Local swelling was the most common adverse effect observed (67%) reaching an average diameter of 5 cm and no significant differences among doses 1, 2 and 3 were noted. Furthermore, in all dogs the local swelling reaction was transient and decreased 24 h after each dose, disappearing five days after injection. Dogs did not vomit or present diarrohea.
The correlation between lymphokine production and clinical score averages 30 days after the end of treatment with rLdccys1 is shown in Figure 4. The correlation between the production of IFN-γ and the clinical score averages is negative because there was a significant increase of IFN-γ followed by low clinical scores in dogs treated with rLdccys1, whereas in the controls the low secretion of IFN-γ was correlated with an increase in clinical signs (Figure 4A). In contrast, there is a direct correlation between IL-10 production and the clinical score average. A significant increase in IL-10 and clinical score averages was observed in controls, while there was a decrease in IL-10 secretion followed by low clinical scores in dogs treated with rLdccys1 (Figure 4B).
The survival of controls and dogs treated with rLdccys1 was followed until the natural death of half of the L. (L.) chagasi-infected animals. Most of the dogs that received saline died between 2.6 and 4.4 months after the end of treatment. Among these animals, one was observed to have the lowest clinical score average and survived until 6 months after treatment. Among the dogs treated with P. acnes, three died between 2.8 and 5.3 months, while two of them survived 6.6 and 6.7 months after treatment. In contrast, dogs treated with rLdccys1 died after 12.3 to 14.2 months, with a mean survival time two times higher than the controls (Figure 5A). Figure 5B shows the parasite load evaluated by limiting dilution analysis in dog spleens after death. Animals treated with rLdccys1 exhibited a seven-log reduction in parasite burden compared to controls that received either saline or P. acnes alone.
The other half of the dogs that were euthanized three months after the end of treatment showed a lower parasite burden compared to animals followed until natural death. However, the parasite burden reduction in rLdccys1-treated dogs was not significantly different between euthanized and non-euthanized animals (data not shown).
The main problems that face the treatment of leishmaniasis are toxicity, high cost and parasitic resistance to leishmanicidal drugs currently available on the market. This scenario is even more pronounced in the case of canine VL due to the low efficacy of the drugs currently in use for treatment of infected dogs [36]. All of these issues have pointed to the immunological stimulation as an attractive option for the treatment of leishmaniasis [37]. With this rationale, the potential of a recombinant cysteine proteinase of L. (L.) infantum chagasi, rLdccys1, for the treatment of canine VL was investigated in this study. The usefulness of this recombinant antigen was first demonstrated in screening naturally infected dogs selected for this study. Similar values were observed when either rLdccys1 or L. (L.) infantum chagasi extracts were used as antigens in ELISA assays for evaluation of humoral responses (data not shown). These findings corroborate previous results that demonstrated the high sensitivity and specificity of rLdccys1 for the diagnosis of canine and human VL [21], [23]. In this study, a significant increase of DTH responses was observed after treatment with rLdccys1, supporting our previous results on the use of rLccys1 to distinguish between asymptomatic and symptomatic canine VL [23].
Data on antibody production showed an increase of total IgG in all animal groups during the treatment; however this increase was higher in the rLdccys-treated dogs. Furthermore, rLdccys1-treated dogs had increased IgG2 production and decreased IgG1, while control animals exhibited lower antibody levels with predominance of IgG1. Although the IgG1 and IgG2 subclasses have been used as more reliable indicators of the CVL status than total IgG [38], recently, the functional characterization of canine IgG subclasses raised doubts about the correlation of IgG subclasses with Th1 or Th2 responses [39]. However, our findings are compatible with those that showed high levels of IgG1 anti-Leishmania antibodies associated with the development of clinical signs in L. (L.) infantum chagasi-infected dogs, while IgG2 antibodies appear to be associated with asymptomatic infection [40]. Furthermore, protective responses in dogs vaccinated with the recombinant A2 protein appear to be associated with increased levels of total IgG and IgG2 but not with those of IgG1 anti-A2 antibodies [41]. Our data on DTH and humoral responses in rLdccys1-treated dogs are also compatible with those reported in dogs infected with L (L.) infantum chagasi subjected to immunotherapy with the Leishmune vaccine [18], [20].
The treatment with rLdccys1 resulted in a significant increase of IFN-γ, reduction in IL-10, and significantly less progressive clinical disease than control groups. In contrast, the control dogs presented the opposite profile of these lymphokines and a significant increase of clinical signs until four months after treatment. Indeed, a significant negative correlation was found between the number of clinical signs and IFN-γ production in dogs treated with rLdccys1, whereas a positive correlation was observed between the production of IL-10 and an increase of clinical signs in control dogs. These results strengthen the implication of IFN-γ and IL-10 in control and progression, respectively of canine VL. The induction of Th1 cells producing IFN-γ, IL-2 and TNF-α has been associated with protection against canine VL [42], [43]. The activation of macrophages by IFN-γ to kill intracellular amastigotes via the L-arginine nitric oxide pathway is the main effector mechanism involved in the protective immune responses of dogs infected with L. (L.) chagasi infantum [44], [45]. In contrast, IL-10 has been correlated with disease progression and an increase in IL-10 levels; this was observed in the spleens of dogs naturally infected with L. (L.) infantum chagasi [46]. IL-10 mRNA transcripts were detected in Con A-stimulated PBMC derived from dogs with VL clinical signs [43], [47]. It is worth noting that among the rLdccys1-treated dogs, the first animal that died exhibited a higher clinical score mean at screening, as well as a lower level of serum IFN-γ one month after the end of treatment (data not shown). These findings indicate that the effectiveness of the rLdccys1 treatment is dependent on the disease progression at the time of inclusion in the study. It is possible that sick animals with lower levels of IL-10 respond better to antigen stimulation. Similar results were observed in dogs developing severe VL that did not respond to Leish-111f vaccine treatment [17]. These considerations point to the treatment of asymptomatic dogs with rLdccys1 and a potential more pronounced Th1 response in these animals.
It is also important to emphasize the choice of the adjuvant used in the present study. P. acnes treatment elicits a type-1 (Th1) immune response involving IL-12 and IL-18 that induces IFN-γ release, enhancement of the IgG2a switch and Th2 expansion inhibition [25], [26]. Administration of killed P. acnes as an adjuvant increased the resistance to infection by Trypanosoma cruzi [48]. In leishmaniasis, the treatment with P. acnes led to the control of L. (L.) major infection in BALB/c mice [49]. Murine vaccination with the A2 antigen from L. (L.) donovani plus P. acnes resulted in a mixed Th1 and Th2 response with a predominance of Th1 responses after the homologous challenge, as well as a significant parasite burden decrease in immunized animals [50]. Our previous data on immunization of BALB/c mice with either the native or recombinant Ldccys1 plus P. acnes followed by challenge with L. (L.) infantum chagasi also showed a predominant Th1 response and a significant parasite burden decrease in immunized animals [24], [27]. Immunization of dogs with rLdccys1 plus P. acnes also resulted in a significant protection against L. (L.) infantum chagasi infection (unpublished data).
The participation of CD8+ lymphocytes in protective immune responses triggered in dogs treated with rLdccys1 was not addressed in this study but cannot be overlooked. Analysis of the predicted amino acid sequence of the L. (L.) infantum chagasi Ldccys1 gene cloned previously showed one potential MHC class I epitope for CD8+ lymphocytes in addition to two MHC class II epitopes [21]. The involvement of CD8+ lymphocytes in canine VL has been demonstrated [51] and increased levels of these cells appear to be the major phenotypic feature of asymptomatic disease [52]. Enhanced expression of CD8+ lymphocytes was also observed in L. (L.) infantum chagasi-infected dogs after treatment with the Leishmune vaccine and resulted in a significant reduction of VL clinical signs and parasite burden levels [18], [20].
Immunotherapy with rLdccys1 increased the survival time of L. (L.) infantum chagasi-infected dogs. Comparable results were reported following the use of Glucantime and the recombinant Leish-110f vaccine for treatment of dogs naturally infected with L. (L.) infantum chagasi [16]. Contrary to our findings, animals subjected to that immunotherapeutic protocol had a reduced number of deaths. Nevertheless, treated animals were followed until 180 days after treatment, while in our study animals were monitored until their natural death, and all of the rLdccys1-treated dogs were alive up until one year after treatment. Use of the Leish-110f vaccine in infected dogs treated for longer period of time also resulted in lower death rates compared to our results [17]. Again, this study was different from our treatment protocol because the authors emphasized the advantage of a weekly vaccine schedule over antigen administration with 3 or 4 weeks intervals [17]. Higher survival rates were also observed in dogs treated with the Leishmune vaccine [18]–[20]. It is important to highlight that the differences among previous studies on immunochemotherapy and the present study, including treatment schedules, adjuvant, geographical factors, and disease status of the dogs, hinder the comparison with our data. It is still important to mention that only symptomatic dogs with a high parasite burden were selected for our study and presumably treatment with rLdccys1 of animals with lower parasite burden may lead to improved results.
Despite the high parasitism levels found within the infected dogs enrolled in the present study, there was a significant lesser parasite burden in dogs treated with rLdccys1 compared to controls. Nevertheless, dog cure and a pronounced decrease of vector infectivity are desirable goals in regions of high VL endemicity. In this context, we believe that the use of booster doses of rLdccys1 associated to allopurinol, the drug recommended by WHO to treat CVL [2], are promising to improve the effectiveness of treating CVL with this recombinant antigen.
In conclusion, our findings showed the potential of rLdccys1 as an additional tool for immunotherapy of canine VL and support further studies to evaluate its healing efficacy with a larger number of dogs, as well as in different regions of VL incidence.
|
10.1371/journal.pgen.1004426 | Identification of Late Larval Stage Developmental Checkpoints in Caenorhabditis elegans Regulated by Insulin/IGF and Steroid Hormone Signaling Pathways | Organisms in the wild develop with varying food availability. During periods of nutritional scarcity, development may slow or arrest until conditions improve. The ability to modulate developmental programs in response to poor nutritional conditions requires a means of sensing the changing nutritional environment and limiting tissue growth. The mechanisms by which organisms accomplish this adaptation are not well understood. We sought to study this question by examining the effects of nutrient deprivation on Caenorhabditis elegans development during the late larval stages, L3 and L4, a period of extensive tissue growth and morphogenesis. By removing animals from food at different times, we show here that specific checkpoints exist in the early L3 and early L4 stages that systemically arrest the development of diverse tissues and cellular processes. These checkpoints occur once in each larval stage after molting and prior to initiation of the subsequent molting cycle. DAF-2, the insulin/insulin-like growth factor receptor, regulates passage through the L3 and L4 checkpoints in response to nutrition. The FOXO transcription factor DAF-16, a major target of insulin-like signaling, functions cell-nonautonomously in the hypodermis (skin) to arrest developmental upon nutrient removal. The effects of DAF-16 on progression through the L3 and L4 stages are mediated by DAF-9, a cytochrome P450 ortholog involved in the production of C. elegans steroid hormones. Our results identify a novel mode of C. elegans growth in which development progresses from one checkpoint to the next. At each checkpoint, nutritional conditions determine whether animals remain arrested or continue development to the next checkpoint.
| Organisms in the wild often face long periods in which food is scarce. This may occur due to seasonal effects, loss of territory, or changes in predator-to-prey ratio. During periods of scarcity, organisms undergo adaptations to conserve resources and prolong survival. When nutrient deprivation occurs during development, physical growth and maturation to adulthood is delayed. These effects are also observed in malnourished individuals, who are smaller and reach puberty at later ages. Developmental arrest in response to nutrient scarcity requires a means of sensing changing nutrient conditions and coordinating an organism-wide response. How this occurs is not well understood. We assessed the developmental response to nutrient withdrawal in the nematode worm Caenorhabditis elegans. By removing food in the late larval stages, a period of extensive tissue formation, we have uncovered previously unknown checkpoints that occur at precise times in development. Diverse tissues and cellular processes arrest at the checkpoints. Insulin-like signaling and steroid hormone signaling regulate tissue arrest following nutrient withdrawal. These pathways are conserved in mammals and are linked to growth processes and diseases. Given that the pathways that respond to nutrition are conserved in animals, it is possible that similar checkpoints may also be important in human development.
| The development of multicellular organisms requires the coordinated differentiation and morphogenesis of multiple cell types that interact to form functional tissues and organs. In favorable environmental conditions, development proceeds in a largely stereotyped pattern. When faced with adverse conditions, tissue growth may slow or arrest until the environment improves [1]–[3]. The most critical environmental factor that regulates development is nutrient availability. Organisms can modulate growth programs in response to changing nutritional conditions [4], although the mechanisms through which organisms sense changes in nutrient availability and alter diverse cellular processes in a coordinated manner are incompletely understood.
The nematode Caenorhabditis elegans is a powerful model for understanding the effects of nutrition on development due to its short life cycle (3–4 days from embryo to adult), simple cellular make-up, and highly stereotyped development. The postembryonic development of C. elegans entails passage through four larval stages (L1–L4) that are separated by molts, before reaching reproductive adulthood. Two alternative pathways of development exist in C. elegans: continuous passage through the four larval stages, or entry into an L3 dauer stage, a growth-arrested state characterized by altered body morphology, elevated stress resistance, and prolonged survival [3]. Entry into dauer is initiated late in the L1 stage in response to unfavorable environmental conditions, in particular high population density, high temperature, and reduced nutrient availability [5]. Additional points of arrest in response to poor nutritional conditions have been identified early in the C. elegans life cycle and in adults. Animals that hatch in the absence of food undergo L1 arrest [6], [7], and animals reared from hatching on a limited supply of heat-killed bacteria arrest in the L2 stage [8]. Finally, adult C. elegans arrest embryo production and shrink their germlines following removal of food [9], [10].
Studies on dauer and L1 arrest have revealed critical roles for the insulin/insulin-like growth factor (IGF) signaling pathway in sensing the nutritional environment and regulating entry into arrest [7], [11], [12]. In C. elegans, insulin-like peptides are generated during feeding and signal through DAF-2, the insulin/IGF receptor. Activation of DAF-2 leads to the phosphorylation and cytoplasmic sequestration of DAF-16, a forkhead box O (FOXO) transcription factor. During conditions of low nutrition, the DAF-2-mediated phosphorylation of DAF-16 is reduced, allowing DAF-16 to enter the nucleus and transcriptionally regulate genes implicated in developmental arrest [7], [13]–[16]. Mutant animals with reduced daf-2 function may arrest in the L1 stage or form dauers constitutively (Daf-c phenotype) [11], whereas daf-16 null mutants continue development past the wild type timing of L1 arrest and are defective in dauer formation (Daf-d phenotype) [2], [12].
In worms, insects, and mammals, insulin-like signaling affects the production of steroid hormones, lipophilic molecules that bind to nuclear hormone receptors and induce cellular responses [17]. In C. elegans, bypassing dauer formation requires the bile-acid like steroid hormone dafachronic acid (DA) [18]. DAF-9, a cytochrome P450 ortholog, is required for the production of DAs, and daf-9 null animals are Daf-c [19]–[21]. The effects of DAF-9 on dauer formation are mediated by DAF-12, a nuclear hormone receptor that binds DAs [18]. The steroid hormone pathway functions downstream of insulin-like signaling during dauer formation, as daf-12 Daf-d alleles suppress the Daf-c phenotype of daf-2 mutants [11].
Despite extensive work on dauer and other arrests, questions remain about the response of C. elegans to nutritional scarcity. Among these are whether the arrests in L1, L2, dauer, and the adult represent unique periods of the life cycle during which animals are sensitive to their nutritional environment, or if arrest can also occur at other times. It is also not known whether bypassing the opportunity to form a dauer leads to continuous development to adulthood, or whether additional opportunities exist to arrest development when faced with nutrient deprivation. Finally, the mechanisms through which numerous tissues and cellular processes are able to coordinately arrest in response to nutrient withdrawal are not well understood.
We sought to address these questions by examining the response of C. elegans to nutrient deprivation during the late larval stages (L3 and L4), after the opportunity to form a dauer has been passed. Several tissues that contribute to the reproductive system undergo differentiation, growth, and morphogenesis during the L3 and L4 stages, making this period an ideal time to determine how ongoing developmental processes are affected by nutrient deprivation. By removing animals from food at different times, we show that specific checkpoints exist in the early part of the L3 and L4 stages that restrict progression through the larval stage and systemically arrest the development of diverse tissues and cellular processes. Insulin-like signaling regulates the response to nutrient deprivation in the L3 and L4 stages through cell-nonautonomous DAF-16 activity in the hypodermis (skin), and functions to suppress DAF-9–mediated signaling activity. Our results identify a mode of metazoan growth in which development proceeds from checkpoint to checkpoint. At these checkpoints, nutritional conditions determine whether animals remain in an arrested state or continue development to the next checkpoint.
To study the effects of nutrient deprivation on tissue development during the late larval stages, we first focused on the hermaphrodite vulva, which develops through a stereotyped pattern of cell specification, cell division, and morphogenesis during the L3 and L4 stages (Fig. 1A) [22]. The vulva derives from three epidermal cells, P5.p–P7.p, which are specified early in the L3 stage to either the 1° vulval precursor cell (VPC) fate (P6.p) or the 2° VPC fate (P5.p and P7.p) (Fig. 1A). The VPCs undergo three rounds of cell division in the L3 stage to generate 22 cells, which differentiate into seven vulval subtypes, vulA–vulF. In the L4 stage, the 22 vulval cells undergo morphogenetic processes that include invagination, migration, and cell-cell fusion (Fig. 1A) [22]. Proper development of the vulva requires the uterine anchor cell (AC), which invades in the mid L3 stage across basement membranes separating the uterine and vulval epithelia to form a connection between the tissues [23]. The AC remains at the vulval apex after invasion until fusing with surrounding uterine cells in the mid L4 stage (Fig. 1A).
We examined the effects of nutrient deprivation on vulval development by growing a synchronized population to late in the L2 stage, prior to the onset of vulval formation, and removing animals from food (Fig. 1B). Part of the population was returned to food to serve as controls, with the remainder maintained in M9, a buffer lacking a carbon source. In addition to the vulva, we also assessed the onset of molting (observable by cuticle covering the mouth; see Fig. 1A, bottom left), which serves as a marker for the transition between larval stages. Results of the experiment are depicted graphically in Fig. 1B; raw data and results of triplicate assays are in Fig. S1. The control group that was returned to food progressed through the stages of vulval development with the predicted timing. The group that remained deprived of food molted into the L3 stage and uniformly arrested prior to the first VPC divisions. No VPC divisions were observed after 10 days in the absence of food (Fig. 1B). Arrested animals were phenotypically distinct from dauer larvae, which arrest after molting into a specialized L3 dauer stage (Fig. S2). When animals were returned to food after 8 d, 97.5% of the population (n = 200) continued development to adulthood, demonstrating that animals retain the capacity to resume development upon re-introduction of food. The median survival of animals under the experimental conditions used was 11.7±1.2 d (n = 3 trials).
In C. elegans, removal of the germline either through ablation or genetic mutation extends lifespan and maintains adult somatic tissues for a longer duration in a youthful state [24]–[26]. This suggests the possibility of a soma-germline tradeoff in which resources are allocated away from germ cells to somatic tissues. We asked if the absence of a germline could promote the continued development of somatic tissues by growing glp-1(e2144) mutants, which do not proliferate germline progenitor cells when reared at 25°C, to late in the L2 stage and removing them from food. We found no difference in the timing of arrest, as all glp-1(e2144) animals (n = 100) arrested prior to VPC divisions. These results demonstrate that the absence of the germline does not alter the timing of somatic tissue arrest in response to nutrient removal.
In addition to cell divisions, fate specification of the vulval cells was also examined in L3-arrested animals. Vulval fates are specified between the late L2 and early L3 stages, when an inductive LIN-3/EGF signal from the AC and lateral LIN-12/NOTCH signaling between VPCs combine to specify the 1° fate in P6.p and 2° fates in P5.p and P7.p [27], [28]. To determine the state of VPC specification in arrested animals, we examined a marker of 1° fate, egl-17>GFP [29], and a marker of 2° fate, lip-1>NLS-GFP [30] (see Materials and Methods for description of transgene nomenclature). All arrested animals expressed egl-17>GFP exclusively in P6.p, and 93% of animals expressed lip-1>NLS-GFP at elevated levels in P5.p and P7.p (n = 30 per assay; Fig. 1C), demonstrating that arrest occurred after 1° and 2° VPC specification. This contrasts with dauer larvae, which are not stably specified to a VPC fate [31], [32]. Coupled with the absence of VPC divisions, these results suggest that, when removed from food late in the L2 stage, vulval development arrests early in the L3 stage in a manner that is distinct from dauer arrest.
The uniform arrest of vulval development early in the L3 stage suggested that a specific developmental checkpoint existed at this time. To determine if this was the case, we asked whether vulval development could arrest at later times in the L3 stage. A synchronized population was grown for 28 h to the mid L3 stage and removed from food. At the time of food removal, 84% had undergone one VPC division, indicating that they had bypassed the L3 arrest point (Fig. 2A; Fig. S1). Animals removed from food continued through the L3 stage, molted into L4, and arrested in L4 after completion of VPC divisions (Fig. 2A). After 48 h in the absence of food, 94% of the population was arrested after VPC divisions in the L4 stage. The remaining 6% of animals was arrested in the L3 stage prior to VPC divisions, and likely represent the youngest members of the population that failed to bypass early L3 arrest. No animals were identified at intermediate stages between the two arrest points, demonstrating that bypass of the L3 arrest point led to invariant progression to the L4 arrest point (Fig. 2A). We examined the effect of the germline on L4 arrest by removing glp-1(e2144) mutants grown at 25°C from food in the mid L3 stage, and found that all animals arrested in early L4 similar to wild type (n = 100). The median survival of populations removed in the mid L3 stage was 11.0 d (n = 3 trials).
All L4-arrested animals completed vulval cell divisions (n = 30), suggesting that arrest in vulval formation could occur at a precise developmental time rather than in a variable manner. To test this notion, we first examined the reporter gene egl-17>GFP, which is expressed in 1° VPC progeny early in the L4 stage and shifts to 2° VPC progeny in mid L4 (Fig. 1A). Expression of egl-17>GFP was exclusively in 1° VPC progeny in arrested animals (Fig. 2B), supporting the hypothesis of a precise timing of arrest early in the L4 stage. A second marker for L4 stage timing in vulval development is cell-cell fusions. Fusions occur between homotypic cells (i.e., vulA with vulA), starting with vulA cells shortly after terminal cell divisions and continuing two hours later with vulC cells (Fig. S3) [33]. Examination of a strain expressing GFP-tagged AJM-1, an apical-membrane–localized protein that delineates the boundaries of vulval cells [33], [34], showed that all L4-arrested animals had undergone vulA fusions but not vulC fusions (Fig. S3). Importantly, the same timing of arrest between vulA and vulC fusions occurred in 97% of the population when animals were grown for an additional four hours prior to removal from food (n = 30 per assay; Fig. S3), demonstrating that the timing of arrest in vulval development in the L4 stage is largely independent of feeding duration. Based on the nutrient removal experiments, we conclude that specific checkpoints exist early in the L3 and L4 larval stages that arrest vulval development at precise developmental times.
Only a single checkpoint on vulval development was identified in the L3 stage, and we wanted to determine whether this was also the case with the L4 stage. Animals were grown on food to the early-to-mid L4 stage and developmental progression examined following food removal (Fig. 2C; Fig. S1). After 48 h in the absence of food, 96% of the population had progressed into adulthood, as evidenced by eversion of the vulva (Fig. 2D), with the remaining animals arrested early in the L4 stage, and no animals at intermediate times (Fig. 2C). Arrest occurred in nearly all adult animals (99/100) prior to oogenesis. Taken together, the nutrient removal assays show that arrest in C. elegans vulval development occurs only at precise checkpoints early in the L3 and L4 stages, and that passage through one checkpoint leads invariantly to progression through the larval stage to the next checkpoint.
There are two alternative possibilities for the timing of arrests observed in vulval formation in the L3 and L4 stages. The first is of a tissue-autonomous program in which arrests occur only at specific times in the developmental program, either prior to cell divisions (early L3) or after cell divisions (early L4). The second is of a global timing mechanism that arrests vulval development at precise times early in each larval stage. To determine which of these was correct, we examined hbl-1(ve18) mutant animals, which have precocious VPC divisions that occur as early as the late L2 stage (Fig. 3A) [35]. We hypothesized that if the vulval cells were regulated by an autonomous program, then shifting the time of cell divisions relative to the L3 larval stage would not affect the all-or-none pattern of cell divisions. If instead a global timing mechanism directed the arrest of vulval development, then cell divisions would be predicted to arrest upon reaching the L3 larval stage checkpoint. Results of the experiment show that after removal from food late in the L2 stage, P6.p divisions continued but stopped prior to completion (Fig. 3B–C). Only 43% of the population was arrested either prior to or after cell divisions, with the remainder at intermediate stages of division (Fig. 3B). These results support the idea of a global timing mechanism that acts on vulval development to arrest it at specific times early in the larval stage.
The experiments on vulval development suggested that checkpoints exist at particular points in the larval stage. We wanted to explore this question in more detail by examining progression through the larval stage in the absence of food. Each C. elegans larval stage comprises a period of foraging for food that lasts for several hours, followed by an approximately two-hour period of lethargus during which pharyngeal pumping stops and animals do not feed. At the end of lethargus, C. elegans undergo molting, the detachment (apolysis) and shedding (ecdysis) of the existing cuticle [36]. We first asked if animals removed from food during the period of foraging underwent lethargus, and found that both the onset and duration of lethargus were similar to a control population that was maintained on food. Further, animals exited lethargus and resumed pharyngeal pumping for at least 24 h after removal from food (Fig. 4A). These results show that lethargus, a key feature of the larval stage, is maintained in the absence of food.
We next examined how nutrient deprivation affected the molting cycle, the oscillatory pattern of gene expression and cuticle replacement that occurs in each larval stage. Cuticle components are synthesized starting in the mid-larval stage and deposited underneath the existing cuticle, which is shed at the end of the larval stage [37], [38]. The timing of the checkpoints in the early part of larval stage suggested that arrest could occur after molting and prior to new cuticle synthesis. We first asked whether this was the case by examining the execution of the molt following removal of food. We found that all L3- and L4-arrested animals completed ecdysis (Fig. 4B), although 17% of adult-arrested animals remained attached to the L4 cuticle after 48 h in the absence of food (n = 100 animals per assay). It is possible that larger animals may not be able to fully shed cuticle in the absence of sufficient feeding. Despite this defect, animals were viable and resumed pharyngeal pumping, with the L4 cuticle remaining attached only in the tail region. These results demonstrate that molting is successfully executed in most instances upon passage through a checkpoint.
We then explored the second part of our hypothesis that arrest occurred prior to new cuticle synthesis. To do this, we examined the expression pattern of mlt-10, a gene required for proper execution of the molt [38], [39]. mlt-10 mRNA increases in the mid-larval stage at the time of new cuticle synthesis, peaks during the molt, and declines upon completion of molting. A destabilized reporter gene, mlt-10>GFP-PEST, recapitulates this oscillatory mRNA expression pattern and serves as a marker for progression through the larval stage [38], [39]. A population of mlt-10>GFP-PEST–expressing animals was removed from food late in the L2 stage and a portion of the population returned to food to serve as controls. The fed and nutrient-deprived groups were then examined for reporter gene expression over an 8 h period. Results show that expression levels were similar in the two groups as they molted and entered the L3 stage (Fig. 4C). Approximately 4 h after molting, the control group increased gene expression, indicating initiation of the L3 molting cycle. The nutrient-deprived group failed to increase expression, however, demonstrating that it had arrested prior to initiation of the L3 molting cycle (Fig. 4C). Similar results were observed when animals were removed from food late in the L3 stage (data not shown). The loss of mlt-10>GFP-PEST expression was unlikely to be due to general transcriptional silencing during nutrient deprivation, as past research has shown that several transcriptional reporters similarly tagged with PEST motifs maintain expression during L1 arrest [40]. Collectively, these results demonstrate that C. elegans arrest development during the L3 and L4 stages at a specific point after molting and prior to new cuticle synthesis.
Our results identified nutrient-sensitive developmental checkpoints in the early part of the L3 and L4 larval stages. We sought to determine the amount of feeding required to pass the checkpoints. To achieve the greatest degree of synchronization and most accurate measurement of timing, individual animals were isolated during ecdysis, the final 10–15 minutes of molting that precede foraging [36]. Animals undergoing ecdysis were either removed from food or allowed to feed for additional 30 min intervals (Fig. S4). Feeding for 30–60 min after ecdysis was required for most animals to pass the L3 and L4 checkpoints within 24 h after food removal, and 90 minutes of feeding resulted in all animals passing the checkpoints (Fig. S4). We conclude that a sufficient duration of feeding is required after molting to advance past the L3 and L4 stage checkpoints.
The insulin-like signaling pathway is a key regulator of growth in response to nutrition [41]. We wanted to determine if insulin-like signaling regulated arrest in the L3 and L4 stages following nutrient removal. We first asked if daf-16, a FOXO transcription factor that is a major target of insulin-like signaling and is required for the proper timing of L1 arrest and dauer formation [7], [12], played a role in L3 and L4 arrest. Animals with the null mutation daf-16(mu86) were removed from food late in the L2 stage, and the developmental stage assessed over time by examination of the vulva and molt. The pattern of growth by 8 h after food removal was similar to wild type: all animals had molted into L3 and 97% were in the early L3 stage (Fig. 5A; raw data and results of replicate assays are in Fig. S5). By 24 h after removal from food, however, 63% of the population had progressed past the L3 checkpoint (compared with 0% of wild type animals removed from food at a similar time; Fig. 1B), ultimately arresting early in the L4 stage (Fig. 5A). A second experiment was performed with daf-16(mu86) animals removed from food late in the L3 stage. Again, the absence of daf-16 caused animals to bypass arrest, with 72% of the population progressing to adulthood after 48 h (Fig. 5B; Fig. S5). The time in the larval stage at which daf-16(mu86) animals arrested was similar to wild type, based on the absence of VPC divisions in L3-arrested animals, the completion of divisions in L4-arrested animals, and no animals at intermediate stages of division (n = 30; Fig. 5C).
In the presence of food, DAF-16 activity is inhibited by a signaling pathway downstream of DAF-2, the insulin/IGF receptor. We hypothesized that animals with reduced DAF-2 function would require a longer duration of feeding to inhibit DAF-16 activity and progress through the L3 and L4 larval stages. To test this hypothesis, we examined the L3 and L4 development of a temperature-sensitive daf-2 mutant, daf-2(e1370), which is Daf-c at 25°C but develops to adulthood at 15°C [11]. Animals were grown at the permissive temperature of 15°C to the mid-L2 stage, bypassing the opportunity to form a dauer, and shifted to the restrictive temperature of 25°C for an additional 24 h feeding (Fig. 5D). Following this regimen, 25% of the population was in the early L3 stage and 42% was in early L4 stage. In contrast, a control wild type population had progressed to the L4/adult molt or beyond (Fig. 5D). The high proportion of the population in the early L3 and early L4 stages suggests that prolonged pausing at the checkpoints may be a factor in the delayed development of daf-2(e1370) animals. This delayed development required the presence of daf-16, as daf-16(mu86); daf-2(e170) double mutant animals advanced through the L3 and L4 stages at a rate comparable to wild type (Fig. 5D). Taken together, the results of the daf-16 and daf-2 experiments demonstrate a role for the insulin-like signaling pathway in regulating progression through the L3 and L4 developmental arrest checkpoints in response to nutritional conditions.
Previous work has shown that DAF-16 functions cell-nonautonomously to regulate multiple physiological processes, including dauer formation, lifespan extension, germline proliferation, and metabolism [42]–[44]. We asked if DAF-16 similarly functioned cell-nonautonomously to regulate L3 and L4 arrest. Plasmids that contained daf-16 cDNA tagged at the N-terminus with GFP and expressed under the daf-16 or tissue-specific promoters [42] (Fig. 6A) were injected into daf-16(mu86); unc-119(ed4) double mutant animals along with an unc-119 rescue plasmid. Animals harboring an extrachromosomal array of the plasmids were identified by rescue of the unc-119(ed4) movement defect and validated by examination of GFP expression (Fig. S6).
When expressed from the daf-16 promoter, GFP::DAF-16 protein localized to neurons, hypodermis, intestine, and body wall muscles (Fig. S6), and rescued the daf-16(mu86) phenotype, in which animals bypass L3 arrest and continue development to the L4 stage (Fig. 6B). GFP::DAF-16 expressed from tissue-specific promoters for neurons (unc-119 and unc-115), muscle (myo-3), and intestine (ges-1) [42] failed to rescue the daf-16(mu86) bypass phenotype. Only GFP::DAF-16 expression from a hypodermis-specific promoter (col-12) significantly rescued the daf-16(mu86) phenotype (Fig. 6B). Similar results were obtained in assays examining bypass of L4 arrest (data not shown). The lower efficiency of rescue by col-12>GFP::DAF-16 compared to daf-16>GFP::DAF-16 could be due to reduced expression of the transgene following removal from food: col-12>GFP::DAF-16 expression decreased 71% following 2 d in the absence of food, whereas daf-16>GFP::DAF-16 expression increased more than twofold during this time (Fig. S7). As in other assays, vulval development was used as the primary marker for developmental stage. Animals that were rescued for the L3 bypass phenotype by daf-16>GFP::DAF-16 or col-12>GFP::DAF-16 did not have detectable GFP::DAF-16 expression in the vulva, consistent with cell-nonautonomous DAF-16 activity regulating L3 and L4 development.
To complement these studies, we carried out tissue-specific RNAi of daf-16. Reducing daf-16 specifically in the hypodermis reproduced the phenotype of systemic loss of daf-16. Targeted reduction of daf-16 in the intestine or muscle did not alter sensitivity to the removal from food (Fig. 6C). We were unable to reduce daf-16 specifically in neurons because of the lower sensitivity of this tissue to RNAi [45], and the inability to directly target neurons by RNAi without off-target effects in the hypodermis [46]. Taken together, the results of daf-16 tissue-specific rescue and RNAi experiments suggest that the hypodermis is a key site of action for the insulin-like signaling pathway in responding to nutritional conditions during the L3 and L4 stages. Our results do not rule out the possibility that DAF-16 functions in other tissues to also regulate L3 and L4 development, either through modulation of hypodermal DAF-16 function or through independent pathways that synergize with hypodermal DAF-16. Previous studies have shown that DAF-16 can function in multiple tissues to regulate dauer formation and metabolism [42], [43], and such a situation could also occur in regulating passage through the L3 and L4 larval stages.
The ability of DAF-16 to affect tissue development cell-nonautonomously implicated the presence of pathways that signal systemically. One such candidate is DAF-9–mediated steroid hormone signaling, which is downstream of insulin-like signaling during dauer formation [11], [19], [21]. A key site of action for DAF-9 during larval development is the hypodermis [19], [21], [47], suggesting that it could similarly function downstream of insulin-like signaling during the L3 and L4 stages. To test this possibility, we depleted daf-9 by dsRNA feeding in daf-16(mu86) animals, and assessed the response to nutrient removal in the L3 and L4 stages. We hypothesized that, if nuclear DAF-16 inhibited L3 and L4 stage progressions through inhibition of DAF-9–mediated steroid hormone signaling, then the bypass of arrest observed in daf-16 null animals would be suppressed by reduction of daf-9. Consistent with this hypothesis, daf-9 dsRNA-fed daf-16(mu86) animals had a 2.6-fold reduction of bypassing L3 arrest and a 1.9-fold reduction of bypassing L4 arrest compared to empty vector controls (Fig. 7A). These results support the idea that the insulin-like signaling pathway regulates DAF-9–mediated steroid hormone production during the L3 and L4 stages.
Since DAF-9 appeared to be involved in generating hormonal signals that promoted progression through the L3 and L4 larval stages, we asked whether increasing the levels of DAF-9 would lead to bypass of arrest in a manner akin to daf-16 null animals. This was tested by examining the response to nutrient removal of a strain overexpressing functional daf-9::GFP (dhIs64) [19]. When removed from food late in the L2 stage, daf-9–overexpressing animals bypassed arrest at high levels, with 90% of the population progressed beyond the early L3 stage after 24 h in the absence of food (Fig. 7B; Fig. S5). In contrast to the phenotype of daf-16(mu86), which paused at the L3 checkpoint before bypassing it (Fig. 5A), daf-9–overexpressing animals continued past the checkpoint with minimal pausing (Fig. 7B). Further, whereas daf-16(mu86) bypassed only one arrest point (Fig. 5A–B), a portion of the daf-9-overexpressing population advanced through both the L3 and L4 arrest points and reached adulthood (Fig. 7B). Thus, the bypass of arrest caused by overexpression of daf-9 is more rapid and robust than that caused by loss of daf-16.
daf-9–overexpressing animals that progressed to adulthood were typically surrounded by undetached cuticle (41/50 animals); in some cases both the L3 and L4 cuticles remained attached (Fig. 7C). The inability to shed cuticle surrounding the mouth led to lethality in a portion of the population within 24 h of food removal (Fig. 7C). In contrast, neither wild type nor daf-16 null animals showed such rapid death. These findings demonstrate that overexpression of daf-9, which forces animals through larval stages in the absence of food, has deleterious effects on the execution of the molt.
Our finding that daf-9 overexpression causes continued development in the absence of food were somewhat surprising since a previous study showed that hypodermal DAF-9::GFP expression is sharply reduced during nutrient deprivation [19]. We examined the expression of hypodermal DAF-9::GFP following removal from food late in the L2 stage, and indeed found a reduction in expression over time (Fig. 7D). Expression persisted at low levels in most animals for at least 8 h after removal from food, however, and was visible in some animals even after 24 h (Fig. 7D). These results suggest that, when expressed at elevated levels, enough DAF-9 protein remains in the hypodermis to promote continued larval stage progressions in the absence of food. It is also possible that the second site of DAF-9 expression during larval development, the two neuronal XXX cells, also contribute to the bypass of arrest, as expression is not reduced in those cells following food removal [19]. Collectively, our results offer evidence that DAF-9 promotes passage through the L3 and L4 developmental arrest checkpoints.
DAF-9 is required for the synthesis of dafachronic acids (DAs), steroid hormones that bind to the nuclear hormone receptor DAF-12 to promote bypass of dauer formation [18]. In the absence of DAs, DAF-12 regulates entry into dauer though its DNA-binding activity [48]. Because our results showed that genes that regulate dauer formation—daf-2, daf-16, and daf-9—also regulate later larval development, we asked if daf-12 similarly had a role in regulating progression through the L3 and L4 stages downstream of daf-9. We first examined the response to nutrient removal of daf-12(rh61rh411), a null mutant that has a Daf-d phenotype [48]. In contrast to daf-16(mu86) Daf-d mutants, which bypass L3 arrest over 60% of the time (Fig. 5B), no daf-12 null mutants bypassed L3 arrest after 48 h in the absence of food (n = 54). We also performed epistasis experiments by generating daf-12(rh61rh411); daf-16(mu86) double mutant animals. The bypass phenotype of daf-16(mu86) was not suppressed by the loss of daf-12 function, as 71% of double mutant animals bypassed the L3 checkpoint after 48 h in the absence of food (n = 68), similar to the percentage of daf-16(mu86) animals that bypass arrest (Fig. 5B). These results suggest that daf-16 functions independently of daf-12 in regulating L3 stage progression. We also asked whether the phenotype of daf-9 overexpression was suppressed by a null allele of daf-12. When removed from food late in the L2 stage, DAF-9::GFP; daf-12(rh61rh411) animals still bypassed arrest to a high degree: 87% of the population had bypassed L3 arrest by 24 h (n = 61), similar to the 90% observed in DAF-9::GFP animals (Fig. 7B). These results provide evidence that DAF-12 does not regulate L3 and L4 stage progressions, and that DAF-9 promotes bypass of the L3 and L4 checkpoints through a different downstream effector.
Our experiments with vulval formation and the molting cycle showed that checkpoints are present early in the L3 and L4 stages that limit continued development. We took advantage of the fact that several additional tissues undergo developmental processes in the L3 and L4 stages to determine if other tissues are also arrested at the checkpoints. We first examined the uterine AC, which becomes polarized early in the L3 stage, when F-actin and actin regulators localize to the basal, invasive cell membrane [49]. The AC breaches the basement membrane in the mid L3 stage, before fusing with the surrounding uterine cells in the mid L4 stage (Fig. 1A). Examination of a marker of AC polarization, the F-actin probe cdh-3>mCherry::moesinABD, showed that animals removed from food late in the L2 stage arrested early in the L3 stage with polarized ACs, but in no instance did invasion occur (n = 100; Fig. 8A). When removed from food in the mid L3 stage, AC invasion occurred in all animals, yet in no instance did the AC fuse with the surrounding uterine cells (n = 100; Fig. 2B). These results suggest that the developmental program of the AC, similar to that the vulval cells and molting cycle, arrests at the early L3 and early L4 checkpoints.
We next examined the two sex myoblast (SM) cells, which divide three times between the mid L3 and early L4 stages, followed by short-range migrations of terminally divided progeny cells in the L4 stage (Fig. 8B). When animals were removed from food late in the L2 stage, no SM cell divisions were observed after 48 h, indicative of arrest early in the L3 stage. When removed from food in the mid L3 stage, SM cell divisions initiated in all animals and typically divided twice, although in some instances fewer or more cell divisions were observed (Fig. 8B). Animals that were grown on food to late in the L3 stage and arrested in the L4 stage with completed cell divisions did not undergo short-range migrations (Fig. 8B), demonstrating that the L4 morphogenetic program did not advance past the early L4 checkpoint. Although cell divisions of the SM cells are not as tightly regulated as the vulval cells, these results suggest that development of the SM cells is also under the control of the L3 and L4 checkpoints.
We then examined the seam cells, stem cells that divide during the L1–L3 molts, generating an anterior daughter that fuses with the surrounding hypodermal syncytium and a posterior daughter that retains stem-like properties (Fig. 8C). Animals that were removed from food in the L3 stage showed variability in the timing of seam cell arrest. Some seam cells failed to divide; others divided but anterior daughters did not fuse; and in the most advanced animals, daughter cells fused but the adherens junctions that connect seam cells did not re-form (Fig. 8C). Thus, removal of animals from food in the L3 stage causes the arrest of a several aspects of the seam cell division program prior to reaching the L4 checkpoint.
We finally looked at elongation of the gonad, which occurs in a continuous manner from the L2 to L4 stages. When animals were removed from food in mid L3 to cause arrested in the early L4 stage, gonad arm elongation was 35% shorter than in a fed control population of early L4 animals (Fig. 8D), indicating that gonadal elongation arrested prior to the L4 checkpoint. Taken together, these results show that diverse cellular processes arrest following removal of food. Although some tissues had a variable pattern of arrest, in no instance did development continue past the early L3 and early L4 stages, demonstrating that the checkpoints limit tissue development in a systemic manner.
Organisms have the ability to sense their nutritional environment and alter growth and metabolism in response. When nutritional conditions are limiting during development, the effects on tissue maturation must be systemic in nature and temporally coordinated in order to maintain the capacity to form functional organs [1]. We examined the C. elegans response to nutrient deprivation during the L3 and L4 larval stages and have uncovered a means by which different tissues are able to arrest in a coordinated manner. We show that distinct checkpoints are present in the early part of the larval stages that regulate development throughout the organism and arrest a range of tissues and cellular processes. At the L3 and L4 checkpoints, nutritional conditions inform a systemic decision to either remain arrested or continue development. This decision is regulated by insulin-like and steroid hormone signaling pathways (see model, Fig. 9).
Previous work on L1 arrest, dauer, and adult reproductive diapause have shown that, in response to unfavorable nutritional conditions, cellular processes can arrest for extended durations and resume upon re-feeding [3], [7], [9], [10], [16]. The nature of the response to nutrient deprivation at other times in development had not been characterized. By focusing initially on the vulva, which has a stereotyped pattern of development during the L3 and L4 larval stages, we show here that checkpoints are present in the early part of the L3 and L4 stages that arrest tissues throughout the organism. The timing of arrest reflects a specific point in the larval stage after molting and prior to initiation of the subsequent molting cycle. A connection between nutrition and the molting cycle has been described in other ecdysozoans [50], [51]. In insects, for instance, molting to a new larval instar occurs only after a sufficient duration of feeding and attainment of a critical weight [50]. It has been speculated that similar nutritional factors impinge on the endocrine signals that trigger the onset of the molting cycle in C. elegans [38]. Our results provide support for this model of nutritional control of molting cycle commitment.
The response to nutrient deprivation in the L3 and L4 stages is systemic in nature and causes the arrest of multiple tissues and cellular processes. Although all tissues arrested prior to or at the checkpoints, vulval formation and the molting cycle were unique in arresting within very narrow developmental windows and in a uniform manner throughout a population. These tissues may have been under selection to arrest in such a precise way. A properly formed vulva is necessary for mating and egg-laying, and a robust developmental program with minimal variation is important for reproductive fitness [52]. For the molting cycle, the inability to shed cuticle surrounding the head during ecdysis causes rapid lethality, as observed in daf-9–overexpressing animals, making it incumbent to execute the molt. A previous study showed that the buccal cavity, which comprises the anterior-most portion of the pharynx and constrains the amount of food consumed with each pumping event, grows only during molts and not between them, which is thought to increase the amount of food that can be consumed during the larval stage [53]. Certain tissues therefore have distinct patterns of growth that are unique to their functions in development and reproduction.
We show that the insulin-like signaling pathway regulates that connection between nutritional conditions and progression past the L3 and L4 checkpoints. In wild type animals, feeding of 30–60 minutes is required after molting to attain a sufficient threshold for bypassing the larval stage checkpoints. Perturbations of key genes in the insulin-like signaling pathway alter the duration of feeding required to bypass arrest. Reduction in the function of daf-2, the insulin/IGF receptor, increases the amount of feeding, such that animals pause at the L3 and L4 checkpoints and have delayed development through the L3 and L4 stages. Loss of function of daf-16, a FOXO transcription factor that is a key target of the insulin-like signaling pathway [54], decreases the amount of feeding required to bypass the checkpoints.
The bypass of arrest caused by loss of daf-16 is partially suppressed by reduced expression of daf-9, a cytochrome P450 ortholog required for the production of certain C. elegans steroid hormones [18], [55]. This result suggests that DAF-16 regulates arrest in the L3 and L4 stages at least in part through inhibition of steroid hormone signaling. DAF-16 has been shown to inhibit daf-9 expression during cholesterol starvation, an unfavorable growth environment that causes larval arrest [32], [56]. It is possible that late larval stage arrest caused by nutrient deprivation also involves DAF-16 inhibition of daf-9 expression. A key site of action for DAF-16 in regulating L3 and L4 arrest is the hypodermis, where daf-9 is also expressed during larval development [47], further suggesting that DAF-16 regulates daf-9 expression. Collectively, these findings support a model in which DAF-16 inhibits daf-9 expression, and possibly other genes involved in steroid hormone production, to limit progression through the L3 and L4 stages (Fig. 9).
The level of DAF-9 protein is a key determinant of L3 and L4 stage progressions in the absence of food, which is demonstrated by the striking ability of overexpressed DAF-9 to promote continued development through one or two larval stages. The hormonal signaling pathway that functions downstream of DAF-9 in the L3 and L4 stages appears to be different from the pathway that regulates dauer formation. During the dauer decision, the DAF-9 biosynthetic pathway produces dafachronic acids (DAs), which bind to the nuclear hormone receptor DAF-12 to promote bypass of dauer [18]. Our experiments with a daf-12 null mutant failed to show a similar role for DAF-12 in the L3 and L4 stages. Consistent with these results, we have also found that supplementation of M9 buffer with DAs does not promote continued development past the checkpoints after food removal (Schindler & Sherwood, unpublished observations), further implicating a mode of hormone signaling during the L3 and L4 larval stages that is distinct from that during dauer formation.
A key implication from these results is that wild type C. elegans arrest development in the L3 and L4 stages despite possessing a sufficient amount of nutrients to continue further development. This is demonstrated by the ability of animals lacking daf-16 or overexpressing daf-9 to bypass one or even two arrest points and progress through the larval stages in the absence of food. Developmental arrest in wild type animals therefore reflects a decision to halt larval stage progressions rather than a lack of available resources to sustain further development. Continued progression in the absence of food appears to have deleterious consequences, as exemplified by the death and molting defects observed in daf-9–overexpressing animals. Limiting progression through the larval stages when nutritional conditions are poor may allow resources to be conserved for survival and tissue homeostasis during prolonged periods of starvation.
This scenario of sensing the environment and arresting development in response to unfavorable conditions also occurs during the C. elegans dauer decision [3]. Both non-dauer and dauer arrest are regulated by insulin-like and DAF-9 signaling pathways, and studies comparing gene expression in dauer and starved animals have revealed overlap between the two types of arrest [57], [58]. From an evolutionary perspective, it is intriguing to speculate that dauer formation, a nematode-specific developmental diapause, evolved from pathways of starvation-induced arrest that are conserved among metazoans.
Two types of growth have been described for ecdysozoans: continuous, with growth occurring throughout the course of development; and saltational, with growth occurring only at distinct times [53]. In organisms with rigid exoskeletons, growth occurs only at molts, an example of saltational growth. C. elegans, with flexible exoskeletons, grow in a continuous manner through the larval stages [53], [59]. By manipulating the nutritional environment, we show that C. elegans growth has an additional saltational aspect to it, with distinct checkpoints present in the early part of the larval stage. At each checkpoint the nutritional environment informs a systemic decision to either proceed through the larval stage or to remain arrested (Fig. 9). Two key pathways that regulate this developmental decision—insulin-like and steroid hormone signaling—are present throughout metazoans [60], [61], suggesting that the mode of growth control described in this work could be conserved. A greater understanding of the mechanisms of growth control could provide insight into aging and metabolic diseases, which are linked to the dysregulation of developmental pathways important for growth [62], [63]. Our work in C. elegans demonstrates a type of saltational growth from checkpoint to checkpoint that may similarly regulate development and physiology in other species.
Nematodes were reared at 20°C on NGM plates seeded with OP50 E. coli using standard procedures. In the text and figures we refer to linked DNA sequences that code for a single fusion protein using a (::) annotation. For designating linkage to a promoter we use a (>) symbol. The wild type strain N2 and the following mutant strains and transgenes used were: dhIs64[daf-9::GFP], qyEx262[unc-119>GFP::daf-16], qyEx263[daf-16>GFP::daf-16], qyEx264[myo-2>GFP::daf-16], qyEx266[GFP::daf-16], qyEx267[ges-1>GFP::daf-16], qyEx268[unc-115>GFP::daf-16], qyEx292[col-12>GFP::daf-16], kbIs7[nhx-2>rde-1], kzIs9[lin-26>rde-1], kzIs20[hlh-1>rde-1]; LG I: daf-16(mu86), ayIs4[egl-17>GFP], syIs78[ajm-1::GFP]; LGII: qyIs17[zmp-1>mcherry]; LG III: daf-2(e1370), unc-119(ed4); glp-1(e2144); zhIs4[lip-1>NLS-GFP]; LG IV: mgIs49[mlt-10>GFP-PEST], ayIs7[hlh-8::GFP], qyIs10[lam-1::GFP]; LG V: rde-1(ne219), qyIs50[cdh-3>mCherry::moesinABD]; LG X: hbl-1(ve18), qyIs7[lam-1::GFP]; daf-12(rh61rh411).
Images were acquired using either a Zeiss AxioImager A1 microscope with a 10×, 20×, or 100× plan-apochromat objective and a Zeiss AxioCam MR charge-coupled device camera, controlled by Zeiss Axiovision software (Zeiss Microimaging, Inc., Thornwood, NJ), or with a Yokogawa spinning disk confocal microscope mounted on a Zeiss AxioImager A1 microscope using iVision software (Biovision Technologies, Exton, PA). Images were processed in ImageJ (NIH Image) and Photoshop CS6 (Adobe Systems Inc., San Jose, CA). Z-stack projections were generated using IMARIS 6.0 (Bitplane, Inc., Saint Paul, MN).
Quantification of fluorescence intensity was performed on images acquired at identical exposure settings using ImageJ. For quantifying GFP::DAF-16 in the hypodermis, the fluorescence intensity in four nuclei (excluding nucleoli) were averaged per animal. All measurement of nuclear GFP::DAF-16 were taken within 5 min of removal from food to minimize relocalization of DAF-16 into the nucleus. For quantifying DAF-9::GFP, a contiguous area of the hypodermal syncytium that excluded nuclei was measured in a region below the pharynx.
Populations containing gravid adults were hypochlorite treated to release embryos, which hatched in M9 buffer and arrested in L1. The duration of L1 arrest did not exceed 20 h. Populations of L1-arrested animals were plated onto NGM plates seeded with OP50 E. coli that covered at least half the plate to minimize the duration of wandering away from food. Maximum population density was 2500 animals/60 mm dish. Animals were reared at 20°C unless indicated otherwise. For removal from food late in the L2 stage, populations were monitored starting at 22 h post-plating. The assessment of developmental age was made by observation of the gonad (which grows through the L2 stage) and the molt (which covers the mouth during the time of molting, see Fig. 1A). Unless a specific duration of growth is indicated, animals were removed from food when the oldest members of the population were molting, and the remaining members were in the late L2 stage, based on gonad size. Populations that contained greater than 5% L3 animals were not used. N2 and daf-12(rh61rh411) populations typically developed in a synchronized manner; hbl-1(ve18), daf-16(mu86), and daf-9::GFP, populations grew more variably and had a wider spread of developmental ages at the time of food removal. To remove food, 1 ml M9 was added to each plate and gently rocked to dislodge worms with minimal removal of E. coli. Animals were transferred to low-retention Eppendorf tubes and centrifuged for 1 min at 500× g, a speed at which C. elegans sank to the bottom but E. coli remained largely in suspension. Liquid was aspirated, and an additional 1 ml M9 added for 6 total washes. Tests of supernatants found that bacteria were removed by the third wash, based on the inability of the supernatant to form colonies on LB plates. Animals were placed in M9 buffer at 0.5–1 animal/µl in 25-ml glass conical tubes and rotated in a roller drum (New Brunswick Scientific, Enfield, CT) at ambient temperature (22°C). For visualization of ecdysis, animals were removed from food late in the L2 stage, anesthetized in levamisole, mounted on agar pads with sealed cover slips, and maintained for 24 h in a humidified chamber.
Developmental stages from L3 to young adult were assessed using the progression of the vulva (see Fig. 1A) and the molt. The two processes occurred synchronously in both fed and nutrient-deprived animals. Statistical significance of differences in arrest response was determined by two-tailed Fisher's exact test. For tissue-specific daf-16 rescue experiments, percentages of L3- and L4-arrested animals were determined for each promoter-driven GFP::daf-16 strains and compared to the promoterless GFP::daf-16 strain (qyEx266). For tissue-specific daf-16 dsRNA feeding, percentages of L3- and L4-arrested animals were compared between animals fed either daf-16 dsRNA or vector control. A similar comparison was made in daf-16(mu86) animals fed either daf-9 dsRNA or vector control. All assays were repeated in triplicate with n≥50 animals per assay.
Populations of wild type animals were removed from food in either late in the L2 or in the middle of the L3 stage and starved in M9 buffer at an approximate population density of 1 animal/µl. Every 2 d, an aliquot of media containing at least 50 animals was removed using Rainex-coated tips to prevent adherence to the plastic, and plated onto NGM plates. After liquid was absorbed into the plate, animals were determined to be alive if they moved or dead if they did not move upon tail poke. Dead animals had a characteristic rod-like appearance. The median survival was determined as the first day at which 50% of the population was dead, for n = 3 trials.
To test recovery from nutrient deprivation, early L3 arrested animals were plated onto NGM+OP50 after 8 d in the absence of food. After 72 h at 20°C, the population was scored for fertile and nonfertile adults.
Synchronous populations were grown to the L2/L3 or L3/L4 molts. Animals in ecdysis were isolated by the appearance of detached cuticle separated from the body, which was most apparent in the head and tail regions. Individual animals were transferred onto NGM+OP50 plates for 30′–90′ further feeding or placed directly in M9 buffer. Animals were maintained in the absence of food for 24 h and the developmental stage assessed.
Transgenic strains expressing promoter-driven daf-16 cDNA fused at the N-terminus with GFP were generated by injection of the following plasmids: pNL205 (promoterless), pNL206 (unc-119 promoter), pNL209 (daf-16 promoter), pNL212 (myo-3 promoter), pNL213 (ges-1 promoter), pNL216 (unc-115 promoter), and pAS10 (col-12 promoter). With the exception of pAS10, plasmids were gifts of the Kenyon lab and are described elsewhere [42]. pAS10 was generated by PCR amplification of 1.1 kb of col-12 promoter 5′ to the start site, which was cloned into the SnaBI restriction sites in pNL205. GFP::daf-16 plasmids were injected at 100 ng/µl into daf-16(mu86); unc-119(ed4) adults with 50 ng/µl unc-119(+) plasmid. Animals carrying extrachromosomal arrays were isolated by rescue of the unc-119 locomotion defect and the expression of GFP::DAF-16 validated. Although unc-115 has been reported to express in both neurons and hypodermis [64], expression was only detected in neurons. With the exception of qyEx266 (expressing pNL205), which did not possess a gene promoter for GFP::DAF-16, and qyEx292 (expressing pAS10), which sometimes had undetectable or minimal GFP expression in the absence of food, animals carrying the array were identified in nutrient removal assays by GFP expression. qyEx266 and qyEx292 animals were plated on NGM plates lacking food, and those that moved freely (indicating presence of the unc-119 rescue array) were selected for analysis.
The generation and validation of strains sensitive to RNAi in the hypodermis (NR222, rde-1(ne219); lin-26>rde-1); muscle (NR350, rde-1(ne219); hlh-1>rde-1); and intestine (VP303, rde-1(ne219); nhx-2>rde-1) are described elsewhere [30], [65]. Strains were fed either daf-16 or L4440 (vector control) dsRNA starting from L1 arrest, grown to late in the L2 stage, and removed from food. After 2 d in the absence of food, the developmental stage was scored.
To assess AC polarization, the average fluorescence intensity was determined from three-pixel-wide linescans drawn along either the basal or apicolateral membranes of Z-stack projections. To determine the movement of SM cell progeny, the distance between the nuclei of the two inner cells from among the four cells on each lateral half were measured. A similar measurement was made to determine the distance between the two outer nuclei. Gonad length was measured from the vulva to the distal end. All measurements were made using ImageJ software.
|
10.1371/journal.pntd.0007260 | A Plant like Cytochrome P450 Subfamily CYP710C1 Gene in Leishmania donovani Encodes Sterol C-22 Desaturase and its Over-expression Leads to Resistance to Amphotericin B | Leishmania donovani is a protozoan parasite, a primary causative agent of visceral leishmaniasis. Sterol produced via the mevalonate pathway, show differences in composition across biological kingdoms. The specific occurrence of Δ22-unsaturated sterols, containing a double bond at the C-22 position in the side chain occurs in fungi as ergosterol and as stigmasterol in plants. In the present study, we report the identification and functional characterization of a plant-like Cytochrome P450 subfamily CYP710C1 in L. donovani as the Leishmania C-22 desaturase.
In silico analysis predicted the presence of a plant like CYP710C1 gene that encodes a sterol C-22 desaturase, a key enzyme in stigmasterol biosynthesis. The enzymatic function of recombinant CYP710C1 as C-22 desaturase was determined. To further study the physiological role of CYP710C1 in Leishmania, we developed and characterized an overexpressing strain and a gene deletion mutant. C-22 desaturase activity and stigmasterol levels were estimated in the wild-type, overexpressing promastigotes and heterozygous mutants.
We for the first time report the presence of a CYP710C1 gene that encodes a plant like sterol C-22 desaturase leading to stigmasterol biosynthesis in Leishmania. The recombinant CYP710C1 exhibited C-22 desaturase activity by converting β-sitosterol to stigmasterol. Axenic amastigotes showed higher expression of CYP710C1 mRNA, protein and stigmasterol levels compared to the promastigotes. Sterol profiling of CYP710C1 overexpressing L. donovani and heterozygous mutant parasites demonstrated that CYP710C1 was responsible for stigmasterol production. Most importantly, we demonstrate that these CYP710C1 overexpressing promastigotes are resistant to amphotericin B, a drug of choice for use against leishmaniasis. We report that Leishmania sterol biosynthesis pathway has a chimeric organisation with characteristics of both plant and fungal pathways.
| The cytochromes P450 (P450s) are ubiquitous heme-containing enzymes that affect a vast range of oxidation reactions in nature. Cytochrome P450s (CYPs) play an essential role in the metabolism of endogenous or xenobiotic compounds and steroid. The sterol compositions among biological kingdoms differ in the specific occurrence of Δ22-sterols. The C22-desaturation reaction is catalyzed by independent cytochrome P450 family proteins, CYP61 in fungi, and CYP710 in plants. Leishmania donovani is a protozoan parasite that causes visceral leishmaniasis (kala-azar). In silico analysis predicted the presence of a plant like CYP710C1 gene in L. donovani that encodes a plant like sterol C-22 desaturase, a key enzyme in stigmasterol biosynthesis. Here, we have characterized CYP710C1 protein of L. donovani. Sterol profile analysis of wild-type, CYP710C1 overexpressing L. donovani and heterozygous mutant parasite showed that CYP710C1 is responsible for stigmasterol production. Amphotericin B has been used in India for treatment of visceral leishmaniasis for over a decade. Our results demonstrated that overexpression of CYP710C1 gene leads to resistance to amphotericin B in L. donovani. Furthermore, characterization of a plant like CYP710C1 gene in Leishmania indicates the presence of a hybrid pathway that shares a resemblance to both fungal and plant pathways.
| Leishmaniasis is caused by an obligate intracellular protozoan parasite of the genus Leishmania and is spread by the sandfly vector. The pathological features of leishmaniasis range from self-healing cutaneous lesions to fatal visceral leishmaniasis. Visceral leishmaniasis control relies mainly on chemotherapy due to problems related to vector control and the lack of an adequate vaccine to treat the disease [1]. The development of resistance against currently available antileishmanial drugs has resulted in a growing need for discovering novel drug targets and developing new inhibitors [2].
The cytochromes P450 (P450s) are ubiquitous heme-containing enzymes that affect a vast range of oxidation reactions in nature. Cytochrome P450s (CYPs) are present in all three domains of life and play an essential role in the metabolism of endogenous or xenobiotic compounds and steroids [3–6].
Cytochrome P450 database (http://drnelson.uthsc.edu/CytochromeP450.html) represents the CYP genome (CYPomes) of various species [7]. According to the accepted standard convention, a family and subfamily use a numeral and a letter, respectively. CYP51E refers to the family 51 and subfamily E. Similarly, CYP710A is the family name and CYP710A1-A4 belong to its subfamily. Arabidopsis has four genes encoding putative P450 protein belonging to CYP710A subfamily. These are CYP710A1, CYP710A2, CYP710A3 and CYP710A4 [4].
CYP51 is present in fungi and has a role in sterol biosynthesis. Cytochrome P450 CYP61 (sterol C-22 desaturase) is yet another superfamily and is present as CYP710 in plants. CYP710 plant P450s are categorized as putative C-22 desaturase that catalyses the synthesis of stigmasterol and brassicasterol/crinosterol from β-sitosterol and 24-epi-campesterol respectively. Both CYP710A1-A2 and CYP710A4 have been characterized in Arabidopsis thaliana and have been shown to have a role in stigmasterol synthesis [4].
Sterols are isoprenoid-derived lipids that are involved in the maintenance of membrane integrity and as biosynthetic precursors of steroid hormones in eukaryotes. The genes encoding enzymes that catalyse the biosynthesis of ergosterol but not cholesterol are present in both Trypanosoma spp. and in Leishmania spp. [8–10]. This characteristic is more similar to fungi than to other eukaryotes including mammals [11]. The primary component of the Leishmania membrane is ergosterol which is functionally related to the maintenance of structural integrity and protection from biotic stress [12]. The enzymes to synthesise cholesterol are not present in Leishmania, but they do have detectable levels of cholesterol which they probably take up from their external environment [13]. Plants, on the other hand, have a more complex sterol composition. In plants, β-sitosterol, stigmasterol, campesterol, and cholesterol form the majority of sterols and are present in membranes [14, 15].
The sterol biosynthetic pathway has been well-studied in fungi, animals and land plants. Squalene synthesised by mevalonate pathway (MVA-pathway) and methylerythritol-phosphate pathway (MEP-pathway) [16] undergoes step-wise oxygenation and cyclization to form either lanosterol (in vertebrates, fungi and Leishmania) or cycloartenol (in land plants) (Fig 1A) [17]. Cycloartenol contains a cyclopropane ring that is subsequently cleaved by plant-specific enzymes in later steps leading to the formation of common sterols with four-carbon rings. In step-wise reactions involving oxidations, reductions and demethylations, lanosterol is converted into cholesterol (in vertebrates) or ergosterol (in fungi and Leishmania) while cycloartenol, on the other hand, is converted into phytosterols like campesterol, sitosterol and stigmasterol in land plants (Fig 1A).
The specific occurrence of Δ22-sterols is one of the most spectacular differences in the sterol composition among biological kingdoms [4]. The C-22 unsaturated sterols are primarily found in fungi and plants. For example, ergosterol in fungi and stigmasterol in land plants are Δ22-unsaturated sterols that are structurally different from sterols of animal origin (cholesterol) (Fig 1B). Sterol C-22 desaturase plays an essential role as a terminal enzyme in the sterol biosynthesis in fungi and plants [18]. The C-22 desaturation reaction is catalysed by independent cytochrome P450 family proteins, CYP61 in fungi, and CYP710 in plants. In Arabidopsis thaliana, the synthesis of stigmasterol and brassicasterol is catalysed by two separate sterol C-22 desaturases, encoded by the genes CYP710A1 and CYP710A2, respectively [4]. Biosynthesis and physiology of the Δ22-sterols are not clearly understood. Leishmania and Trypanosomes being non-photosynthetic protozoans synthesise not only lanosterol but also have CYP710C (L. major, XP_001684965) related gene [19, 20]. The presence of CYP710C related gene suggests a possible role of this C-22 desaturase in the synthesis of stigmasterol in these parasites (Fig 1C).
The composition of sterols inside fungi and Leishmania is a significant determinant of the action of anti-fungal/anti-leishmanial polyene antibiotic amphotericin B (AmB). AmB binds with its mycosamine appendage to ergosterol [21], a significant sterol in fungi, Leishmania and Trypanosomes. This binding leads to the disruption of membrane integrity coupled with an extensive manipulation of redox balance resulting in the induction of cell death [22]. AmB also causes ergosterol sequestration and induces an oxidative burst, though the exact mechanism of this induction remains mostly unexplored [23]. The AmB can directly interfere with the normal metabolic state of the mitochondria thus influencing the redox state of the cell [24, 25].
We for the first time report the presence of a CYP710C1 gene that encodes a plant like sterol C-22 desaturase, leading to stigmasterol biosynthesis in L. donovani [17]. Characterization of a plant like CYP710C1 protein in Leishmania was performed. L. donovani overexpressing CYP710C1 and heterozygous mutant parasite showed that CYP710C1 is responsible for stigmasterol production. Overexpression of CYP710C1 gene resulted in resistance to AmB in L. donovani.
All restriction enzymes and DNA modifying enzymes were obtained from New England Biolabs (MA, USA). Plasmid pET-30a was obtained from Novagen (Germany). Escherichia coli DH10β and BL21 (DE3) were used as the host for plasmid cloning and protein expression, respectively. Nickel-nitrilotriacetic acid-agarose was purchased from Qiagen (USA). DNA and protein markers were acquired from New England Biolabs, β- NADPH, β-sitosterol and stigmasterol were acquired from Sigma-Aldrich (USA). Other materials used in this study were of analytical grade and were commercially available.
L. donovani Bob (LdBob/strain/MHOM/SD/62/1SCL2D) was obtained from Dr Stephen Beverley (Washington University, St. Louis, MO). Wild-type L. donovani Bob (WT) promastigotes were cultured at 22°C in M199 medium (Sigma-Aldrich, USA) supplemented with 100 mg/ml of penicillin (Sigma-Aldrich, USA), 100 mg/ml of streptomycin (Sigma-Aldrich, USA), and 5% heat-inactivated fetal bovine serum (FBS; Gibco/BRL, Life Technologies, Scotland, UK). WT parasites were routinely cultured in M199 media with no drug supplementation whereas the genetically manipulated heterozygotes (CYP710C1/NEO), in which one allele of the LdCYP710C1 gene has been replaced with neomycin phosphotransferase gene were cultured in 300 μg/ml paromomycin whereas CYP710C1 overexpressing (CYP710C1 OE) parasites were maintained in 5 μg/ml of blasticidin.
The axenic amastigotes were prepared according to the standard protocol [26]. The late-log promastigotes were adapted in an acidic media (RPMI-1640/25 mM 2-(N-morpholino) ethane sulfonic acid (MES)/pH 5.5), at 26°C. These parasites were then grown in RPMI-1640/MES/pH 5.5 at 37°C with 5% CO2.
The gene for CYP710C1 (LDBPK_303610) was amplified by PCR using forward primer with a flanking BamHI site (5’ TTTGGATCCATGGCGAAGAAGAAGAAGAAATTCAAGATGGCT -3’) and reverse primer with a flanking HindIII site (5’—TTTAAGCTTCTACACCTTCTCAGCCTTGGGTT -3’) from L. donovani genomic DNA. The CYP450s are heme containing proteins. Since the standard laboratory strains of E. coli have a limited capacity to take up heme from the extracellular environment, therefore, several strategies, have been developed for the expression of active CYP450s in E. coli [27]. One such strategy is N-terminal modification and co-expression of auxillary protein. In the present study, we have used both N-terminal modification and addition of auxillary protein (Cytochrome P450 reducatse) in the enzyme assay. The highest expression levels were observed from the expression constructs that were modified by eliminating a stretch of hydrophobic residues and replacing the first 41 residues with the fragment MAKKKK. The 1527-bp amplification product encompassing the entire CYP710C1 open reading frame (ORF) was cloned into the pET-30a vector (Novagen) using BamHI and HindIII restriction sites. This construct containing a His6-tag at the N terminus was transformed into the E. coli BL21 (DE3) strain (Novagen). Protein expression was induced with 1 mM isopropyl b-D-thiogalactopyranoside at 28°C for 5 h. Bacterial culture was then pelleted down by centrifugation at 5000 g for 10 min, and the cell pellet was suspended in lysis buffer (20 mM Tris-Cl, pH 8.0, 10 mM imidazole, 500 mM sodium chloride, 0.4% Triton X-100, 0.5 mM EDTA, 1 mM DTT, 10% glycerol, 2 mM phenylmethylsulfonyl fluoride and protease inhibitor mixture). The overexpressed protein was purified using Ni2+ -NTA-agarose resin (Qiagen) by eluting with increasing concentrations of imidazole. The purified protein was found to be 95% pure as judged by SDS-PAGE.
Total RNA from approximately 2 x 106 parasites was isolated using TRIZOL reagent (Sigma, USA). RNA was precipitated by phenol-chloroform treatment and dissolved in DEPC treated RNase free water and quantified with spectrophotometric analysis. cDNA was prepared from 4 μg of total RNA using First Strand cDNA Synthesis Kit (Thermo Scientific, USA) and random hexamer priming. The subsequent cDNA was analysed by Quantitative real-time PCR experiments (Q-PCR) using SYBR Green PCR Master Mix (Applied Biosystems, CA, USA). Fold change was determined using the comparative (2-ΔΔCt) method. In the comparative or ΔΔCt method of qPCR data analysis, the Ct values obtained from two different experimental RNA samples are directly normalized to a reference gene and then compared. The CT value is inversely proportional to the abundance or relative expression level of the gene of interest. In the present study, the reference gene was kinetoplast minicircle DNA specific for L. donovani with primer sequences as follows: forward primer 5’ CCTATTTTACACCAACCCCCAGT 3’ [JW11] and reverse primer 5’ GGGTAGGGGCGTTCTGCGAAA 3’ [JW12].
Late log phase promastigotes (109 parasites) or axenic amastigotes were taken and washed with PBS. The resulting cell pellet was resuspended in 20 ml of dichloromethane: methanol solvent (2:1 v/v), mixed vigorously and incubated for 24 h at 4°C. After centrifugation at 11,000 x g for 1 h at 4°C, the extract was evaporated under vacuum. The extract was saponified with 30% KOH in methanol at 80°C for 2 h. Sterols were then extracted using n-hexane and evaporated. The dried residues were dissolved in dichloromethane. Two volumes of N,O-bis (trimethylsilyl)trifluoroacetamide (BSFTA) was added to an aliquot of the extracted sterol solution, and the tubes were sealed and heated at 80°C for 1 h.
The sterols were subjected to gas chromatography/mass spectrometry (GC/MS) analysis using Shimadzu TD 1020 GC Mass Spectrometer QP2010 Plus that was equipped with DB5 columns (dimethylpolysiloxane/diphenyl ratio, 95/5; dimension 30 m by 0.25 mm). The gas carrier was helium. For analysis, the column was kept at 270°C, the injector and detectors were kept at 300°C. The linear gradient was from 150°C to 180°C at 10°C/min for methyl esters. MS conditions were 280°C; electronic ionization was at 70 eV and an emission current of 2.2 kV. NIST (National Institute for Standards and Technology) library was used for obtaining the Retention time (RT), mass/charge (m/z) values and peak assignment. The Retention time (RT) for a compound is not fixed as several factors can influence it. The Retention time (RT) depends on the matrix effects of analytes. Therefore, the RT can be different between pure standards and cell lysates containing many kinds of sterols. Sterols structures were identified by reference to relative RT and mass spectra (m/z) [4]. In mass spectrometry, several molecules are ionised by the high energy electron beam and form cations thereby leading to different fragmentation pattern. The same compound can, therefore, have different m/z values.
The concentration of the sterols is calculated by comparing the peak area of the analyte in the sample with the peak area of the standard sterol of a known concentration [28].
Purified recombinant CYP710C1 protein (4 mg) was subcutaneously injected in mice using Freund’s complete adjuvant (Sigma, USA), followed by three booster doses of the recombinant protein (2 mg) in Freund’s incomplete adjuvant (Sigma, USA), at 2-week intervals. The mice were sacrificed after the last booster and serum were collected for western blot analysis. Early log phase wild-type promastigotes, axenic amastigotes, CYP710C1 overexpressing strain and heterozygous mutant parasites were harvested, and the resultant cell pellets were re-suspended in lysis buffer (10 mM Tris-Cl, pH 7.2, 5 mM DTT, 10 mM NaCl, 1.5 mM MgCl2, 0.1 mM EDTA, 0.5% Triton X-100, 0.3 mM phenylmethylsulfonyl fluoride (PMSF)). The cell pellets were lysed by freeze-thaw cycles and sonicated on ice followed by centrifugation at 10,000 g.
Total soluble cell extracts of promastigote (40 μg) were fractionated on a 10% SDS-PAGE gel and blotted onto a nitrocellulose membrane using an electrophoretic transfer cell (Bio-Rad). After blocking with 5% BSA, the membrane was incubated for 2 h at room temperature with anti- CYP710C1 antibody (1:3000) generated in mice. In case of recombinant protein, incubation was done with anti- CYP710C1 antibody (1:5000). The membrane was then washed with Tris-buffered saline (TBS) containing 0.1% Tween 20 (TBS-T) and incubated with horse-radish peroxidase (HRP)-conjugated anti-mouse antibody (Cell Signaling Technology number 7076S) (1:5000). The blot was developed using 3, 3 -diaminobenzidine (DAB) tablets (Sigma) or ECL kit (Amersham Biosciences) according to the manufacturer’s protocol.
Desaturase activity of rCYP710C1 was determined according to the coupled enzyme assay procedure [4]. The reaction mixture (0.5 mL) consisted of 50 mM potassium phosphate, pH 7.25, recombinant CYP710C1 (100 μg protein/mL), 100 mM NADPH, and different substrate concentrations ranging from 75 to 150 μM of β-sitosterol. In order to monitor cytochrome P450 activity, 0.1 unit/mL of a purified recombinant NADPH-P450 reductase (Cytochrome P450 Reductase) was added to the reaction mixture. To check in vitro cytochrome P450 enzymatic activity of cell lysates (100 μg) of WT, CYP710C1 OE and CYP710C1/NEO mutants were used along with recombinant NADPH-P450 reductase (0.1 unit/mL). The incubation was done up to 90 min at 30°C. The reactions were stopped by adding 50 μL of 1 N HCl. The reaction products were extracted 4x with an equal volume of ethyl acetate. The ethyl acetate extracts were evaporated to dryness, and then dried residues were dissolved in dichloromethane followed by addition of 2 volumes of BSTFA at 90°C for 1 h. The extracts were completely evaporated in a stream of nitrogen gas and were immediately stored at -80°C. These dried lipid residues were dissolved in hexane and subsequently used for GC-MS analysis.
GC-MS, results are based on retention times of standards and NIST mass spectral library match. The availability of a mass spectral library makes this technique attractive as qualitative analysis can be performed despite the lack of commercially available standards.
The susceptibility profile of the L. donovani wild-type and transgenic promastigotes to the drugs was determined using MTT [3-(4, 5- dimethylthiazol-2-yl) -2, 5- diphenyltetrazolium bromide] (Sigma) assay [29]. Briefly, log-phase promastigotes (5 × 104 cells/well) were seeded in a 96-well flat-bottomed plate (Nunc) and incubated with different drug concentrations at 22°C. After 72 h of incubation, 10 μL of MTT (5 mg/ml) was added to each well, and the plates were incubated at 37°C for 3 h. The reaction was stopped by the addition of 50 μL of 50% isopropanol and 20% SDS followed by gentle shaking at 37°C for 30 min to 1 h. Absorbance was measured at 570 nm in a microplate reader (SpectraMax M2 from Molecular Devices).
All sequences were accessed from TriTrypDB (http://tritrypdb.org/tritrypdb/). For inactivation of the CYP710C1 gene, a targeted gene replacement strategy based on PCR fusion was employed. The flanking regions of the CYP710C1 gene (5’UTR and 3’UTR) were amplified and fused by PCR to the neomycin phosphotransferase gene (NEO). The 5’UTR (770 bp) of the CYP710C1 gene was obtained from WT L. donovani genomic DNA by PCR amplification using primers A and BNeo (Table 1). The NEO gene was amplified from pX63-NEO with primers CNeo and DNeo. The 3 ‘UTR (590 bp) of the CYP710C1 gene was obtained from L. donovani WT genomic DNA by PCR amplification using primers ENeo and reverse primer F (Table 1).
The 5’UTR of L. donovani CYP710C1 gene was then ligated to the antibiotic resistance marker gene (NEO) by PCR using primers A and DNeo. This fragment (5’UTR marker gene) was then fused with the 3’UTR using primers A and F, yielding the linear replacement cassette, 5’UTR’-Neo-3’UTR. The fragment was gel purified, and about 2 μg of the fragment were transfected by electroporation into the wild-type L. donovani promastigotes according to the standard protocol [30]. The transfectants were selected in the presence of 300 μg/ml paromomycin (Sigma-Aldrich, USA). The cells resistant to antibiotic selection were checked by PCR-based analysis for the correct integration of the replacement cassette using primers shown in (Table 2). The genotype of the LdCYP710C1 mutants was confirmed by Southern blotting analysis using a standard protocol [31]. The constructs were sequenced to confirm the correct orientation and sequence fidelity. Our attempts to delete the second copy of the CYP710C1 gene were not successful. Hence, we proceeded with the characterization of the heterozygous mutant strain.
In order to overexpress CYP710C1 gene in L. donovani, 1500-bp CYP710C1 was amplified using the forward primer 5’-TCTAGAATGGACTACAAAGACGATGACGACAAGATGGCAGCGTTTAGTCGTCTCCTCG-3’ with a flanking XbaI site and the reverse primer 5’- AAGCTTCTACACCTTCTCAGCCTTGGGTTC -3’ with a flanking HindIII site. The amplified DNA fragment CYP710C1 was cloned into the XbaI-HindIII site of pSP-α-blast-α- (Leishmania-specific vector) containing a blasticidin acetyltransferase gene as the selection marker. The recombinant vector pSP72α-zeo-α-CYP710C1 was transfected by electroporation into wild-type L. donovani promastigotes according to the standard protocol [30], and selection of transfectants was made in the presence of 5 μg/ml blasticidin (Sigma-Aldrich, USA).
6 week old female Swiss Albino mice were used for the generation of anti- CYP710C1 protein antibody. Animal experiments were performed according to the guidelines approved by the Committee for Control and Supervision of Experiments on Animals (CPCSEA), Ministry of Environment and Forest, Government of India. The protocol was approved by the Institutional Animal Ethics Committee (IAEC) of Jawaharlal Nehru University (JNU) (IAEC Code Number: 15/2017).
Graph Pad Prism Version 5.0 was used for the statistical analysis. Data shown are representative of at least three independent experiments unless otherwise stated as n values given in the legend. All the experiments were set in triplicate, and the results are expressed as the mean ± S.D. Student's t-test was employed to assess the statistical significance of the differences between a pair of data sets with a p-value of < 0.05 considered to be significant.
In our earlier report, we have presented an un-rooted phylogenetic tree comparing CYP710, CYP51 of Leishmania and CYP51 of Trypanosoma with fungi, plant, bacteria and animal counterparts. Leishmania CYP710 and plant CYP710 reside in one branch indicating that they are highly similar and are also found to be similar to CYP61 of fungi indicating their common origin [17]. Multiple sequence alignment of different CYP710 with the Leishmania CYP710C1 is shown in Fig 2.
Two annotated cytochrome P450 proteins from Leishmania with accession number, DQ267494 (CYP5122A1) and UniProt ID: A2TEF2 (X-ray structure available in PDB (http://www.rcsb.org/pdb/) (PDB ID: 3L4D) have been reported in the literature. Our studies identified another cytochrome P450 protein-encoding gene, CYP710C1, in the genome of Leishmania donovani (UniProt ID: E9BMA4), Leishmania infantum (UniProt ID: A4I646), Leishmania major (UniProt ID: Q4Q6T3), and Trypanosoma cruzi CL Brener (UniProt ID: Q4DX81). This gene is present on chromosome 30 in Leishmania while in Trypanosoma cruzi it is present on chromosome 32. The CYP710C1 gene sequence is not present in Trypanosoma brucei.
Amino acid sequence alignment of the CYP710C1 protein of L. donovani with homologous sequences from other strains using Clustal Omega (http://www.ebi.ac.uk/Tools/msa/clustalo/) revealed that CYP710C1 of L. donovani shares 72% identity with Trypanosoma cruzi (CYP710C1), 40% identity with Selaginella moellendorffii (Uniprot ID:D8QPW3) (CYP710A21v1) and 38–39% identity with Arabidopsis thaliana (CYP710A1-4) (Arabidopsis thaliana CYP710A1 (Uniprot ID: O64697), Arabidopsis thaliana CYP710A2 (Uniprot ID: O64698), Arabidopsis thaliana CYP710A3 (Uniprot ID:A0A178VR86), Arabidopsis thaliana CYP710A4 (Uniprot ID: A0A178VVZ9) (Fig 2). The predicated protein sequence of Arabidopsis thaliana CYP710A1 is 81.9%, 77.7%, and 76.1% identical to CYP710A2, CYP710A3 and CYP710A4 respectively.
The CYP family of proteins are reported to have four conserved regions. Motif ‘a’ containing the conserved motif AGXDTT contributes to oxygen binding and activation. The second conserved motif EXLR (motif ‘b’) and the third consensus PER (motif ‘c’) form E–R–R triad that is important for locking the heme pocket into position and assuring stabilization of the core structure. Importantly, motif d containing the conserved sequence FXXGXRXCXG corresponds to the heme-binding domain (Fig 3).
Comparison of the sequence logo of all the four conserved motifs of L. donovani with T. cruzi and plants (Arabidopsis thaliana and Selaginella moellendorffii) showed that the proteins show a high degree of conservation. However, there were a few noticeable differences among the motifs (Fig 3). The motif “a” (oxygen binding) showed complete conservation between L. donovani and T. cruzi but not between L. donovani and S. moellendorffii where in case of plants an alanine is present instead of serine at the second position Motif “b” (EXXR) of L. donovani was fully conserved and showed similarity with T. cruzi and plant. The only noticeable difference observed was in the case of plants where the second position had Val in place of Gln. The motif “c” of L. donovani and T. cruzi were identical. It also showed similarity with plants except in the first and the sixth position where it has Phe in place of Tyr and Met. Motif “d” of L. donovani was well conserved and was identical to T. cruzi. It was also identical to plants except in the third and the ninth position. In the third position, Ala (S. moellendorffii) and Asn (A. thaliana) were present in place of Val. In the ninth position, the only difference was in the case of S. moellendorffii where Leu was present in place of Val. These studies led us to conclude that CYP710C1 from Leishmania and plants have highly similar conserved motifs.
In order to characterize L. donovani CYP710C1, the gene was cloned and expressed as His-tagged fusion protein in E. coli (Fig 4A). The induced His6-tagged recombinant CYP710C1 had an estimated molecular size of ~ 61 kDa. The size of rCYP710C1 correlated with the size of the CYP710C1 protein (~ 55 kDa) and His6 tag (~ 6 kDa). The recombinant protein was purified to homogeneity by Ni2+-NTA metal affinity chromatography (Fig 4A, Lane 4). The expression of the full-length CYP710C1 recombinant protein (~ 61 kDa) was confirmed by immunoblot analysis using anti-CYP710C1 antibody (Fig 4B). The presence of CYP710C1 protein in the promastigotes and the amastigotes was also checked by western blot analysis using anti-CYP710C1 specific antibodies. The anti-CYP710C1 antibody detected a 55 kDa band in the cell extracts of both the promastigotes and the amastigotes (Fig 4C). A higher expression level of the CYP710C1 protein was observed in amastigotes as compared to the promastigotes. A two-fold increase in CYP710C1 mRNA levels was observed in amastigotes in comparison to promastigotes by semi-quantitative RT-PCR (Fig 4E). The CT values of CYP710C1 and reference gene is reported in Supplementary Information S1 Table. The sterol composition of promastigotes and axenic amastigotes cells was also examined (Table 3). Cell pellets of promastigotes and amastigotes were resuspended in dichloromethane: methanol solvent and incubated for 24 h at 4°C. Sterols were then extracted using n-hexane and evaporated. The dried residues were derivatized and used for GC-MS analysis as reported in the methods section. The major sterols found in promastigotes of L. donovani were cholesterol (40%) and ergosterol (48%). The axenic amastigotes of wild-type cells contain ~40% cholesterol, ~12.5% ergosterol, ~10% stigmast-5-en-3-ol and ~13% stigmast-5, 22-diene-3-ol of total sterols. The amount of stigmasterol was higher in axenic amastigotes as compared to the promastigotes.
The desaturase activity of rCYP710C1 was assessed using a coupled enzyme assay as explained in the materials and methods. The reaction was carried out with rCYP710C1 in the presence of NADPH and purified recombinant NADPH dependent cytochrome P450 reductase using different concentrations of β-sitosterol, the potential substrate. The gas chromatography analysis of the reaction products from the rCYP710C1 with β-sitosterol as the substrate showed specific peaks at the same retention time (25.09 min) and m/z (484) as that of stigmasterol (Fig 5A) indicating that the CYP710C1 protein catalyzed the C-22 desaturase reaction to produce stigmasterol from β -sitosterol in vitro. A concentration-dependent increase in the area of the peak of stigmasterol with varying concentration of β-sitosterol was observed. A maximum increase was observed with 125 μM concentration of β-sitosterol (Fig 5B and Table 4). The results are representative of three different replicates.
The CYP710C1 encoding sequence was cloned in Leishmania-specific pSP-α-blast-α vector (Fig 6A) and transfected into L. donovani. Both western blot and RT-PCR confirmed that the expression of CYP710C1 was higher in the overexpressing strain (CYP710C1 OE) as compared to the wild-type promastigotes (Fig 6B and 6D). The western blot showed some non-specific bands other than the CYP710C1 (55 kDa) band possibly due to the polyclonal nature of the antibody. Further, cell lysates of CYP710C1 OE promastigotes showed an increased conversion of β-sitosterol into stigmasterol as indicated by enhanced levels of stigmasterol (Fig 6E and Table 5).
The sterol compositions of wild-type and CYP710C1 OE L. donovani promastigotes is reported in Table 6. The significant sterols found in promastigotes of L. donovani were cholesterol (40%) and ergosterol (48%). In contrast, in CYP710C1 OE promastigotes, the significant sterols were a mixture of cholesterol (~33%), ergosterol (~42%) and stigmasterol (~21%). The possible reason for the peak of cholesterol could be due to the increased uptake from the culture medium as Leishmania promastigotes do not synthesize cholesterol but can take it from the surrounding medium [32]. The percentage of stigmasterol in CYP710C1 OE promastigotes was about ten times higher than the wild-type promastigotes (Table 6). We also examined the sterol composition of axenic amastigotes cells derived from the wild-type and CYP710C1 OE L. donovani promastigotes. The ergosterol contents in axenic amastigotes are low (~12%) compared to wild-type promastigotes. The axenic amastigotes of wild-type cells contain ~42% cholesterol, ~12.5% ergosterol, ~10% stigmast-5-en-3-ol and ~13% stigmast-5, 22-diene-3-ol of total sterols. In contrast, the axenic amastigote of CYP710C1 OE cell lines contained ~67% of stigmasterol and ~17% β-sitosterol-3 acetate. The amount of stigmasterol found in axenic amastigotes was about three times higher than promastigote overexpressing cells and about thirty-four times higher than wild-type promastigotes (Table 7).
The essentiality and function of LdCYP710C1 in the life cycle of the parasite were determined by replacing one allele of this gene with cassettes harbouring drug-resistance marker gene using targeted gene replacement strategy. This was achieved by generation of inactivation cassettes having neomycin phosphotransferase (NEO) as selection markers along with 5’UTR and 3’UTR of the CYP710C1 gene, as described under Materials and Methods (Primers position marked in Fig 7A). Linear replacement cassettes made by PCR-based fusion were transfected into wild-type L. donovani promastigotes, leading to the generation of heterozygous parasites. Replacement of a single copy of the LdCYP710C1 gene was confirmed by PCR-based analysis using primers external to the transfected inactivation cassette of the CYP710C1 gene. DNA from the transfected L. donovani promastigotes was isolated and subjected to PCR-based analysis which showed 1.1 (with primer 1 and primer 6) and ~1.2 kb (with primer 5 and primer 4) bands in the case of the NEO cassette (Fig 7B). Similarly, a band of ~1.25 kb was observed in the case of WT with WT specific primers (1–2, 3–4) given in Table 2. However, no band was observed when PCR reaction was set up with WT genomic DNA and NEO specific primers (5–4, 1–6) confirming that one allele of the WT LdCYP710C1 gene had been replaced in these heterozygous mutant parasites.
The genotype of the heterozygous (CYP710C1/NEO) mutant parasites was further confirmed by Southern blot analysis. Genomic DNA from WT and transgenic parasites were digested with the XhoI and BglII restriction enzymes. In the WT cells, digestion of the LdCYP710C1 gene locus with XhoI yielded a ~1.26-kb band after probing with the 5’UTR of the LdCYP710C1 gene (Fig 7C). Integration of NEO cassette in single transfectants was expected to yield ~4.35 kb 5’UTR hybridizing bands, respectively, in addition to the 1.26 kb wild-type band (Fig 7C). Our attempts to delete the second allele of CYP710C1 were not successful.
Both western blot and RT-PCR data further confirmed CYP710C1 heterozygous mutants had lower expression of CYP710C1 as compared to the wild-type promastigotes (Fig 7D and 7F).
The cell lysates (100 μg) of CYP710C1 heterozygous mutants (CYP710C1/NEO) had lower levels of stigmasterol as compared to the cell lysates of wild-type promastigotes (Table 8) leading us to conclude that deletion of one allele of CYP710C1 led to the decreased conversion of β-sitosterol into stigmasterol. In vitro study on the survival of WT, CYP710C1 OE and CYP710C1/NEO inside murine macrophages J774A.1was performed. Virulence studies were carried out to detect the effects of genetic deficiency of CYP710C1 inside murine macrophages. A murine macrophage cell line was infected with WT, heterozygous mutant parasites CYP710C1/NEO and CYP710C1 OE at a multiplicity of infection (MOI) of 20:1. WT parasites were capable of infecting and sustaining robust infection in murine macrophages, whereas the parasitemia of the heterozygous mutants was reduced by around 40% relative to that of WT parasites both at 24 h and 48 h post-infection (Fig 7G). The infectivity of CYP710C1 OE was equivalent to WT.
We also assessed if reduced expression of CYP710C1 compromised the cellular growth of single knockouts. Growth kinetics was studied. Heterozygotes parasites exhibited delayed growth as compared to wild-type parasites (Fig 7H). Our data suggest that the LdCYP710C1 gene has a significant role in the growth and infectivity of amastigotes.
The sterol compositions of wild-type and heterozygous CYP710C1 mutant (CYP710C1/NEO) L. donovani promastigotes strains are shown in Table 6 and Table 7. Typically, the significant sterols found in promastigotes of L. donovani were cholesterol (40%) and ergosterol (48%). In contrast, in CYP710C1 heterozygous (CYP710C1/NEO) promastigotes, the significant sterols were a mixture of cholesterol (~23%) and ergosterol (~52%). No stigmasterol was observed in these mutant promastigotes. The sterol composition of axenic amastigotes cells derived from the wild-type and CYP710C1/NEO L. donovani promastigotes was also examined. The ergosterol contents in axenic amastigotes were low (~12%) when compared to wild-type promastigotes. The axenic amastigotes of wild-type cells contain ~ 42% cholesterol, ~12.5% ergosterol, ~10% stigmast-5-en-3-ol and ~13% stigmast-5, 22-diene-3-ol of total sterols (Table 7). On the contrary, the axenic amastiogotes of CYP710C1/NEO cell lines contained only ~1.7% of stigmasterol. This data indicated that heterozygous mutant of CYP710C1 had altered sterol levels in L. donovani.
As sterol composition is a significant determinant for AmB action, we finally asked whether overexpression of CYP710C1 alters the resistance of the parasite to this drug. MTT assays showed that CYP710C1 OE was ~5-fold more resistant to AmB as compared to wild-type (Fig 8). However, there was no significant change in the sensitivity profile of CYP710C1 OE promastigotes to potassium antimonyl tartrate trihydrate (PAT) and miltefosine, suggesting that the increase stigmasterol composition is responsible for resistance to AmB (Table 9). We also checked the susceptibility of CYP710C1 heterozygous L. donovani promastigotes to AmB. The heterozygous mutant strain displayed increased sensitivity to AmB. Heterozygous promastigotes (CYP710C1/NEO) and WT promastigotes showed that CYP710C1/NEO was ~1.6 -fold more sensitive to AmB as compared to that of WT (Fig 8).
Leishmaniasis is a significant public health problem in several parts of the tropical and subtropical world. The disease treatment entirely relies on a limited repertoire of anti-leishmanial chemotherapy. However, the associated toxicity and the emergence of drug resistance have necessitated the identification of new drugs and drug targets [1].
In recent years, AmB has gained popularity as the first line treatment for leishmaniasis, particularly in its liposomal formulation. AmBisome has several advantages over non-liposomal formulation such as less toxicity, a lower dose (single dose) and no prolonged hospitalization [33].
We report here the presence of a CYP710C1 P450 family gene that encodes a plant-like sterol C-22 desaturase, leading to stigmasterol biosynthesis in Leishmania. Cytochrome P450 (CYP) is a superfamily of heme-containing monooxygenases that is involved in the metabolism of endogenous or xenobiotic compounds. We reported earlier that different classes of CYPs (CYP51E1, CYP710C1, CYP5123A1 and CYP5122A1) exist in various strains/ species of Leishmania and Trypanosomes [17].
Multiple sequence alignment of CYP710C1 protein of L. donovani with homologous sequences from other strains revealed that it shares 72% with T. cruzi (CYP710C1), 40% with Selaginella moellendorffii (CYP710A21v1) and 38–39% identity with Arabidopsis thaliana (CYP710A1-4). The motifs unique to CYP719C10 are highly conserved amongst the various organisms. Despite the high degree of conservation, we were able to identify few discernible differences between CYP710C of various species; however, further studies would reveal the significance of these differences.
CYP710C1 is an NADPH-dependent C-22 desaturase that catalyses the conversion of β-sitosterol into stigmasterol. The gene present in L. donovani encodes a functional protein capable of catalyzing this reaction. Further, the expression of CYP710C1 is higher in amastigotes as compared to promastigotes, suggesting that stigmasterol may confer some benefit to the parasite when it is resident in the host macrophage.
Here we have shown the sterol composition of WT, CYP710C1 OE and CYP710C1/NEO promastigotes and axenic amastigotes have dramatic differences among sterol levels (Table 6 & Table 7). Overexpression of CYP710C1 results in higher level of stigmasterol in promastigotes as compared to their respective wild-type. The percentage of stigmasterol in overexpressing promastigote was about ten times higher than the wild-type promastigotes. The promastigotes of wild-type cells revealed ~40% cholesterol, ~48% ergosterol, and ~2% stigmast-5-en-3-ol of total sterols. The CYP710C1 OE cell lines contributed 33% cholesterol, ~42% ergosterol, ~21% stigmasterol and ~17% β-sitosterol-3 acetate. In the case of CYP710C1 heterozygous (CYP710C1/NEO) promastigotes, the significant sterols were a mixture of cholesterol (~23%) and ergosterol (~52%). No stigmasterol was observed in these mutant promastigotes. Furthermore, ergosterol contents in axenic amastigotes were low (~12%) when compared to wild-type promastigotes. The axenic amastigotes of wild-type cells contain ~ 42% cholesterol, ~12.5% ergosterol, ~10% stigmast-5-en-3-ol and ~13% stigmast-5, 22-diene-3-ol of total sterols (Table 7). On the other hand, the axenic amastiogotes of CYP710C1/NEO cell lines contained only ~1.7% of stigmasterol.
Sterols are considered as membrane reinforcers as they help to sustain the domain structure of cell membranes. Sterols are critical for the formation of liquid-ordered (lo) membrane states (lipid “rafts”) that are supposed to play an essential role in fundamental biological processes such as signal transduction, cellular sorting, cytoskeleton reorganization and asymmetric growth [34]. They have been proposed as critical molecules to maintain membranes in a state of fluidity adequate for function. A higher level of stigmasterol within plant plasma membrane has been correlated with altered fluidity and permeability characteristics [35, 36]. The sterol composition is different in CYP710C1 OE as compared to WT, possibly resulting in altered membrane integrity and membrane dynamics in CYP710C1 OE. There are reports showing parasites with altered sterols composition respond differently to anti-leishmanial drugs [15, 37].
Amphotericin B is known to mediate changes in the membrane permeability thereby inducing a leakage of important cellular constituents and ultimately lysis and death of the cell. They promote such an effect because of the presence of sterols in the cell membrane. The polyene antibiotics and sterols are reported to form molecular complexes to create channels or solid patches that disrupt the membrane [38]. The actual mechanism of action of AmB may be more complex and multifaceted.
The basis of the antifungal and antileishmanial selective toxicity of AmB is proposed to be due to the prevalence of ergosterol on the surface membranes of these organisms and a significant interaction between AmB and ergosterols. AmB binding to ergosterol leads to disruption of the osmotic integrity of the membrane in target cells and is a potent agent used for the treatment of leishmaniasis [21]. Some standard features have been detected in AmB resistant Leishmania such as changes in sterol metabolism, where ergosterol, the primary sterol of wild-type Leishmania cell membranes is reduced or lost and replaced by different cholestane-type sterols [33]. However, in mammalian cells the principal sterol is cholesterol.
Interestingly, overexpression of CYP710C1 led to ~5-fold increase in resistance to AmB in contrast to CYP710C1/NEO which showed ~1.6 -fold more sensitivity to AmB as compared to the wild-type (WT), suggesting that this protein is possibly involved in modulation of target of AmB in Leishmania, and therefore, is possibly essential for the parasite to withstand the drug. The precise mechanism by which stigmasterol regulates membrane properties in Leishmania is not yet understood. Further studies are required to check the exact role of stigmasterol in AmB resistance. The presence of stigmasterol, ergosterol and their corresponding isoforms in L. donovani indicates that Leishmania has acquired both the fungus and plant pathways for sterol biosynthesis. The present study will offer the ability to monitor the emergence and spread of resistance to AmB drug in the field.
|
10.1371/journal.pgen.1001127 | Identification of New Genetic Risk Variants for Type 2 Diabetes | Although more than 20 genetic susceptibility loci have been reported for type 2 diabetes (T2D), most reported variants have small to moderate effects and account for only a small proportion of the heritability of T2D, suggesting that the majority of inter-person genetic variation in this disease remains to be determined. We conducted a multistage, genome-wide association study (GWAS) within the Asian Consortium of Diabetes to search for T2D susceptibility markers. From 590,887 SNPs genotyped in 1,019 T2D cases and 1,710 controls selected from Chinese women in Shanghai, we selected the top 2,100 SNPs that were not in linkage disequilibrium (r2<0.2) with known T2D loci for in silico replication in three T2D GWAS conducted among European Americans, Koreans, and Singapore Chinese. The 5 most promising SNPs were genotyped in an independent set of 1,645 cases and 1,649 controls from Shanghai, and 4 of them were further genotyped in 1,487 cases and 3,316 controls from 2 additional Chinese studies. Consistent associations across all studies were found for rs1359790 (13q31.1), rs10906115 (10p13), and rs1436955 (15q22.2) with P-values (per allele OR, 95%CI) of 6.49×10−9 (1.15, 1.10–1.20), 1.45×10−8 (1.13, 1.08–1.18), and 7.14×10−7 (1.13, 1.08–1.19), respectively, in combined analyses of 9,794 cases and 14,615 controls. Our study provides strong evidence for a novel T2D susceptibility locus at 13q31.1 and the presence of new independent risk variants near regions (10p13 and 15q22.2) reported by previous GWAS.
| Type 2 diabetes, a complex disease affecting more than a billion people worldwide, is believed to be caused by both environmental and genetic factors. Although some studies have shown that certain genes may make some people more susceptible to type 2 diabetes than others, the genes reported to date have only a small effect and account for a small proportion of type 2 diabetes cases. Furthermore, few of these studies have been conducted in Asian populations, although Asians are known to be more susceptible to insulin resistance than people living in Western countries, and incidence of type 2 diabetes has been increasing alarmingly in Asian countries. We conducted a multi-stage study involving 9,794 type 2 diabetes cases and 14,615 controls, predominantly Asians, to discover genes related to susceptibility to type 2 diabetes. We identified 3 genetic regions that are related to increased risk of type 2 diabetes.
| Type 2 diabetes (T2D) is a common complex disease that affects over a billion people worldwide [1]. Through genome-wide association studies (GWAS), at least 24 genetic susceptibility loci have been reported for T2D [1]–[9], including a SNP, rs7593730, at 2q24 near the RBMS1 and ITGB6 genes that was associated with diabetes risk in a recent report from the Nurses' Health Study/Health Professionals Follow-up Study (NHS/HPFS) [2]. However, most of the reported genetic variants have small to moderate effects and account for only a small proportion of the heritability of T2D, suggesting that the majority of inter-person genetic variation in this disease remains to be determined. Over the last two decades, China, like many other Asian countries, has experienced a dramatic increase in T2D incidence. Cumulative evidence suggests that Asians may be more susceptible to insulin resistance compared with populations of European ancestry [10]. However, among the previously reported T2D genetic markers, only three SNPs – including two reported very recently – have been identified in populations of Asian ancestry [8], [9]. SNP rs2283228 in the KCNQ1 gene was identified in a 3-stage study that included 194 diabetes patients and 1,558 controls and 268,068 SNPs in the first (discovery) stage [8]. A study conducted among Han Chinese in Taiwan recently identified two additional novel loci in the protein tyrosine phosphatase receptor type D (PTPRD; P = 8.54×10−10) and serine racemase (SRR; P = 3.06×10−9) genes [9].
Large genetic studies conducted in Asian populations will facilitate the identification of additional genetic markers for T2D, particularly for markers with a higher frequency in Asians than in other populations. We recently completed a GWAS of T2D in Shanghai. We report here our first effort, using a fast-track, multiple-stage study approach, to identify novel genetic markers for diabetes.
The study protocol was approved by the institutional review boards at Vanderbilt University Medical Center and at each of the collaborating institutes. Informed consent was obtained from all participants.
This study consisted of a discovery stage and two validation stages, i.e. an in silico and a de novo validation study. The overall study design is presented in Figure S1.
The discovery stage included 1,019 T2D cases, 886 incident T2D cases from the Shanghai Women's Health Study (SWHS), an ongoing, population-based, prospective cohort study of women living in Shanghai, and 133 prevalent T2D cases identified among controls of the Shanghai Breast Cancer Study (SBCS), who were recruited in Shanghai during approximately the same period as the SWHS [11]. Controls for the discovery phase were 1,710 non-diabetic female controls from the SBCS (for further details, see Text S1, online). The biologic samples used for genotyping in this study were collected by the SWHS and SBCS.
DNA samples were genotyped using the Affymetrix Genome-Wide Human SNP Array 6.0. Extensive quality control (QC) procedures were implemented in the study. In the SWHS/SBCS GWAS scan, three positive QC samples purchased from Coriell Cell Repositories and a negative QC sample were included in each of the 96-well plates of the Affymetrix SNP Array 6.0. SNP data obtained from positive quality control samples showed a very high concordance rate of called genotypes based on 79,764,872 comparisons (mean, 99.87%; median, 100%). Samples with genotyping call rates less than 95% were excluded. The sex of all study samples was confirmed to be female. The identity-by-descent analysis based on identity by state was performed to detect first-degree cryptic relationships using PLINK version 1.06 [12]. We excluded from the study 21 samples that had: 1) call rate <95% (n = 5); 2) samples that were contaminated or had mixed-up labels or that had been duplicated (n = 12); 3) first-degree relatives, such as parent-offspring or full siblings (n = 4).
We also excluded from the analysis SNPs that met any of following criteria: 1) MAF <0.05; 2) call rate <95%; 3) P for Hardy-Weinburg equilibrium HWE <0.00001 in either the case or control groups or in the combined data set; 4) concordance rate <95% among the duplicated QC samples; 5) significant difference in allele frequency distribution (P<0.00001) between the 886 T2D cases from the SWHS and the 133 T2D cases from the SBCS; 6) significant difference in missing rates between cases and controls (P<0.00001). After applying the QC filter, 590,887 SNPs remained for the analyses.
Because of financial constraints, we conducted a fast-track validation study using an approach that combined in silico and de novo replication. We selected a total of 2,100 SNPs from the discovery phase that had P-values of 1.3×10−9 to 5.0×10−3 derived from the additive model and that were not in linkage disequilibrium (LD; r2<0.2 based on the HapMap CHB dataset) with any previously reported T2D GWAS SNPs for an in silico replication using the GWAS scan data from the NHS/HPFS [2]. We used the NHS/HPFS T2D GWAS scans for our first step of validation, because the Shanghai T2D GWAS was conducted concurrently and used the same genotyping platform as the NHS/HPFS T2D GWAS and a priori arrangement was made for the two studies to exchange the top 2,000 SNPs for in silico replication. The NHS/HPFS T2D GWAS included 2,591 cases and 3,052 controls of European ancestry. We recognize that this approach may have reduced our chances of finding ethnicity-specific T2D markers, however, this approach had the advantage of enhancing our ability of finding true genetic markers. From the first in silico replication, 65 SNPs with the same direction of association in both studies and with a MAF >20% were chosen for a second in silico replication using GWAS scan data from a Korean T2D study, which included 1,042 cases and 2,943 controls genotyped with the Affymetrix Genome-Wide Human SNP Array 5.0 platform. In order to improve yield, only the top SNPs that are included in Affymetrix 5.0 (N = 56) or that are in high LD (r2>0.8) with at least one SNP on Affymetrix 5.0 (N = 9) were selected for replication (Table S1). Of the 65 SNPs, the top 8 SNPs replicated in the Korean T2D study were further investigated using GWAS data from a T2D study conducted among Singapore Chinese (2,010 cases and 1,945 controls) who were genotyped by using Illumina HumanHap 610 or Illumina Human1M (Table S2). Four of the 8 SNPs were not directly genotyped in the Singapore study, so instead, we selected SNPs that are in strong LD with these 4 SNPs (imputed SNP information became available recently and is presented in this report). Finally, the 5 top SNPs (rs2815429, rs10906115, rs1359790, rs10751301, and rs1436955) were selected for de novo genotyping in an independent sample set of 1,645 T2D cases and 1,649 controls identified from the SWHS and Shanghai Men's Health Study (SMHS). Four of these SNPs (rs10906115, rs1359790, rs10751301, and rs1436955) were selected for the final stage of de novo genotyping replication in two independent Chinese studies, the Wuhan Diabetes Study (WDS; 1,063 cases and 1,408 controls) and the Nutrition and Health of Aging Population in China (NHAPC) study (424 cases and 1,908 controls). Detailed descriptions of the study designs and populations for each of the participating studies are presented in Text S1 online.
Genotyping for the 5 SNPs included in the SWHS and SMHS sample set was completed using the iPLEX Sequenom MassArray platform. Included in each 96-well plate as quality control samples were two negative controls, two blinded duplicates, and two samples included in the HapMap project. We also included 65 subjects who had been genotyped by the Affymetrix SNP Array 6.0 in the Sequenom genotyping. The consistency rate was 100% for all SNPs for the blinded duplicates, compared with the HapMap data and compared with data from the Affymetrix SNP Array 6.0. Genotyping for the final 4 SNPs in the WDS and NHAPC was completed using TaqMan assays at the two local institute laboratories using reagents provided by the Vanderbilt Molecular Epidemiology Laboratory. Both laboratories were asked to genotype a trial plate provided by the Vanderbilt Molecular Epidemiology Laboratory that contained DNA from 70 Chinese samples before the main study genotyping was conducted. The consistency rates for these trial samples were 100% compared with genotypes previously determined at Vanderbilt for all four SNPs in both local laboratories. In addition, replicate samples comparing 3–7% of all study samples were dispersed among genotyping plates for both studies.
The imputation of un-genotyped SNPs in all participating GWASs was carried out after the completion of the current study using the programs MACH (http://www.sph.umich.edu/csg/abecasis/MACH/) or IMPUTE (https://mathgen.stats.ox.ac.uk/impute) with HapMap Asian data as the reference for Asians and CEU data as the reference for European-ancestry samples. Only data with high imputation quality (RSQR >0.3 for MACH) were included in the current analysis.
PLINK version 1.06 was used to analyze genome-wide data obtained in the SBCS/SWHS GWAS scan. Population structure was evaluated by principal component analysis using EIGENSTRAT (http://genepath.med.harvard.edu/~reich/Software.htm).A set of 12,533 SNPs with a MAF ≥10% in Chinese samples and a distance of ≥25 kb between two adjacent SNPs was selected to evaluate the population structure. The first two principal components were included in the logistic regression models for adjustment of population structures. The inflation factor λ was estimated to be 1.03, suggesting that population substructure, if present, should not have any appreciable effect on the results.
Pooled and meta-analyses were carried out in SAS to derive combined odds ratios (OR) by using data from studies of all stages. We applied the weighted z-statistics method, where weights are proportional to the square root of the number of subjects in each study. Results from both random and fixed effect models are presented.
ORs and 95% confidence intervals (CI) were estimated using logistic regression models with adjustment for age, BMI, population structure (for GWAS data), and gender, when appropriate. Analyses with additional adjustment for smoking were conducted by pooled analysis whenever possible and by meta-analysis when KARE data were included in order to examine the confounding and modification effects of these factors (Table S2). Genotype distributions for the top 4 SNPs included in the final de novo genotyping were consistent with HWE (P> 0.05) in each study. All P values presented are based on two-tailed tests, except where indicated otherwise.
The general characteristics of the participating study populations are presented in Table 1. T2D cases had a higher BMI than controls across all studies. Except for the SWHS, SMHS, and Shanghai Nutrition Institute (SNI) validation studies, where cases and controls were matched on age, cases were older than controls in all other studies. A difference in gender distribution was also seen in several studies. These variables were adjusted for in subsequent analyses.
Table 2 presents the results of analyses of associations of T2D with previously reported, GWAS-identified genetic markers in our discovery samples [1]–[9]. Of the 24 SNPs reported by previous GWAS, 15 were directly genotyped by the Affymetrix SNP Array 6.0. One SNP (rs7578597) showed a MAF = 0 in HapMap CHB data and was not included on the Affymetrix 6.0 chip. The remaining 8 SNPs, including rs2943641, rs10010131, rs13266634, rs12779790, and rs4430796, as well as the newly identified markers rs391300 and rs17584499, were imputed. SNP rs4430796 showed low imputation quality (RSQR = 0.06) in the SBCS/SWHS GWAS and was excluded from the analysis.
We found that 8 of these SNPs showed an association consistent with initial reports at P<0.05, including rs4402960 (3q27.2, IGF2BP2), rs10946398 (6p22.3, CDKAL1), rs13266634 (8q24.11, SLC30A8), rs10811661 (9p21.3, CDKN2A/B), rs5015480 (10q23.33, HHEX), rs7901695 (10q25.2, TCF7L2), rs2283228 (11p15.5, KCNQ1), and rs5215 (11p15.1, KCNJ11). Among the remaining 11 SNPs, 4 SNPs had a MAF of 3–7% in our study population. Thus, our study did not have sufficient statistical power (statistical power range: 19–45%) to replicate these markers (Table 2). Associations of T2D with SNPs that are in LD with the reported T2D SNPs discovered in European-ancestry populations or in Asians are presented in Table S3.
Multidimensional scaling analyses of the GWAS scan data showed no evidence of apparent genetic admixture in our study population (Figure S2). The observed number of SNPs with a small P value was larger than expected by chance (Figure S3). We found that rs10906115 (10p13), rs1359790 (13q31.1), and rs1436955 (15q22.2) were consistently associated with T2D across all studies, although the 95% CI for the per allele ORs in several studies included 1.0 (Table 3; Figure 1). P-values for trend tests (per allele OR, 95% CI) from meta-analyses of data from all studies were highly statistically significant for these associations: 1.45×10−8 for rs10906115 (1.13, 1.08–1.18), 6.49×10−9 for rs1359790 (1.15, 1.10–1.20), and 7.14×10−7 for rs1436955 (1.13, 1.08–1.19). These P-values were below (for rs1359790 and rs10906115) or near (for rs1436955) the genome-wide significance level of 5.0×10−8. SNP rs10751301 (11q14.1) was not replicated in the Singapore or de novo genotyping studies; the P-value for the meta-analysis was 1.31×10−4 in the fixed effect model and 0.004 in the random effect model. Additional adjustment for smoking history did not appreciably change the point estimates described above, although the P-values were slightly elevated (Table S2).
In an exploratory analysis stratified by smoking, BMI, family history of T2D, and age at diagnosis, SNP rs1359790 showed a slightly stronger association with T2D risk among non-smokers (per allele OR = 1.19, 95% CI = 1.12–1.26, P = 6.4×10−8) than among smokers (OR = 1.09, 95% CI = 1.00–1.19, P = 0.044) with a P value of 0.11 for interaction (Table S4). None of the SNPs were related to age at onset of T2D. Neither family history of T2D nor BMI altered the SNP-T2D associations under study.
Using the GWAS data from our discovery stage samples, we were able to validate 8 of 22 previously reported, GWAS-identified T2D SNPs, lending strong support to the validity of the initial discovery samples and methodologies. Applying a fast-track validation study approach, we also identified three promising new T2D markers.
The most significant association identified by our study was for rs1359790 (13q13.1), a novel genetic susceptibility locus identified for T2D (Figure 2). Several transcription factors, such as NIT-2, CdxA, GATA-2, and CDP, bind to this polymorphic site. The C to T transition eliminates a GATA-2 binding site and creates a TATA binding site. The closest known gene, sprouty homolog 2 (Drosophila) (SPRY2), is located 193 kb upstream of rs1359790. The SPRY2 gene encodes a protein belonging to the sprouty family and inhibits growth factor-mediated, receptor tyrosine kinase-induced, mitogen-activated protein kinase signaling [13]. The encoded protein contains a carboxyl-terminal cysteine-rich domain essential for the inhibitory activity of receptor tyrosine kinase signaling proteins and is required for growth factor-stimulated translocation of the protein to membrane ruffles [13], [14]. SPRY2 also modulates the apoptotic actions induced by the pro-inflammatory cytokine, tumor necrosis factor-alpha [15]. SPRY4, a homolog of SPRY2, inhibits the insulin receptor-transduced MAPK signaling pathway [16] and regulates development of the pancreas [17].
SNP rs10906115 is located on chromosome 10p13 (Figure 2), 13.0 kb from rs12779790, which was reported by a previous GWAS of T2D [1]. These two SNPs, however, are in low LD in both Chinese (r2 = 0.06) and European populations (r2 = 0.19) based on HapMap data. SNP rs12779790 was not included in the Affymetrix SNP Array 6.0, Illumina HumanHap 610-Quad, or Human1M-Duo; thus, it was imputed for both the SBCS/SWHS and the NHS/HPFS by using MACH with RSQR>0.9 and for the Singapore studies using IMPUTE with PROPER_INFO >0.85. The imputed SNP rs12779790 was associated with a per allele OR of 1.10 (95% CI = 1.01–1.19, P = 0.035) in the analysis of pooled data from three studies. However, when both rs12779790 and rs10906115 were included in the same logistic model, the association with rs10906115 remained statistically significant (per allele OR = 1.09, 95% CI = 1.02–1.16, P = 0.007), while the association with rs12779790 was no longer statistically significant (per allele OR = 1.04 [95% CI = 0.96–1.12], P = 0.38; Table 4). These data provide strong evidence that rs10906115 is a new genetic variant at 10p13 independent of the previously-identified SNP rs12779790.
SNP rs10906115 is located 22.4 kb downstream of the cell division-cycle 123 homolog (S. cerevisiae) (CDC123) gene and 76.6 kb upstream of the calcium/calmodulin-dependent protein kinase ID (CAMK1D) gene (Figure 2). The CDC123 gene encodes a protein involved in cell cycle regulation and nutritional control of gene transcription [18]. The CAMK1D gene encodes a member of the Ca2+/calmodulin-dependent protein kinase 1 subfamily of serine/threonine kinases. The encoded protein may be involved in the regulation of granulocyte function through the chemokine signal transduction pathway [19]. The role of the CDC123 and CAMK1D genes in the etiology of T2D is unclear.
SNP rs1436955, located on chromosome 15q22.2 (Figure 2), is 51.4 kb downstream of a C2 calcium-dependent domain containing the 4B gene (C2CD4B; also known as NLF2 or FAM148B). C2CD4B is up-regulated by pro-inflammatory cytokines and may play a role in regulating genes that control cellular architecture [20]. The role of inflammation in the pathophsyiology of T2D has been suggested previously [21]–[25]. C2CD4B and SPRY2 are both highly expressed in human pancreatic tissue [26]. Intriguingly, a very recent report from the Meta-Analysis of Glucose and Insulin-related traits Consortium (MAGIC) found that a SNP (rs11071657) near the C2CD4B gene was associated with fasting glucose (P = 3.6×10−8) and T2D (P = 2.9×10−3) [27]. SNPs rs11071657 and rs1436955, however, are not in LD (r2 = 0.04) in Asians, although they are weakly related (r2 = 0.25) in Europeans, according to HapMap data. SNP rs11071657 is not included in the Affymetrix SNP 6.0 array. Imputed data from the SBCS/SWHS GWAS showed that this SNP was not significantly associated with T2D risk (per A allele OR = 1.06, 95% CI = 0.94–1.19), although the direction of the association was consistent with that reported by the MAGIC consortium [27]. Adjusting for rs11071657 did not alter the association of T2D risk with rs1436955 (per allele OR = 1.21, 95% CI = 1.06–1.39, P = 0.006). Again, these data strongly imply that rs1436955 may be a new genetic risk variant for T2D at 15q22.2 independent of the recently reported SNP rs11071657.
In summary, in this first GWAS of T2D conducted in a Chinese population, we identified a novel genetic susceptibility locus for T2D, rs1359790, at 13q31.1. Furthermore, we revealed two new genetic variants (rs10906115 at 10p13 and rs1436955 at15q22.2) near T2D susceptibility loci previously reported by GWAS of T2D conducted in European-ancestry populations. Our study demonstrates the value of conducting GWAS in non-European populations for the identification of novel genetic susceptibility markers for T2D.
|
10.1371/journal.pcbi.1002057 | Attracting Dynamics of Frontal Cortex Ensembles during Memory-Guided Decision-Making | A common theoretical view is that attractor-like properties of neuronal dynamics underlie cognitive processing. However, although often proposed theoretically, direct experimental support for the convergence of neural activity to stable population patterns as a signature of attracting states has been sparse so far, especially in higher cortical areas. Combining state space reconstruction theorems and statistical learning techniques, we were able to resolve details of anterior cingulate cortex (ACC) multiple single-unit activity (MSUA) ensemble dynamics during a higher cognitive task which were not accessible previously. The approach worked by constructing high-dimensional state spaces from delays of the original single-unit firing rate variables and the interactions among them, which were then statistically analyzed using kernel methods. We observed cognitive-epoch-specific neural ensemble states in ACC which were stable across many trials (in the sense of being predictive) and depended on behavioral performance. More interestingly, attracting properties of these cognitively defined ensemble states became apparent in high-dimensional expansions of the MSUA spaces due to a proper unfolding of the neural activity flow, with properties common across different animals. These results therefore suggest that ACC networks may process different subcomponents of higher cognitive tasks by transiting among different attracting states.
| For understanding how neural processes give rise to cognitive operations, it is essential to understand how aspects of the underlying neural network dynamics reconstructed from neurophysiological measurements relate to behavior. For instance, different actions may be represented by neural states characterized by stable population patterns to which activity converges in time, called attractors in the language of dynamical systems. However, experimental demonstrations of neural attractors associated with cognitive entities have been rare so far. One problem may have been that in behaving animals, in-vivo one can access only a relatively small fraction of the total number of neural units comprising the whole system, even with modern multiple single-unit (MSU) recording techniques. Therefore, the neural activity dynamics are necessarily projected from a very high-dimensional into the empirically accessible much lower-dimensional space in which attractor properties may be lost due to ambiguities and entanglement in the flow of trajectories. In the present study, principles from nonlinear time series analysis and statistical learning are applied to MSU recordings from the rat's prefrontal cortex during decision-making tasks. By expanding the empirically accessed neural state space (semi-) attracting properties of neural states corresponding to cognitively defined task-epochs became apparent, in line with many neuro-computational theories.
| To fully understand how neural processes give rise to cognitive operations, it is essential to reconstruct the underlying neural network dynamics from electrophysiological or neuroimaging measurements in relation to behavior. A common theoretical idea is that these dynamical properties of the nervous system, like the convergence of activity to specific stable population patterns (attractors), are what ultimately implement the computational operations that link inputs to outputs [1]–[6]. For instance, different attracting states may represent different active memories or cognitive entities, and movement between these states may correspond to the recall of a memory sequence or the execution of a behavioral or motor plan. Attractor states as a basis for cognition received particular attention in the context of working memory [2], [4], [7]–[9] and decision making tasks [5], [10]–[12].
Especially in recent years, along with the advances in multiple single-unit recording techniques [13], there has been a dramatic rise in the attempts to reconstruct cognitively relevant aspects of the population dynamics. Many of these relied on methods from multivariate statistics and machine learning (as reviewed in [14], [15]). These studies gave a number of valuable insights into mechanisms of neural information processing like the information content of the transient dynamics connecting steady states [16], [17], the representation or processing of stimuli by reproducible sequences of states [18], or the sudden nature of transitions among representational states during learning [19]. Several experimental studies also suggested that spatial representations in the rodent hippocampus [6], [20]–[22] or olfactory representations in zebrafish [16], [23] may have attractor-like properties with sometimes stochastic transitions among them [24], [25]. In these studies, attractor states were indicated by discrete switches in the population activity patterns eventually attained (after some transient) when stimulus parameters were continuously varied. Strictly speaking, however, these studies did not attempt to explicitly demonstrate a convergent flow of neural trajectories (as sometimes pointed out by the authors themselves, [23]), as another important signature of attracting states. Moreover, they mostly focused on (stimulus-driven) sensory or spatial representations rather than on presumably intrinsically-driven higher cognitive processes. In addition, since most of these previous approaches worked directly in the space of observed variables, i.e. the recorded units' firing rates or spike times, they could potentially miss some important structural details of neural space organization, especially in high-noise situations, as they try to infer the dynamics of a large complex system by selecting only a few of its dimensions (recorded neurons). Thus, experimental evidence for the hypothesis that higher cognitive processes proceed by moving among attracting states is still sparse.
Here we combined and adapted two approaches well established in statistical learning theory [26], [27] and nonlinear time series analysis [28], [29] in an attempt to move beyond some of the limitations that could arise in previous analyses of electrophysiological data. These methods were applied to multiple single-unit recordings from the rat anterior cingulate cortex (ACC) during a complex memory-guided decision making task in a radial arm maze (Figure S1). The ACC is assumed to play a key role in higher-level cognitive processes like monitoring of behavior [30], processing error feedback [31], making choices [32] and dissecting task structure [33]. Thus, the ACC is a brain area with complex intrinsic dynamics and computational properties that presumably demand a sophisticated multivariate analysis to much larger degree than comparatively simpler early sensory systems (e.g. [16], [23]). The present analysis was designed to be more sensitive to potential state space structure, suggesting previously unrecognized convergence properties of ACC neural ensemble states associated with cognitive processing steps and stable across multiple trials.
A state space is a coordinate map spanned by all relevant dynamical variables of a system (e.g. the membrane voltages or firing rates of neurons). A single (vector) point in this space represents the whole state of the recorded neural system at a given point in time (e.g. the current firing rates of all neurons), while a trajectory in this space charts how its state changes over time. Most computational theories of the brain work by linking geometrical objects in these spaces (e.g. attractors) and the temporal evolution of neural activity (the trajectories) to specific computational and cognitive functions (e.g. [2], [4], [34]–[37]). However, inferring the dynamics of a large complex system from experimental data by selecting only the observable dimensions (recorded neurons) can lead to incorrect conclusions [28], [29]: Neural trajectories may not be sufficiently “unfolded”, i.e. may follow apparently convoluted patterns where they frequently “intersect” themselves and exhibit ambiguities with regards to their direction of flow (Figure 1A, left). This is due to the fact that other dimensions along which the flow would have been disambiguated are missing (e.g. the third axis in Figure 1A, top left; [28], [38]). Thus, a state space construed solely from the activities of the simultaneously recorded units (termed multiple single-unit activity, MSUA, space in the following) is not guaranteed to properly represent the geometry of the underlying dynamical system's attractors.
A potential solution to this problem was provided by time series embedding theorems [28], [38] which demonstrated that the structure of the underlying attractor dynamics could be fully recovered (under ideal, noise-free conditions) if the dimensionality of the space is expanded by adding a sufficient number p of time-lagged versions ν(t-τi) of the present observations ν(t) as new variables to the space, where the time lags τi are determined such that these new variables do not contain redundant information with respect to the original MSUA axes, i.e., are only weakly correlated with them (Figure 1A, center). In principle, the optimum number of delay axes is constrained by the dimensionality of the underlying attractor of the system [28]. Unfortunately, however, due to the sparseness of the MSUA spaces and the noise levels in these data it cannot be reliably computed. Moreover, given that for neural systems the true dimensionality could be (much) higher than the number of dimensions one has experimentally access to, the number of time lags required for a statistically optimal disambiguation of trajectory flows may be so high that it cannot be accommodated by the (experimentally) limited length of the time series (Materials and Methods).
Therefore, it may be necessary to consider also other types of state space expansion that allow to effectively discern the neural dynamics associated with different cognitive events. Adding interactions between units' firing rates as dimensions to the space seems a particularly suitable choice since neuronal cross-correlations have often been postulated to play an important role in cognitive processes (e.g. [39]–[43]). From a mathematical point of view products of neural firing rates would correspond to terms of a multinomial basis expansion frequently employed in statistical classification procedures [27]. Hence, such an expansion would have both a neuroscientific meaning and a theoretical foundation. Therefore, in our approach the delay-coordinate (DC) map of the MSUA space (DC-MSUA space) is further expanded by adding pairwise and higher order cross-products of the recorded units' firing rates, up to some order O, as new dimensions. For example, an expanded state space of 3rd order will contain all the original MSUA axes, plus time-delayed versions of the firing rates of all n recorded units, ν1(t-τ1), ν2(t-τ2),…, νn(t-τn) as well as new axes corresponding to third order products like ν1(t-τ1) ν2(t-τ2) ν3(t-τ3) or ν1(t-τ1)2 ν3(t-τ3). Vectors in these high-dimensional spaces will be denoted by Φ(t) – each such vector corresponds to a specific (spatio-temporal) pattern of neural firing rates and firing rate correlations up to the order set by the expansion. Since the dimensionality of such spaces can be extremely high, specialized algorithms (so-called kernel-methods [44]–[46]) were used for the statistical analyses, as discussed below. As illustrated in Figure 1A (right), adding these cross-product terms can help to further disentangle neural trajectories by amplifying small differences present in the DC-MSUA space.
Why this 2-stage process in expanding the original MSUA space? If trajectories in the originally recorded MSUA space are already nicely disentangled and noise levels are very low, no further expansion may be necessary. However, many of the simultaneously recorded neurons may fire very sparsely, or may otherwise be non-informative about the system's dynamics, or there may simply not be enough of them which access “sufficiently different aspects” of the system's dynamics. Adding delay coordinates (with delays chosen such as to minimize cross-correlations among the firing-rates of different neurons, see Materials and Methods) will increase the amount of information about the neural dynamics captured by the space by removing ambiguities in the neural flow which may occur in the MSUA space (Figure 1A, center). Adding product terms, on the other hand, may not add further information about the dynamics to the space (although it may make information contained in neuronal correlations explicitly accessible), but it will help to pull trajectories apart and thus enhance task-related differences in the activity flow in situations of high noise (Figure 1A, right; see also Materials and Methods). It may also take care of the fact that putative attractor geometries may be highly nonlinear structures that are not easily captured by linearly separating hyperplanes. Hence, by combining these two types of expansion we arrive at a space which should be both, more informative due to the addition of delay coordinates, and at the same time “less noisy” and more apt for detecting nonlinear structures. Here we show that the identification of ensemble dynamics for different animals and behavioral performance levels will, in general, indeed significantly improve by combining both types of expansion.
As an example, Figure 1B shows a single trial of an animal performing a higher cognitive task explained in the next section. A type of principal component analysis (PCA) suitable for very high-dimensional Oth order spaces, termed kernel-PCA [47] (for O = 1 equivalent to conventional PCA), was used to visualize the neural dynamics in the 3 most variance-explaining dimensions. While for both the MSUA and O = 5 spaces the two illustrated task phases (blue and red dots in Figure 1B) can be clearly discerned, the actual trajectories (the lines connecting the dots) are quite entangled in the MSUA space but are nicely unfolded for high-order expansion spaces, exposing attracting orbits and properties of the two task phases (Figure 1B; see also Video S1).
The techniques introduced above were used to analyze MSU recordings obtained from the rat ACC (Figure S1) while the animals were situated in a radial-arm-maze decision-making task with temporal delay (Figure 2A). This task is considered to be ecologically valid in the sense that it mimics key aspects of rats' natural foraging, food-hording, and retrieval behavior (e.g. [48], [49], [33]). The entire time on task was divided into six epochs with differing cognitive demands as illustrated in Figure 2A (see Materials and Methods for precise definition of the cognitive epochs). Two data sets were available for the present analyses: 1) Three animals recorded for up to 15 trials solely for the purposes of the present study. From these, only trials with good performance were selected ( = less than 3 test phase errors; median errors across all trials were 1, 0.5 and 2 for respectively for each animal), with an error defined as re-entrance into an arm from which food was already retrieved. 2) Six animals recorded for one or two trials from a previous study [33], which will be used to further confirm the results obtained with the “multiple-trial animals” and to conduct an explicit comparison of high (<2 errors) vs. low (>4 errors) performance trials. Average trial duration (±SEM) was 159.3±19.7 s across all trials and animals. With a standard binning for the spike density estimates of 0.2 s, this resulted in an average of 797±99 firing-rate vectors per trial (see further below for a discussion on data size effects).
To provide a direct comparison with previous approaches for constructing neural state spaces, Figure 2 shows three dimensional projections obtained in different ways from the first five trials of one of the multiple-trial animals which performs the task with less than three errors per trial. Consistent with our previous observations [33], the MSUA space shows a visually apparent segregation among the different task epochs (indicated by the color-coding), using either PCA (Figure 2B, left) or multi-dimensional scaling (MDS; Figure 2B, right) for the 3-dimensional reconstruction.
Figure 2C shows the same data projected into a 3-dimensional space using a Fisher discriminant analysis technique (FDA; see e.g. an application to MSUA spaces in [19]). Like PCA, FDA amounts to just a linear transformation of the original variables. However, unlike PCA, the directions sought are such that the differences between group means are maximized while at the same time within-group jitter is minimized along them (for the Oth order higher-dimensional spaces we used a regularized kernel-FDA which is equivalent to a standard (regularized) FDA for O = 1 [50]; see Materials and Methods and Text S1). The figure displays the flow field in addition to the data points, i.e. the speed and direction of movement of the neural population state at each time bin (computed as the difference between temporally consecutive vector pairs). While the flow field in the FDA-reduced original MSUA space may appear relatively disordered (Figure 2C, left), in the expanded space (Figure 2C, right) a consistent movement into each of the task related clusters at points far from any cluster center appears to occur (as will be statistically confirmed below). In summary, these 3-dimensional visualizations seem to suggest that different cognitively defined task epochs are associated with different population states which exhibit attractor-like properties (convergence of flow), a phenomenon that becomes apparent only after expanding the spaces to sufficiently high dimensionality using the techniques outlined in the previous section.
We stress that, in principle, expansion of spaces to much higher dimensionality is a well-known technique in statistical classification approaches to improve the linear separability of classes [27]. However, a serious statistical issue with such approaches is the potential problem of “over-fitting” the data: For instance, n+1 points can always be perfectly linearly separated in a n-dimensional space, even if their configuration is purely random. To circumvent this problem, two approaches which are standard in statistics (e.g. [46]) and machine learning (e.g. [44]) were employed here: First, a regularization term (fixed throughout the study; Eq. S3 in Text S1), which penalizes model complexity and thus reduces the efficient dimensionality of the fitted classifier (typically way beyond the nominal dimensionality), was included in the optimization criterion for the kernel-FDA. The technique of cross-validation (e.g. [44], [46]) is used in the next section for deriving this regularization term and the expansion order optimal for across-trial predictions (Materials and Methods). Over-fitting would imply poor generalization to new data sets not used for fitting the classifier, i.e. a high out-of-sample prediction error across trials. Second, the performance of the classification statistics on the original data was compared to bootstrap data in which the relation between neural population vectors and cognitive-class labels has been randomized. Such bootstrap samples have to be devised carefully such that they retain features of the original time series (like their temporal autocorrelations) which are not necessarily related to task-imposed structure, as explained in the sections to follow.
For determining the optimal state space we assessed whether the assignment of population-interaction patterns to task epochs could be correctly predicted in a test set of trials based on information obtained solely from a non-overlapping training set of trials, or, from another perspective, how stable the task-epoch-specific clusters in Oth-order expansion space are across multiple trials. To these ends, state spaces were reconstructed exclusively from the first set of 4 to 8 well-performed trials, and data points from the (non-overlapping) set of the last 4–8 well-performed trials were projected into this space (“forward predictions”). Vice versa, “backward predictions” from the last to the first trials were also obtained. If the neural dynamics remain largely invariant across multiple trials, then vector points on any subsequent trial should fall into the same clusters derived only from the first few trials. This analysis was performed for any pair of task epochs using the most discriminating direction as obtained by kernel-FDA within the expanded high-dimensional spaces. Assuming that the projections of the Oth-order population vectors from any two task epochs onto this maximally separating direction are normally distributed (which will almost inevitably be the case due to the central limit theorem, as the projections are sums of many random variables), for each population pattern ν(t) the probabilities P(ν(t)|C1) and P(ν(t)|C2) that it comes from one task-epoch or the other can be evaluated. Assigning population vectors to task epochs based on these probabilities yields a segregation error (SE) for each pair of task epochs defined as the relative number of misclassified population patterns ν(t) (see Materials and Methods for discussion of further advantages this brings over other kernel-based approaches). By chance this misclassification rate will be 50% since we fixed the prior probabilities P(C1) and P(C2) at 0.5 for any pair of epochs, such that the results would not be biased towards the longer-lasting epochs. Note that all time bins (population vectors) from a given task epoch class were entered into this analysis, regardless of whether they came from the same or from different trials.
For checking predictability across trials, the crucial aspect now is that the optimal discriminant direction was solely obtained from the first (or last) couple of (reference) trials, and then fixed and used for out-of-sample predicting the corresponding misclassification rate SEpredic (for “predicted SE”) of population interaction patterns to task-epochs for the non-overlapping set of last (or first, respectively) prediction trials (see Materials and Methods for more details). To evaluate the significance of the observed SEpredic, bootstrap data were constructed by randomly shuffling stretches of the ν(t) vector time series that retained entire trajectories form a given specific task epoch, i.e. each bootstrap replication preserved all temporal autocorrelations up to the length of the relevant task epochs. Consistent with the visual displays presented above, for O∼5 SEpredic was significantly lower (p<0.01) in the original as compared to the bootstrap data (Figure 3A; see Figure S2 for a schema on bootstrap construction). Note, however, that SEpredic for the bootstraps is also less than what would be expected by chance, i.e. <0.5, such that prediction accuracy in the bootstraps is above chance level. This is because the bootstraps retain original auto-correlations as indicated above, which by themselves may induce some state space clustering, irrespective of task-epoch membership. Surprisingly, in contrast to the case O∼5, for O = 1 (i.e., within the DC-MSUA space) predictability across trials was not significantly better in the original than in the bootstrap data. Thus there does not seem to be sufficient information in the lower-dimensional state spaces to allow prediction of population pattern assignments across trials. Rather, given the experimental noise and the potentially nonlinear state space structures, neural interactions have to be included to establish stable associations between task epochs and population patterns, or, in other words, further trajectory separation beyond the one achieved by delay-coordinates is indeed necessary to reveal across-trial stability. Specific comparisons for each pair of task epochs are shown in Figure 3B. Finally, for O>5 predictability starts to deteriorate again. Hence, it seems that there is a maximum order of activity products which would be required to optimally resolve task-epoch-related structure in the neural state spaces, a finding consistent across the different data sets studied (Figure 3C). We emphasize that this result does not imply that neural activity interactions up to some precise order (3rd–5th) are important– it only shows that below or above a certain expansion order generalization performance degrades, which can be the case for purely statistical reasons (i.e., simply because there are too few data or too few simultaneously recorded neurons to reliably estimate the optimum order of interactions).
On the other hand, the optimal orders we obtained do not seem to be completely arbitrary (in the sense of being determined purely by the number of data points and recorded units): First, similar optimal orders were also observed for the other two animals (Figure 4) which differed in the number of recorded units (18, 13 and 21, respectively) and the size of the training and prediction sample sets (5, 8 and 4 trials, respectively). Second, we performed additional controls by including subsets of neurons of differing size (Figure 4, upper left) and by artificially augmenting or decimating the data sets in a way that preserved the original distributions (Figure 4, right). Hence, we conclude that there is an organization of task-related population interaction patterns predictable across many trials which is optimally revealed by expanding the MSUA space by taking higher orders of activity interactions into account.
In a previous study [33] we had compared animals performing well on the task to animals which committed a lot of behavioral errors. We observed that in animals performing poorly state space segregation (task-epoch-dependent clustering) was generally comprised compared to trials on which only few (0 or 1) errors were committed. Here we re-addressed this issue using the methods developed above (Figure 5). Data from 8 trials (coming from 4 different animals) performing with less than two errors ( = “good performers”) and 8 trials (coming from 5 animals) with more than four errors ( = “bad performers”) were used. These two groups of trials were combined into two separate data sets for analysis (termed “single-trial” datasets). This works since the basic structure of the cognitively-defined classes was the same for all animals, i.e., the task obviously was the same for all animals, and population patterns specific for different task episodes like choices, rewards, or the delay phase, were a common feature of ACC activity. Since only a single trial with electrophysiological recordings, however, was generally available from each of these animals, results were cross-validated by removing each single one of the animals from the data set in turn (i.e., a jackknife validation [51]).
Consistent with our previous observations [33], discriminability in the MSUA space is significantly worse (Wilcoxon rank-sum test T13 = 113, p<0.05) for “bad performers” (Figure 5A, dark curve for O = 1) when compared to “good performers” (Figure 5A, gray curve for O = 1). However, as Figure 5A shows, for both groups discriminability significantly increases just up to expansion orders of about 5, i.e. the segregation error (SE) as defined further above (computed from FDA with the same regularization as above, see Materials and Methods) significantly decreases (Wilcoxon ranksum tests, p<0.03; see details in Figure 5 legend). Thus, as the maximum order O of the reconstructed state space is increased, cognitively relevant features of the neural dynamics are increasingly better resolved to the extent that an organized dynamics becomes evident even in situations where previous methods had failed (see [33]). However, as for the multiple-trials data analyzed in the previous section, SE for O>5 grows again for both groups (Figure 5A), suggesting once again that there may be a maximum order of activity interactions for which trajectories are optimally resolved.
Finally, and again consistent with previous results [33], although SE decreases for both groups, there still remains a significant difference between the low and the high performance groups even for O>3 (Wilcoxon test, p<0.04), confirming that still some of the state space organization is corrupted in bad performers. Detailed task-epoch comparisons are shown in Figure 5B. Similar results were obtained with information-theoretic measures of task-epoch segregation like the relative entropy (Kullback-Leibler divergence, e.g. [44]) between the conditional probability distributions of task-epochs given a specific firing-rate vector (Figure 5C; see Materials and Methods section). Moreover, further control analyses indicated that results are not significantly altered by using state spaces constructed by using different types of expansion, other classification criterions, or other smoothing parameters for the spike trains (as shown in Figure S3).
The most interesting aspect of the present methodological approach is that it permits to examine the flow of neural trajectories during performance of a cognitive task, dynamical properties that may not be well accessible in the unprocessed representation of MSU activity as demonstrated in the previous sections (Figure 3B, left). Here we analyzed the attracting behavior suggested by the three-dimensional visualizations more systematically. First, a simple statistical approach was taken. Activity flows were evaluated in the low-dimensional kernel-PCA projections of task epochs, since velocity vectors cannot be reliably obtained in the extremely high-dimensional expanded spaces (for similar reasons for which we used kernel methods before; see Figure S4 and Text S2 for further discussion). Figure 6 displays the speed of movement at each data point in these projections as a function of the likelihood of a population pattern given the task epoch to which it belongs, i.e. p(ν(t)|correct task-epoch classification), evaluated using FDA in the high-dimensional Oth-order spaces for the prediction set of trials (see Figure 3). If the task-epoch states have indeed attracting properties, one would expect that vector points which exhibit little movement should have a high likelihood of correct classification, reflecting the fact that these points should be found close to the cluster centers. Consistent with the idea that in low-order spaces trajectory flows should appear convoluted and disordered, for O = 1 velocities were evenly distributed across all regions of the state space, i.e. the velocity of movement of the neural state was largely independent of the likelihood of correct classification (Figure 6, left-top; O = 1). In contrast, for higher-order expansions the likelihood of correct classification rapidly falls off as the speed of neural state changes increases (Figure 6, left-bottom; O = 5), confirming that regions where trajectories move quickly are on average far from the cluster centers.
Although these results are suggestive, they by themselves do not conclusively rule out alternative explanations unrelated to the potentially attracting nature of the task-specific ensemble states, e.g. the tendency of extreme values to be followed by values closer to the mean simply by laws of probability (“regression to the mean”), auto-correlative properties of the time series, or by systematic deformations of the flow field induced by PCA. To statistically control for such alternatives, we performed a bootstrap test. The right column of Figure 6 shows results from the same analysis as performed on the bootstrap data when the temporal sequence of binned firing rates was inverted for all neurons within task-epochs. Therefore, task-epoch-specific lengths are preserved, but any causal relationships in the original time series are destroyed. For O = 1, the correct classification likelihood as a function of velocity behaves similar for bootstrap and original time series, but at higher expansion orders the fall-off of correct classification likelihood with vector velocity is significantly less steep in the bootstrap than in the original time series (paired t-test between the two slopes, p<0.001 for O = 5, see Figure 6 caption) as demonstrated by the linear fits to the log-linear graphs. In summary, different cognitively defined task epochs may potentially act as attracting states of the neural dynamics, i.e. regions of state space towards which all trajectories tend to converge with high likelihood and within which they remain bounded for some time.
While this analysis suggests attracting behavior related to the task epochs, it was performed on a three-dimensional representation in which velocity vectors could still be reliably determined. We therefore next sought to precisely quantify within the full high-dimensional spaces to which degree the (mathematical) conditions defining attracting states were met in the empirical data, with the statistical analysis based on the task-epoch boundaries defined previously. As the definition of these boundaries did not include any knowledge about putative attractor states, there is no a-priori reason why there should be strong convergence over time towards the center of these states. Attracting state conditions are illustrated in Figure 7A which shows a schema of different kind of convergent trajectories in the high-dimensional state spaces. Figure 7B shows within the 3-dimensional PCA projections some empirical examples of such trajectories which either cycle within or return to the task-epoch-specific population states. Figure 7C precisely quantifies, both for the single-trial data sets (red bars, left y-axis) and for the prediction-sets of trials in the multiple-trial data (blue bars, right y-axis), the fraction of trajectories which escaped again from the task-epoch specific clusters without returning to them within the given period (i.e. trajectories which are not of the kind “a” or “b” in Figure 7A). For O≈3–5, consistently across all task epochs this was only the case for ∼15% of the trajectories (across all 3 animals) when escape behavior was determined in the prediction trials while event boundaries were those defined in the non-overlapping reference set of trials, as shown in Figure 7C (blue bars, right y-axis; and ∼8% of the escaped trajectories when assessed within the reference set of trials, see red bars, left y-axis). Thus, these results further support the hypothesis that the task-epoch clusters constitute regions of convergence with >80% of trajectories returning to these states or bound within them. In summary, the quantitative analysis of trajectory flows in the optimal state spaces seems to confirm that different cognitively defined task epochs of the present memory-based decision making task act as high probability regions of convergence.
According to many neuro-computational theories, cognitive processes in the brain are implemented through the system's dynamical properties, i.e. the movement of neural trajectories among different attracting states that represent the contents of cognition (e.g. [2], [6], [34]–[36]). A number of previous experimental observations have therefore suggested the existence of attractor-like behavior in the nervous system, or were at least interpreted this way: Many of these studies dealt with forms of persistent [52], [53] or reoccurring spatio-temporal activity patterns [54] as they may be relevant to computational demands in working memory, e.g. temporary active maintenance of stimulus information required in a forthcoming choice situation [2], [5]. Other studies tried to establish direct links between neural attracting behavior and sensory or environmental representations [16], [20], [23]. With the recent progress in multiple single-unit recordings there has also been a rise in the application of advanced techniques from multivariate statistics and machine learning for reconstructing properties of the neural system dynamics or identifying re-occurring patterns, including different dimensionality reduction [16], [33], [55]–[57], pattern classification [33], [55], [58], and time series analysis approaches [18], [19], [59]. However, most of the previous experimental studies inferred attracting dynamics indirectly from, e.g., the property that neural activity after some time settled into one of several discrete states (e.g. [23]). In contrast, a more direct demonstration of a convergent flow of neural trajectories as a defining property of attractor-like structures has been, to our knowledge, mostly lacking so far. This may at least partly be attributed to the methodological difficulties associated with revealing the flow of trajectories directly in the experimentally accessed low-dimensional subspaces, i.e. the spaces spanned by the spiking activities of the set of recorded neurons (cf. Figure 1).
Here we therefore combined well-established approaches from nonlinear dynamics [28], [29] and statistical learning theory [27], [45], [46] for expanding spaces to a sufficiently high dimensionality such that the flow of trajectories becomes resolved. Since the expanded Oth-order embedding space can have very high dimensionality (for instance, for O = 5 and n = 30 neurons the dimensionality would be on the order of 106), specialized and strongly regularized algorithms (kernel-methods) were necessary to perform the relevant computations in these spaces [45]. However, in the present study this was done solely for computational tractability and numerical stability: The kernel function employed here is mathematically equivalent to vector products in the high-dimensional expanded spaces (see Materials and Methods), and hence does not change the nature of any of the results or arguments. We furthermore emphasize that kernel algorithms designed for very high-dimensional systems [26] are well-benchmarked techniques developed during the last decade [44], as are the delay embedding [38] and multinomial basis expansion [45] procedures employed here. All of these methods have been extensively tested with both simulated and real data in many areas of science [26], [44]–[46]. For instance, in functional neuroimaging the usage of kernel methods and high-dimensional classifiers becomes more and more of a routine now (e.g. [60]–[63]). The particular combination of these techniques for their application to electrophysiological data, on the other hand, to our knowledge presents a novel aspect of this work.
Using those approaches, we found that by augmenting the space with dimensions defined as products of neural firing rates, population interaction patterns belonging to distinct, cognitively defined task epochs were maximally separated and predictive of neural-behavioral state associations on future trials (cf. Figures 3, 4 and 7). More importantly, a consistent flow of neural trajectories and their convergence to task-epoch-specific ensemble states became apparent that was not obvious in the lower-dimensional embeddings of neural activity. Thus, the present memory-based decision making task seems to involve different (semi-)attracting states (in a statistical, probabilistic sense) among which neural activity may travel to implement task-related cognitive processes. These states had a cognitive interpretation as they were specific to particular task epochs. The organization of neural activity into different attracting states was furthermore related to behavioral performance: In animals exhibiting a high number of behavioral errors this structure was significantly degraded (Figure 5; see also [33]), perhaps reflecting a general “flattening of attractor basins” associated with diminished memory- and choice-related functions [64]. Therefore our results seem to support long-standing computational theories about the neural implementation of cognitive functions [20], [35], [52]–[54].
We observed that unfolding of trajectories and separation of task-epoch clusters became stable across trials when higher-order activity products were taken into account, but did not improve further when moving to arbitrarily high expansion orders. This, in other words, seems to imply that considering the joint activity constellations of a couple of neurons will still add information about the neural dynamics not easily or directly available from single unit activities, while still higher-order interactions may not be relevant: For sub-optimal state spaces the clustering into task-epoch-specific patterns was either unclear (O = 1) or had no predictive power across trials (O>6; cf. Figure 3). Note, however, that higher-order activity products are used here mainly as a statistical tool for disentangling trajectory flows and not for assessing the cognitive relevance of neural correlations. Thus, we cannot conclusively rule out, for instance, that adding many more neurons and data points to the state spaces than were available in the present study would shift the optimal expansion dimensionality to different orders. The specific value for the optimal expansion order obtained here may just reflect the well-known (in statistics; e.g. [46]) “bias-variance tradeoff” for our data set (in the sense of yielding low generalization errors, i.e. without over-fitting the data).
Nevertheless it is still remarkable that for all the different types of data sets studied here (multiple-trials vs. many animals), different numbers of recorded units, and different numbers of trials (and hence data points) a similar order of activity interactions appeared to be optimal. Similarly, the control studies reported in Figure 4 suggest that sample size effects cannot completely account for the specific optimality value obtained here. Indeed, a recent study, performed in visual cortex, revealed the importance of higher-order correlations in local neural ensembles like recorded here, while only second-order correlations seemed to be the relevant for information transmission across larger cortical distances [65]. The importance of higher-order correlations among neurons for information processing has also been stressed by many previous authors [43], [66], [67], e.g. by relating multiple-spike coincidence statistics to significant behavioral events [40], [42], [68], or by computing the information gained from correlations while decoding the current stimulus from the neural activity [69]. Some research had suggested that higher than second order correlations are redundant, at least in some preparations like the retina which may strongly differ in their structural and computational properties from the neocortex [67]. On the other hand, most recently it was suggested that some of the low bounds found in earlier studies may be an artifact of the limited number of experimentally accessed units [69]. Finally, studies in somatosensory cortex also found similar bounds on the maximum order of perceptually relevant neural activity interactions as suggested here [66].
Within the optimal order expansion spaces, the stable and attracting nature of the task-epoch-specific states became apparent (cf. Figure 7): The neural dynamics progressively slows down as trajectories approach the cluster centers (Figure 6) and the majority of trajectories cycles within or returns towards these states (Figure 7), indicating that there should be bounded regions of the neural state space which capture and contain neural trajectories. Just like in most previous studies indicating attractor-like dynamics (e.g., [16], [20], [23]), we cannot rule out, however, that these states are stimulus-driven, i.e. become attracting states only under the influence of certain (sensory or motor) stimulus conditions, rather than being a property of the intrinsic (autonomous) dynamics. For instance, in Wills et al. [20] or in Niessing and Friedrich [23] the different “categorical” steady state population responses which reflect attracting dynamics are observed for different types of external stimuli (spatial layout of a maze in the first and olfactory composite stimuli in the second case). Likewise, in our case specific spatial, motor, olfactory, or visual properties may be associated with the choice and reward periods.
There are three observations, however, which make it less likely that only external factors account for establishing different attracting states: First, also the delay period where the animals are confined to one arm of the maze and lights are switched off approximately acts as an attracting set of the dynamics, just like the other task epochs (Figures 3B and 5B). Second, the training and test epoch choice periods act as separate attracting states although they should share all sensory and motor features, but differ only in their memory requirements. Third, task-epoch specific states break down if the animals commit a lot of behavioral mistakes in the test period, yet one would assume that they experience similar sensory input and perform similar movements at each choice point. Thus, there must be some internal component in the generation of task-epoch specific states.
Nevertheless, true attractor states as mathematically defined (e.g. [70]) may be unlikely to exist in such an extremely non-stationary and high-dimensional complex system like the neocortex – rather, it seems more likely that neural information processing proceeds by stochastically itinerating among “semi-attracting” states which, for instance, may attract trajectories along most dimensions yet allow them to escape again along others [71]. This idea underlies many more recent conceptualizations of neural information processing (e.g. [35], [72]), and has also been advanced as a theoretical explanation of experimental results on sensory processing in locusts [16], [73]. For instance, a specific population activity pattern may be temporarily stable until some slow negative feedback mechanism has build up sufficiently to inhibit this currently active configuration [74], or until noise has driven the system out of this state again, i.e. until a stochastic transition between states has occurred [24], [25], [75]. It will be very difficult or even impossible to experimentally prove in such a high-dimensional and almost never stationary system under constant bombardment from external sources that any neural activity configuration is formally an attractor. Moreover, whether physiological phenomena as the ones reported here really match formal definitions of attracting states may be largely irrelevant from a computational perspective [35]. Rather, neural objects with semi-attracting properties as shown here could serve equally well (or even better, e.g. with regards to sequence processing) in most computational ideas about cognitive processing.
Does the high expansion order needed to fully reveal the converging dynamics of neural trajectories imply that the attracting states are very high-dimensional? Not necessarily: The key point of the delay embedding is to add more dimensions which are informative about the dynamics; many of the single-unit firing rate dimensions may be non-informative, i.e. may not contribute much to disentangling trajectories [28], and thus in principle could be omitted. The multinomial expansion on the other hand primarily serves to optimally pull apart noisy trajectories [45]. In a purely deterministic, noise-free system these dimensions would not be needed either to reveal the attractor. Indeed, the fact that convergent properties of the dynamics could be reasonably well evaluated in the 3-dimensional projections obtained by kernel-PCA suggests that the attracting states may in fact live in much lower dimensional subspaces [57]; which however were only fully revealed by properly expanding the space first before reducing it to the most informative dimensions by using kernel-PCA [76].
Finally, we stress that methods like the ones introduced here are widely applicable to almost any multivariate neural time series, including those obtained from various optical or functional imaging techniques, EEG, MEG [60], [63] or electrochemical techniques generating spatio-temporal time series. Thus, they may allow to address a number of previously unanswered questions about neural dynamics in many fields that require a proper unfolding and detailed resolution of trajectories not aided by across-trial averaging. Such techniques may also aid the discovery of common dynamical phenomena across tasks, species, and recording techniques. Here they revealed that ACC networks move among different state space regions, defined by specific population constellations of neural firing rates and their interactions, with a high likelihood of attracting neural trajectories. In this manner ACC networks may parse experience into meaningful task-relevant subcomponents.
All animals in this study were treated in accordance with the ethical guidelines set forth by the University of British Columbia and Canadian Council for Animal Care.
Briefly, animals were placed on a reverse light cycle upon arrival and given ad libitum access to food for one week. Surgery was then performed and the animals were allowed two weeks of recovery before maze training. For an in depth description of the multi electrode array fabrication and surgical procedures see Lapish et al. [33].
After recovery from surgery, all animals were trained on the delayed spatial win shift run on an eight arm radial maze. Each trial consisted of a training and test phase separated by a one minute delay phase. Prior to the task, the terminal end of all eight arms were baited with a sugar pellet (Research Diets, Inc., New Brunswick, NJ, USA). The training phase commenced by opening four of eight arms. Upon retrieval of the fourth sugar pellet in the training phase, the animal was locked in the last arm visited and the lights were extinguished for the delay. After the delay, the test phase began by allowing access to all eight arms and errors were scored as re-entries into previously visited arms. Upon completion of the trial by retrieving all eight sugar pellets, all arms were closed and the animal was re-confined to the center of the maze. Animals received one trial per day until they made one error or less for two days in a row, and then received a minimum of 10 trials per day. Data sets for the multiple trials analysis were selected from animals that were able to remain vigilant and attend to the task for ∼15 trials as evidenced by uninterrupted foraging.
In order to assess the population dynamic as the cognitive demands of the task vary, the whole time on task was divided into the following six epochs (Figure 2A): reward epochs (dark gray and red dots) during the training or test phases, respectively, correct choice epochs during training and test phases (blue and green, respectively), incorrect arm choice periods (yellow) during the test phase; and the entire delay period (light gray). Reward epochs were defined as the 1 s periods starting 200 ms before the points where the animal's nose reached a food cup during the training and test phase, respectively. Choice epochs were defined as periods starting 1.5 s before the arm choice and finishing 500 ms after it or before the reward period starts (assessed by visual video inspection).
Behavioral data were captured via a video camera (Cohu, Poway, CA, USA), recorded in Noldus Ethovision (Noldus, Leesburg, VA, USA), and exported via voltage in real time as Cartesian coordinates to the Neuralyx recording system and then scored offline. All data was acquired with arrays of 24 single-wire tungsten (diameter = 25 µm, impedance = 150–300 kΏ, California Fine Wire) electrodes implanted into the ACC (Figure S1A). Recordings were sampled at ∼30 kHz, band-pass filtered from 600–6000 Hz, and stored off-line for sorting and analysis. Spike channels were amplified 5,000–10,000 times and thresholds for detection were set to ∼50 µV, which corresponded to >5 times the root mean squared noise amplitude for the system. Spike sorting and classification was performed in Neuralynx Spikesort 3D (Neuralynx, Bozeman, MT, USA). Spike cluster assignments were based upon investigation of numerous principle components of the waveform (), and clusters lacking a well-defined boundary were excluded After classification, unrealistically low ISIs (≤10 ms) were removed as well as neurons with unrealistically high cross-correlations indicating the same neuron may have been captured on two different channels.
An intuitive introduction to our statistical methodology was provided at the beginning of the Results section, while most of the mathematical details can be found in the Supplementary Material.
Spike-trains from the n simultaneously recorded units were convolved with Gaussian functions to obtain statistically reliable estimates of spike densities from single trials (checking the range σ = 5–200 ms, see Figure S3 for values from 5–50 ms), normalized to the length of the whole trial (to yield a true probability density) and then summed and binned at 200 ms (approximately the inverse of the average single unit firing rate). Single unit spike densities were then combined into n-dimensional population vectors with components νi(t) for each unit i (e.g. [33], [59] as a function of time bin t. Small bin sizes (<50 ms) produce extremely sparse νi(t) series which became computationally prohibitive for the exact algorithm described below, and numerical approximations were required [45] Units for which 〈νi〉 <2% of the most responsive unit were excluded.
For the across-trial analysis, three different datasets consisting of 15 trials recorded on the same day were obtained from 3 animals. For each animal, only trials with ≥20 responsive units (see criterion above) were selected (10 trials from animal #1, 16 trials from animal #2, and 8 trials from animal #3). The first set of trials obeying above criteria constituted the reference set (trials 1–5 for animal #1, trials 1–8 for animal #2, and trials 1–4 for animal #3), while the last set of trials in the sequence formed the prediction set selected such that it had no overlap with the reference set. Furthermore, for each task-epoch, time series from the reference and prediction sets were constrained to have about the same length (number of vectors). For each of these two (reference and prediction) data sets, for each animal firing-rate vectors were then concatenated across trials to yield the two data matrices which entered into the analysis described further below. From our previous study [33] where animals were run only on a single trial after reaching criterion, two separate data sets from six animals were constructed from 8 trials performing with less than two errors (“good performers”) and 8 trials with over four errors (“bad performers”). Neurons from different networks were ordered according to their mean firing-rate, while low-responsive units were excluded as in the multiple-trial dataset.
For standard parametric testing, statistical test details can be found in the corresponding figure captions. For testing attracting properties of the task-epoch sets, nonparametric tests were used based on conservatively designed bootstrap data (100 replications used for one-sided comparisons at p = 0.01) as explained in the corresponding text sections and in Figure S2. For the control analyses shown in Figure 4, original task epochs were artificially augmented 5–20 times (generating ∼104 data points) and decimated by a factor of 0.8-0.6. This process did not significantly alter the original distributions, auto- and cross-correlations for all units.
An instantaneous population firing rate vector in MSUA space, obtained by convolution of the spike trains with Gaussian functions as described above, is given by [33]. For univariate time series ν(t), delay embeddings are usually constructed by forming vectors from temporally delayed values with delays (lags) τi. These are typically chosen to correspond to p successive minima of the autocorrelation function (or mutual information) where p would be high enough to unfold (pull apart) trajectories within this delay-coordinate space [38], [77]. In general, the reconstructed spaces should have dimensionality p = 2×D+1, where D is the attractor dimension [28]. Similar ideas can be applied to multivariate systems [29]. The attractor dimension is often estimated via the correlation dimension, which, however, will not provide sensible answers in sparse high-dimensional spaces as the ones examined here (see [29]; Figure S4 and Text S2). Moreover, the use of large delay coordinate maps would result in an extensive loss of data and hence poor statistics. Therefore, additional non-delay variables are sought to effectively disentangle noisy trajectories.
The first step in our approach is to construct a reduced multivariate delay-coordinate map which should simply ensure that trajectories do not significantly cross each other. This auxiliary DC-MSUA space, defined by vectors , contains only a single lag for each unit optimized to be the first minimum of the average cross-correlation between pairs of units' firing-rates. The resulting lags ranged from just one time bin to <5% of the task-phase length. Note that the main purpose of these lagged variables is to add axes to the space which contain information about the system dynamics not captured by the current state of the firing rate variables, therefore the choice of lags such as to minimize cross-correlations. In fact, the use of more than one delay per unit did not improve across-trial predictions (data not shown).
After this step, differences between trajectories were further amplified by combining these variables into new functional forms, in accordance with ideas from statistical learning theory [27]. As we were specifically interested in functional forms with a biological meaning, this was done by adding higher-order products of the units' firing rates as new coordinates to the neural state space. The oth-order interaction of n-units omitting lags for notational convenience, is defined by(1)By construction of the smoothed firing rate vector (see above) each axis φ(t) is the net sum of probabilities of o multiple spikes independently occurring across n neurons (e.g. [78]). For the frequent case that a single spike is contained within a single bin, and for small smoothing windows σ, φ(t) tends to represent a pattern of multiple spike-co-occurrences (a “poly-synchronous” pattern [79]). Now, the Oth-order delay-interactions coordinate map consists of all oth-order firing-rate products with o = 1…O. Vectors in this high-dimensional space will be denoted by Φ(t). For instance, a vector in the space corresponding to O = 2 is defined by(2)The dimensionality p(O) of such a space is typically p∼105–109, much larger than the number of task-epoch vectors which is on the order of ∼103. Note that this approach sharply contrasts with other methods where the MSUA space dimensionality is instead further reduced by exploiting correlations among units (e.g. [14], [15], [80]). As was noted further above, the delay-coordinate map suffices to remove overlap between trajectories in an ideal, purely deterministic system. On the other hand, the multinomial basis expansion defined above helps to achieve an optimal separation in a statistical learning sense when dealing with highly noisy systems (e.g. Figure 3).
Explicit computations in such extremely high-dimensional spaces are associated with numerical and computational problems which can be solved by the so-called “kernel trick” [45]. In this context a kernel is a function which represents a vector product in a high-dimensional space without explicitly computing the dot product of the vectors. Here, for any two high-dimensional vectors Φ(t) from the expanded Oth-order space occurring at times ta and tb, respectively, the kernel function is given by
(3)Thus, the function on the right hand side operating on the low-dimensional firing rate vectors ν(t) is mathematically equivalent to (and uniquely defined for) a dot product between vectors Φ(t) from the much higher-dimensional Oth-order space [45]. See Text S1 and [81] for further motivation for the use of this kernel.
Within the mathematical framework of kernel algorithms, high-dimensional covariance matrices are replaced by kernel matrices in the reformulation of classical statistical procedures like PCA or FDA. Kernel matrices were computed for each possible pair of task epochs (such that ta and tb in Equation 4 may correspond to two different time points of the same epoch, or to time points from two different epochs), and then used to build a classifier using Fisher's discriminant (FD) criterion. FD analysis works by maximizing the difference between task-epoch means while minimizing within-task-epoch (co-) variances, i.e., by finding the direction Ω of the high-dimensional Oth-order space along which the overlap between two task-epoch distributions is minimized [45], [50]. Since in the expanded spaces the number of dimensions (variables) d is extremely high, in fact much higher than the number of observations m, means and covariance matrices cannot be explicitly computed, as stated above, and thus for the FD analysis all computations on high-dimensional vectors are reformulated in terms of a kernel matrix K of much smaller dimensionality (equal to m2<<d2; see Text S1). By usage of the kernel matrix K, the projections x(ti) of high-dimensional vectors Φ(ti) onto the optimally discriminating direction Ω are obtained by(4)where the m elements of the vector α are derived as e.g. explained in Schölkopf and Smola, [45] and in Text S1. Since the projected values x(ti) on the most discriminating axis represent linear combinations of up to 109 random variables (one variable per dimension), the projected data will be approximately normally distributed according to the central limit theorem (e.g. [44]). Hence, building on this assumption of approximate normality, a Bayes-optimal classifier (the one with theoretically best performance) can be defined on this most-discriminating axis (where equal priors were used here for not biasing the results according to the lengths of the sampled task epochs). From this, classification (separation) errors (SE; cf. Figure 5), likelihoods p(ν(t)|C) of classification into task-epoch C, posterior probabilities P(C|ν(t)) (using Bayes criterion), and 99% confidence intervals are straightforward to obtain. The (discretized) Kullbach-Leibler divergence [e.g. 44] was computed as a measure of the distance between these Gaussian posterior distributions corresponding to any two tasks epochs C1 and C2. It was estimated for each Oth order expansion (Figure 5C) and is given by(5)The utilization of normal probability theory represents a fundamental advantage over other approaches specialized for high-dimensional spaces (e.g. support-vector-based classifiers [27]) which may have similar classification performance [45] but do not easily permit other aspects of the present statistical analysis.
For the across-trial analyses, optimal directions Ω for each task epoch pair were obtained using exclusively the first set of (reference) trials, Φref. This direction was then fixed for computing the projections xpredic(ti) of vectors Φpredic(ti) from the prediction set onto Ω to yield the predicted SE (SEpredic):(6)where the vector αref is the one obtained from the reference set and K represents projections of prediction set vectors into the reference space. A brief summary of these algorithms can be found in Text S1 [45].
A regularization penalty was furthermore added to the kernel matrices to ensure a low generalization error (loosely speaking, a regularization factor automatically constrains the number of free parameters to reduce out-of-sample prediction errors; e.g. [27], [45]). This regularization was optimized such that SEpredic was minimal for animal #1 and then it was fixed for all other analyses (because of this regularization, for instance, in-sample SE never decreases to zero for the expanded spaces in Figures 3A and 5A). Prediction errors were found to be invariant for large enough values of this regularization penalty (as demonstrated in Figure S3). The robustness of the present approach with regards to different basis functions used in the expansion (and thus different definitions of the kernel) is also discussed in Figure S3. Finally, we also investigated how unsupervised clustering approaches perform on the DC-MSUA spaces, and noticed that they reliably pick up only the delay vs. training/test phase differences in this lower-dimensional representation (see Figure S5 for an example).
Kernel-FDA [50] and kernel-PCA [47] were used to obtain three-dimensional visualizations for each high-dimensional task-epoch state. Three-dimensional projections were also used for determining velocity vectors (cf. Figure 6), as these cannot be efficiently computed in high dimensions (a problem running under the label “curse of dimensionality”; e.g. [44]). Kernel-PCA proceeds in much the same way as ordinary PCA, except that – like kernel-FDA – it works on the kernel matrices defined above instead of directly on the high-dimensional covariance matrices (see brief summary in Text S1). Thus, the three orthogonal dimensions capturing the largest amount of data variance in the high-dimensional spaces were obtained. Additional discussion about the adequacy of these three-dimensional velocity vectors as obtained by kernel-PCA can be found in Figure S4 and in Text S2. Finally, note that, apart from the convergence analysis shown in Figure 6, these three-dimensional reductions served only for the purpose of visualization, while all statistical analyses were performed on the full high-dimensional spaces (see Figure 7C).
Analysis software was implemented in MatLab (Mathworks Inc., MA, USA) and is freely available in http://www.bccn-heidelberg-mannheim.de under the terms of the general public license (http://www.gnu.org/licenses/).
|
10.1371/journal.pgen.1003764 | Endogenous Stress Caused by Faulty Oxidation Reactions Fosters Evolution of 2,4-Dinitrotoluene-Degrading Bacteria | Environmental strain Burkholderia sp. DNT mineralizes the xenobiotic compound 2,4-dinitrotoluene (DNT) owing to the catabolic dnt genes borne by plasmid DNT, but the process fails to promote significant growth. To investigate this lack of physiological return of such an otherwise complete metabolic route, cells were exposed to DNT under various growth conditions and the endogenous formation of reactive oxygen species (ROS) monitored in single bacteria. These tests revealed the buildup of a strong oxidative stress in the population exposed to DNT. By either curing the DNT plasmid or by overproducing the second activity of the biodegradation route (DntB) we could trace a large share of ROS production to the first reaction of the route, which is executed by the multicomponent dioxygenase encoded by the dntA gene cluster. Naphthalene, the ancestral substrate of the dioxygenase from which DntA has evolved, also caused significant ROS formation. That both the old and the new substrate brought about a considerable cellular stress was indicative of a still-evolving DntA enzyme which is neither optimal any longer for naphthalene nor entirely advantageous yet for growth of the host strain on DNT. We could associate endogenous production of ROS with likely error-prone repair mechanisms of DNA damage, and the ensuing stress-induced mutagenesis in cells exposed to DNT. It is thus plausible that the evolutionary roadmap for biodegradation of xenobiotic compounds like DNT was largely elicited by mutagenic oxidative stress caused by faulty reactions of precursor enzymes with novel but structurally related substrates-to-be.
| Many bacteria have acquired the capacity of metabolizing chemical compounds that have never been in the Biosphere before the onset of contemporary synthetic chemistry. However, the factors that shape the new metabolic properties of such microorganisms remain obscure. We examined the performance of a still-evolving metabolic pathway for biodegradation of 2,4-dinitrotoluene (DNT, an archetypal xenobiotic compound) borne by a Burkholderia strain isolated from soil in an ammunition plant. The biodegradation pathway likely arose from a precursor set of genes for catabolism of naphthalene (although Burkholderia does not degrade this compound any longer), and is now advancing towards the new substrate, DNT. We found that the action of the first enzyme of the biodegradation pathway, a Rieske-type dioxygenase, on the still-suboptimal substrate (DNT) generates a high level of endogenous reactive oxygen species. This, in turn, damages DNA and increases mutagenesis, ultimately resulting in the creation of novelty that may foster evolution of xenobiotic-degrading variants of the strain hosting the biodegradation pathway. The very metabolic problem thus somehow seems to stimulate the exploration of the solution space. Our data is fully consistent with the notion that stress caused by faulty dioxygenation of DNT accelerates the rate of bacterial evolution.
| A diversity of xenobiotic compounds has made it to the environment in large amounts since the onset of synthetic chemistry due to urban and industrial activities [1]. Although many of such compounds bear bonds that are rare or non-existing in the natural realm, it is not infrequent to isolate bacteria able to use them as C, N and/or energy sources, especially in places with a history of chemical pollution [1]–[3]. This bear witness of the ability of existing metabolic networks to conquest new and diverse chemical landscapes. The process by which a given bacterium comes across a genetic and enzymatic solution to the challenge of degrading a new chemical is not trivial [4], [5]. Not only enzymes and complete pathways have to evolve new substrate specificities, but also they should be expressed at the right time and location. In addition, the resulting metabolic currency should be wired to the central biochemical network for eventual biomass buildup. This surely requires a mutual adaptation between the host and the pathway itself that involves multiple changes in both the catabolic genes and the rest of the cell transcriptome. On this basis, it comes as a surprise that many xenobiotic compounds containing strong chemical bonds and/or substituents can be degraded by environmental bacteria after being in the biosphere for only a relatively short period of time. One conspicuous case in this respect is that of nitroaromatic compounds. Although many chemical structures with C-NO2 bonds are found in natural products [6], [7], nitroaromatics could never become a significant selective pressure before their massive production and release in connection to the modern industry of explosives [1], [8]. Yet, a number of bacterial strains have been isolated that can mineralize many of such unusual chemical structures [1], [2], [9]. However, the mere genetic drift of pre-existing enzymes and operons that catabolize similar compounds towards new substrates can hardly explain the high rate at which new strains and catabolic operons appear [10]. The question thus arises on whether evolution of given catabolic properties benefit from some inherent accelerator of the process somehow encoded in the enzymes and genes involved.
Environmental bacteria capable of biodegradation of 2,4-dinitrotoluene (DNT) afford an exceptional opportunity to examine the factors at play in the emergence of new metabolic abilities. One of such isolates, Burkholderia sp. strain DNT, was isolated from contaminated surface water of an ammunition waste plant on the basis of utilizing DNT as C and N sources [11]. This capacity is due to the action of a route for catabolism of the nitroaromatic compound (Fig. 1A). The phylogeny of the dnt genes and the biochemical properties of the encoded products indicate that this cluster for DNT biodegradation has originated from a precursor pathway for catabolism of the naturally-occurring hydrocarbon naphthalene (Fig. 1B; [12], [13]). Still, the same data indicates that the pathway is even now evolving since the kinetic coupling of each of the biochemical steps is not well balanced [11], the regulation of the system keeps the same effector profile than the precursor operon [14], there are transposon remnants and vestigial genes in the genetic cluster [12] and the strain hardly grows on the substrate of interest [15]. These are all indicators that this strain has started its itinerary towards DNT catabolism but has not yet found an optimal solution to the multi-tiered problem of its viable degradation.
In this work we have examined possible physiological and biochemical drivers that frame the evolution of Burkholderia sp. DNT towards DNT catabolism and, by extension, many other environmental bacteria towards new chemical compounds. As shown below, reactive oxygen species (ROS) brought about by the faulty performance of the first enzyme of the pathway (DntA dioxygenase) on the xenobiotic substrate seem to be a key agent of the process. This is because ROS is translated into DNA mutagenesis and diversification of the host strain that foster the emergence of novel adaptive phenotypes. Moreover, the results provide an evolutionary logic to the abundance of enzymes containing Rieske-type Fe centers (as DntA) in pathways for biodegradation of xenobiotic compounds recently released to the environment [16]. This work showcases an evolutionary scenario in which the physiological stress caused by a metabolic problem triggers also the genetic diversification necessary for exploring the solution space to the same problem.
The experimental system of choice for examining the transition of the DNT pathway from degrading naphthalene to metabolizing DNT involves only three players: strain Burkholderia sp. DNT and the two chemicals at stake. It should be noted that, according to the very high similarity between the leading ring oxygenases of the nag route in Ralstonia sp. U2 for naphthalene degradation and the dnt pathway of Burkholderia sp. DNT for DNT catabolism (Fig. 1B), the extant DntA enzyme retains the ability to act on the first substrate to produce 1,2-dihydroxy-1,2-dihydronaphthalene [13], [17]. Owing to its current catabolic role, this multi-component enzyme dioxygenates DNT in positions 4 and 5 to yield 4-methyl-5-nitrocatechol (4M5NC, Fig. 1A). In order to have a gross estimate of the physiological state of the cells when exposed to the former and current DntA substrates, we simply examined the survival of Burkholderia sp. DNT pre-grown on M9 minimal medium containing succinate and then exposed respectively to either chemical at a final concentration of 0.5 mM. In order to differentiate between the intrinsic effect of these compounds from their interplay with the dnt-encoded enzymes, we also employed a variant strain cured of the dnt genes (Burkholderia sp. DNT Δdnt), that has no chance of transforming the substrates at stake. While both naphthalene and DNT led to a very significant loss of viability of Burkholderia sp. DNT (Fig. 2A), which went down to <30% in the presence of DNT, addition of the same chemicals to the strain lacking the dnt genes had a considerably lesser toxicity (Fig. 2B). It should be noted that neither naphthalene nor DNT are inducers of the dnt pathway [14], which is expressed through its native promoter at a basal level under these experimental conditions. The detrimental effect of the former (naphthalene) and new (DNT) substrate can thus be both traced to the presence of the dnt genes and plausibly to the activity of their encoded products on each of the chemicals. In order to qualify the observed toxicity of the DntA substrates we measured the activity of the enzyme glucose-6-phosphate dehydrogenase (G6PDH) in cells exposed or not to DNT. The activity of this enzyme is a physiological stress marker in most bacteria [18], [19] and its increased activity reports a general response to metabolic hardship. When Burkholderia sp. DNT cells were exposed to DNT, the level of G6PDH activity more than doubled in the presence of DNT (Fig. 3). In contrast, the equivalent strain lacking the dnt genes had a G6PDH activity minimally affected by addition of the nitroaromatic compound. Taken together, these data suggested that activity of the dnt-encoded products on the two substrates tested caused physiological stress that was more pronounced in the case of DNT.
What could be the reason for the remarkable toxicity of DNT in cells expressing dnt genes? Since an increased activity of G6PDH would lead to enhanced NADPH turnover rates required to counteract oxidative stress [18]–[20], one plausible scenario was that the encounter of DNT with the dnt-encoded oxygenases could release ROS as a side product of the oxidative steps that shape the pathway (Fig. 1A). This is a frequent situation in evolving biodegradative routes of aerobic bacteria [21]–[24] and it often becomes a veritable bottleneck in the catabolism of some compounds [19], [25], [26]. On this background, we set out to measure ROS generation in Burkholderia sp. DNT cells treated with DNT along with their Δdnt counterparts. To quantify oxidative stress we resorted to staining cells with the ROS-activated green fluorescent dye 2′,7′-dichlorodihydrofluorescein diacetate (H2DCF-DA) followed by quantitative flow cytometry [19]. The minimum and maximum response levels of reference were fixed using succinate-grown cells added with either the H2DCF-DA solvent carrier (dimethyl sulfoxide, DMSO) or the same with 1.5 mM H2O2. Fig. 4 accredits the performance of the test and shows raw data obtained when cells were exposed to the toxicants at stake. Perusal of the results indicated that DNT caused an extraordinary and somewhat unexpected level of intracellular ROS, the intensity of which nearly equaled that produced by straight addition of H2O2 to the medium. In comparison, cells lacking the complement of dnt genes showed a much lower level of ROS in the presence of DNT, indicating that both the substrate and the genes were required for such a surprising synergy to cause oxidative stress. Addition of the ancestral substrate naphthalene also caused a lower, but still significant, level of ROS in Burkholderia sp. DNT (Fig. 5A and B). Even though permeability of the dye can be affected by the aromatic compounds tested [27], [28], pairwise comparisons of the different strains and substrates revealed the extension of the stress produced by each compound. Since the DntA originates in a precursor naphthalene dioxygenase [12], [13], we hypothesized that some of the ROS could be originated from an uncoupled, non-productive oxygen-delivery reaction between the evolved enzyme and the suboptimal substrate DNT, as reported for similar Rieske-type dioxygenases [29], [30]. This possibility placed the focus of the ensuing experiments on the first step of the DNT biodegradation route.
Since the most critical move in any biodegradative route of aromatics is the initial step that activates the ring for overcoming the resonance energy that stabilizes their structure [31], we concentrated on the leading reaction that converts DNT to 4M5NC (Fig. 1A). This compound, as is the case for other catechols, is highly toxic [32], [33]. Furthermore, it is known [11] that such yellow-colored intermediate transiently accumulates in the medium prior to be channeled towards the next step of the degradation route (Fig. 6A). The toxic effect of DNT on Burkholderia sp. DNT could thus originate in either faulty reactions of DntA on its substrate, in accumulation of 4M5NC or in both. Our strategy to distinguish these possibilities was to generate a strain that overproduced DntB (Fig. 1A). This enzyme is a monooxygenase that eliminates the remaining nitro substituent of 4M5NC to produce 2-hydroxy-5-methylquinone (2H5MQ; [34], [35]). Higher levels of DntB are thus predicted to drain 4M5NC faster towards 2H5MQ -thereby allowing us to separate the intrinsic physiological effect of DntA action from that of the toxic catechol that results from the reaction. To this end, we generated strain Burkholderia sp. DNT dntB↑, in which extra copies of the gene encoding the second oxygenase of the pathway were expressed from a constitutive promoter in a broad-host-range vector. Fig. 6B verifies such a prediction, as cultures of Burkholderia sp. DNT dntB↑ released a much lower amount (<20%) of 4M5NC to the medium. In the same setup, strain Burkholderia sp. DNT Δdnt did not accumulate any intermediate, while naphthalene would be biotransformed to 1,2-dihydroxy-1,2-dihydronaphthalene by any of the dntA+ strains. Once we had the three isogenic variants in hand, we set out to measure intracellular ROS production in wild-type, Δdnt and dntB↑ cells exposed to either DNT or naphthalene (Fig. 5A, B and C, respectively). These results provide an answer to the questions raised above. First, that DNT causes high ROS levels in both the wild-type and dntB↑ strains (Fig. 5A and C) clearly indicates that the bulk of oxidative stress can be traced to the very reaction of DntA with DNT, not to the product of the catalysis. Second, the effects of naphthalene were comparatively lower than those elicited by DNT, but they were still significant in respect to the untreated control conditions. ROS generated from the encounter of this substrate with the dnt-encoded enzymes must stem from uncoupled, faulty reactions with the leading DntA oxygenase, as there is no other enzyme that can recognize this aromatic compound as a potential but ultimately non-productive substrate. Since both the ancestral and the new substrate of the first dioxygenase release ROS when facing DntA, the cognate reaction has probably left behind a former biochemical optimum (i.e., when acting on naphthalene) but has not yet reached a new one with DNT.
In a further effort to identify the specific types of ROS that result from the above mentioned processes, we tested cells treated with the various substrates with nitro blue tetrazolium (NBT), a reagent that is specific for superoxide production. Consistently with the flow cytometry data above, the highest indications of superoxide presence were found in wild-type and dntB↑ cells exposed to DNT, while a lower level was detected in the Δdnt strain and in bacteria exposed to naphthalene (Fig. 7). Since superoxide originated in defective redox reactions promotes hydroxyl-radical formation and consequent DNA damage [36], [37] we wondered whether ROS stemming from the faulty oxygenations discussed above could eventually translate into an insult to the genome of Burkholderia sp. DNT. One direct way of quantifying such damage is the measurement of the 8-hydroxy-2′-deoxyguanosine (8-oxoG) content of genomic DNA as a coarse descriptor of HO• attack to purines [37], [38]. On this background we resorted to an immunoassay for quantifying 8-oxoG levels in cells exposed to DNT, using H2O2 as a positive control of oxidative stress. As shown in Fig. 8A, DNT indeed caused a 1.4-fold increase in the share of damaged purines in the Burkholderia sp. DNT genome (as opposed to a 2-fold increase elicited by H2O2). As such a chemical damage to DNA bases triggers the SOS response and eventually increases the mutation rate of the bacteria that undergo the insult [36], we next wondered whether the ultimate consequence of the uncoupling of the ring-hydroxylating reaction performed by DntA with their current and ancestral substrates was to increase genetic diversity by enhancing mutation rates. Since virtually nothing is known about the SOS response in Burkholderia sp. DNT, we directly measured such mutation rates as a descriptor of emerging genetic novelty. For this, we employed the standard test of appearance of rifampicin-resistant (RifR) clones under the various conditions assayed. As shown in Fig. 8B, both naphthalene and DNT triggered a considerable increase in the appearance of RifR colonies which was not noticeable in the strain deleted of the dnt genes The mutagenic effect of naphthalene in this context was slightly more pronounced than what could be expected from the sheer data on ROS production shown above. We speculate that ROS could react chemically with this bicyclic aromatic compound and generate additional DNA-intercalating agents (e.g. naphtoquinones and naphtodiols [39]) with a separate mutagenic action on DNA. In any case, it is worth noticing that maximum novelty (as measured with this procedure) is accompanied by an acute lethality (Fig. 2), so that the survivors to DNT exposure are more capable to explore the possible solution space to the next adaptive challenge (metabolic or otherwise) than those which had not been diversified because of the phenomena described here.
Bacteria that inhabit environmental niches with a history of pollution by xenobiotic compounds offer a phenomenal experimental system for examining the expansion of the existing metabolic networks into new chemical spaces [10]. At least three bottlenecks must be overcome for the successful emergence of a novel catabolic pathway able to deal with a new-to-nature chemical structure. First, suitable enzymes must develop the right substrate specificity towards the new compound [40]. In most cases, such enzymes stem from precursors that act on structurally related chemicals but shift specificities through intermediate steps where the old and the new activities coexist in the same protein at no fitness cost [40], [41]. Second, the genes encoding the new enzymatic activities should be expressed only when required i.e. when the substrate-to-be is present in sufficient concentrations to grant a suitable return. More evolved catabolic systems are typically subject to a tight regulation by inducer substrates [42], while more recent counterparts are often expressed at low constitutive levels. Not infrequently, still evolving biodegradative operons carry along regulatory systems that respond to former substrates and not the new ones [43]. Finally, the metabolic action of the evolved pathways must be connected to some growth benefit that ultimately fosters its selective advantage [43]. This is brought about directly by entering metabolic currency from the biodegradation process into the central metabolism, or indirectly, by making cells more resistant to otherwise detrimental endogenous or exogenous stress. Even horizontal gene transfer granted, the number of mutations and genetic events that might be required for the emergence of a fully competent biodegradative strain able to deal with a novel xenobiotic compound can be considerable, far more than those necessary e.g. for new antibiotic resistances [44]. It thus comes as a surprise that a good number of xenobiotic compounds are indeed degraded by environmental strains not much after their production by the chemical industry [8], [10]. In view of all this, it could well be that evolution of biodegradative routes for xenobiotics is accelerated by factors inherent to the nature of the enzymatic reactions involved in the process.
The metabolic pathway of Burkholderia sp. DNT that acts on DNT has all what is biochemically needed for its complete biodegradation but it is clearly not optimal yet for coupling the process to efficient growth. Furthermore, we show above that the DNT dioxygenation reaction that constitutes the first step of the pathway exacerbates ROS production (Fig. 4 and Fig. 5). It is very likely that such reactive species result from the failure of the Fe-containing center of the α subunit of DntA to deliver active oxygen atoms to positions 4 and 5 of the DNT aromatic ring (Fig. 1A). Since the DntA enzyme complex originates in a precursor naphthalene dioxygenase (Fig. 1B), it is likely that such uncoupling of the reaction stems from a still poor geometry of the substrate-enzyme recognition. On the other hand, the enzyme still recognizes naphthalene as a substrate, but the reaction is not metabolically productive and also causes considerable ROS. These data not only pinpoints the DntA complex to be in the midst of a transition between the old and the new specificities, but also suggest key events in evolution of new xenobiotic-degrading strains. Since ROS mediate DNA damage, which in turn brings about a mutagenic SOS response [36], [37], [45], it seems that the generation of diversity associated to a faulty dioxygenation reaction could contribute to the exploration of the metabolic novelty space in a faster fashion that could be expected by a background, spontaneous mutagenesis. This evolutionary scenario is reminiscent - but not exactly identical, to that proposed for the rapid evolution of antibiotic-resistant bacterial strains, which is thought to be fostered as well by the boost of ROS and the ensuing mutagenesis and novelty generation that precedes death of bacteria treated with antimicrobials [46]–[49]. In the case of the DNT-degrading strain presented above, we could in fact trace the entire process to the master dioxygenation reaction that initiates the metabolic pathway. Because of its inherent enzymatic mechanism (e.g. Rieske-type active center) this reaction is prone to uncouple dioxygenations of suboptimal substrates and release ROS [29], [30]. In the case of strain DNT, the process might also be assisted by reactive nitrogen species connected to the nitrite [50] released as a by-product of the DntA and DntB-catalyzed reactions (Fig. 1). This issue deserves further studies, as nitrite has been claimed to be detrimental for in situ DNT degradation by environmental bacteria [51]. Additional mutagenic effects of biodegradation intermediates downstream of 4M5NC cannot be ruled out. However, since the only metabolite that is significantly accumulated during DNT turnover is precisely 4M5NC (the substrate of DntB, [11]), it is unlikely that the toxicity generated by such intermediates could be ultimately meaningful.
How general could be such a substrate-driven generation of diversity for evolution of other xenobiotic-biodegradation routes? A good number of results are found in the literature that can be interpreted under this light. For instance, the abundance of Rieske-containing enzymes in biodegradative operons [16] could reflect a key role in generation of novelty both in the catabolic genes themselves and in the host. This is an important detail, as changes that originate a better biodegradative strain involve genes other that those directly involved in the corresponding pathway. It is remarkable also that the transcriptomes and proteomes of strains degrading a suite of recalcitrant aromatics (e.g. polychlorobiphenyls and polyaromatic hydrocarbons) are overrepresented with functions that counteract oxidative stress [21]–[23]. Finally, it is not unusual to find oxidative pathways for xenobiotics in bacteria which are typical macrophage-resistant pathogens (e.g. Burkholderia and Mycobacterium) and thus adapted to endure a strongly oxidative milieu [52], [53]. Such routes might only develop in hosts able to endure the ROS-mediated lethality associated to the evolutionary roadmap from the enzymatic recognition of one substrate to another.
In sum, the data above strongly argue that evolution of DNT biodegradation pathways and possibly of many other catabolic systems, ultimately benefit from the mutagenic stress caused by faulty reactions of pre-existing enzymes on suboptimal substrates. The term anti-fragility (as opposed to robustness, [54], [55]) has been recently coined for describing such systems that reach a higher peak of efficacy after having nearly collapsed with stress and shocks (Fig. 9). The results discussed in this work seem not only to place the appearance of new xenobiotic-biodegradation strains within such an evolutionary scenario, but also suggest experimental approaches to accelerate their emergence in the Laboratory.
Burkholderia sp. DNT has been described before as a DNT degrading specimen [11]. The sequence, organization and function of each of the dnt genes, involved in mineralization of this aromatic compound, have been reported elsewhere [14]. The action of Burkholderia sp. DNT on the nitroaromatic substrate is very slow i.e. depletion of 1 mM DNT takes 3–5 days [15]. A spontaneous derivative of Burkholderia sp. DNT lacking the DNT-degradative phenotype was isolated by three consecutive cultivation rounds on LB medium, and the absence of key dnt genes (dntA, dntB and dntD) in this derivative was confirmed by PCR (data not shown). Accumulation of the yellow intermediate 4M5NC was spectrophotometrically quantified by absorbance readings at 420 nm [14]. The high segregational instability of the DNT plasmid, carrying the dnt genes, has been previously reported when cells are grown in rich media [56], [57]. Maintenance of the dnt+ phenotype (and thus of the DNT plasmid) was verified prior to each of the experiments by examining accumulation of 4M5NC (the first biodegradation intermediate; Fig. 1) in 200 clonal cultures added with 1 mM DNT. This control indicated that growth in minimal medium with succinate as the sole C source caused <1% loss of the catabolic ability of processing DNT. A strain overexpressing dntB, encoding the second enzyme of the pathway that removes 4M5NC, was constructed as follows. Oligonucleotides dntB-attB1-F (5′-GGG GAC AAG TTT GTA CAA AAA AGC AGG CTG CTC TGT CGA TGA TTT GAG GA-3′) and dntB-attB2-R (5′-GGG GAC CAC TTT GTA CAA GAA AGC TGG GTG TTG CGC ACC TGT CAT CG-3′), including the attB secuences required for Gateway-based cloning (shown in italics in the corresponding sequences), were used to amplify dntB from Burkholderia sp. DNT. The amplicon was cloned in the pDONR™/Zeo vector (Life Technologies Corp.) according to the manufacturer's instructions, and moved to the Gateway vector pBAV226 [58] (kindly provided by Jean T. Greenberg, University of Chicago, USA). This plasmid includes the nptII promoter to constitutively drive expression of dntB. The resulting construct was transformed into Burkholderia sp. DNT, giving rise to strain DNT dntB↑.
All strains were grown in 250-ml Erlenmeyer flasks containing 50 ml of M9 minimal medium [59] added with 2.5 ml/l of a trace elements solution [60] and 0.3% (w/v) sodium succinate as the sole C source. Overnight-grown Burkholderia cells were used as the inoculum by diluting them to a starting OD600 of ca. 0.1. Cells were grown in the same culture medium until they reached an OD600 of ca. 0.5, at which point the cell suspension was split into two 125-ml Erlenmeyer flasks, one culture served as a control experiment and was added with DMSO (the vehicle for both DNT and naphthalene), and the other one was amended with either DNT or naphthalene (from concentrated solutions in DMSO) to a final concentration of 0.5 mM. These conditions were used throughout the work. The solubility in H2O of DNT at 20°C is 1.7 mM, and that of naphthalene at 25°C is 0.2 mM; which results in dosages of 1.48 and 0.23 µmol/ml for DNT and naphthalene, respectively, assuming similar solubility values in water and in M9 minimal medium. Flasks were incubated at 170 rpm and 30°C, and samples collected at the times indicated in each case. Abiotic controls were run in parallel to estimate the possible disappearance of the aromatic compounds in the absence of cells. Non-inoculated flasks containing naphthalene (known to be a volatile compound) were incubated during 3 h as explained above, and the remaining substrate content was determined by gas chromatography. The determinations showed that ca. 88% of the initial naphthalene remained at the end of the experiment in these abiotic controls.
Fluorescence-activated cell sorter cytometry analysis was performed using the ROS-sensitive green fluorescent dye H2DCF-DA (Sigma-Aldrich Co.), essentially as described previously [19], [61], [62]. Cells were grown and treated with either DNT or naphthalene as explained above. In the cases indicated, 1.5 mM H2O2 was separately added as a positive control of oxidative stress conditions. After being incubated for 3 h, cells from 1.5–5.0 ml culture were pelleted by centrifugation at 5,000×g during 5 min, washed once with phosphate-buffered saline (PBS, pH = 7.5), and resuspended in PBS to adjust the OD600 to ca. 0.1. This cell suspension was added with 40 µM H2DCF-DA (added from a freshly-prepared 4 mM stock solution in DMSO), and then incubated in the dark for 15 min at room temperature. Flow cytometry analysis of fluorescence levels was performed in a Gallios™ flow cytometer (Beckman Coulter Inc.) equipped with an argon ion laser of 15 mW at 488 nm as the excitation source. The H2DCF-DA fluorescence emission at 525 nm was detected using a 530/30-nm band pass filter array. Size-related forward scatter signals gathered by the cytometer were used by the Cyflogic™ 1.2.1 software (CyFlo Ltd.) to gate fluorescence data from only bacteria in the stream, thus avoiding the mixing of data from bacteria with data from smaller, non-living particles in the suspension. Data for at least 25,000 cells per experimental condition were collected, and the Cyflogic™ 1.2.1 software was used to calculate the geometric mean of fluorescence per bacterial cell in each sample as well to generate histogram plots with a measure of -fluorescence intensity shown on the x-axis and the number of bacteria (i.e., events) counted at the specific fluorescence intensity used for H2DCF-DA detection shown on the y-axis.
Cell suspensions were treated with the compounds indicated before, and 1.5-ml culture aliquots were exposed to 350 µg/ml NBT during 30 min at 30°C in the dark. HCl was then added at 7.5 mM and the resulting mixture was centrifuged at 2,500×g for 15 min at room temperature. Pellets were treated twice with 0.35 ml DMSO to extract reduced NBT, and supernatants were pooled and mixed with 1 ml of 50 mM phosphate buffer (pH = 7.5). The NBT color change was monitored by absorbance at 575 nm and superoxide production was normalized to the protein concentration, measured by the Bradford method [63]. As the cultures had a similar OD600 value at the time of harvesting and processing, the amount of total protein used in these assays was considered to be roughly the same for each sample. Results are expressed as the fold-change in the ratio A575/mg protein among different strains and growth conditions as described elsewhere [64].
G6PDH activity was analyzed by following the rate of NADP+ reduction at 340 nm at 30°C in a reaction mixture (1 ml final volume) containing 50 mM phosphate buffer (pH = 7.5), 10 mM MgSO4, 0.75 mM NADP+, 2 mM glucose-6-phosphate and 15–100 µl of the cell-free extract, obtained as previously described [19]. An extinction coefficient (εNADH) of 6.22 mM−1 cm−1 was used to calculate the specific enzymatic activity, and the protein concentration in cell-free extracts was measured by the Bradford method. One unit of enzymatic activity was defined as the quantity of enzyme that catalyzed the formation of 1 µmol product in 1 min at 30°C.
One hundred microliters of the 10−5, 10−6 and 10−7 dilutions of treated 24-h cultures were plated by quadruplicate onto LB agar, and 100 µl of undiluted cultures were spread (by quadruplicate also) onto LB-rifampicin plates (250 µg/ml). Colony counts were performed after 48 h of incubation. Mutation frequency values are reported as the dimensionless ratio between the number of RifR colonies and the total viable count [65].
A 25-ml aliquot of the cultures treated with either DNT or 1.5 mM H2O2 for 6 h was spun at 4°C at 4,500×g for 15 min. Cells were washed twice in cold 50 mM phosphate buffer (pH = 7.5) and genomic DNA was extracted using the UltraClean™ microbial DNA isolation kit (MoBio Labs Inc.). Genomic DNA was reconstituted in H2O, quantified by horizontal gel electrophoresis and using a NanoVue micro-volume UV/Vis spectrophotometer (GE Healthcare Bio-Sciences AB), and immediately hydrolyzed as follows. A suitable volume of the suspension, containing 1 µg DNA, was boiled at 97°C during 5 min, and immediately placed in an ice bath. Digestion reactions (1 ml) contained 1 µg denatured genomic DNA, 0.75 mM ZnCl2, 10 mM CH3COOK (pH = 5.5), 2.5 U/ml nuclease P1 from Penicillium citrinum, and 1.75 U/ml phosphatase acid from Ipomoea batatas (enzymes purchased from Sigma-Aldrich Co.). Digestions were carried out during 18 h at 37°C, after which the reaction was loaded into a 2-ml Centricon YM-10™ device (Millipore Corp.) and clarified by centrifugal ultrafiltration. Hydrolysis of genomic DNA was assessed by horizontal gel electrophoresis and spectrophotometry. Samples were processed within 6 h after the hydrolysis procedure. The content of 8-oxoG (in equilibrium with the keto form 8-oxo-2′-deoxyguanosine) was assessed by competitive ELISA using a commercial kit (Cayman Chemical Co.). The assay is based on the competition of free 8-oxoG and an 8-oxoG-acetylcholinesterase conjugate for a limited amount of 8-oxoG monoclonal antibody. The free acetylcholinesterase activity was evaluated with the Ellman's reagent [66]. The detection limit for 8-oxoG is 30 pg/ml, and calibration curves were run in parallel in each set of experiments using an authentic 8-oxoG standard.
All reported experiments were independently repeated at least twice (as specified in the figure legends), and the mean value of the corresponding parameter ± SD is presented. Statistical significance was assessed using analysis of variance (ANOVA). For flow cytometry experiments, the median value is reported in box plots along with the 1st and 3rd quartiles, and the statistical significance was evaluated with the Mann-Whitney U test. In all cases, data were considered statistically significant when P<0.05.
|
10.1371/journal.pntd.0007094 | Economic performance and cost-effectiveness of using a DEC-salt social enterprise for eliminating the major neglected tropical disease, lymphatic filariasis | Salt fortified with the drug, diethylcarbamazine (DEC), and introduced into a competitive market has the potential to overcome the obstacles associated with tablet-based Lymphatic Filariasis (LF) elimination programs. Questions remain, however, regarding the economic viability, production capacity, and effectiveness of this strategy as a sustainable means to bring about LF elimination in resource poor settings.
We evaluated the performance and effectiveness of a novel social enterprise-based approach developed and tested in Léogâne, Haiti, as a strategy to sustainably and cost-efficiently distribute DEC-medicated salt into a competitive market at quantities sufficient to bring about the elimination of LF. We undertook a cost-revenue analysis to evaluate the production capability and financial feasibility of the developed DEC salt social enterprise, and a modeling study centered on applying a dynamic mathematical model localized to reflect local LF transmission dynamics to evaluate the cost-effectiveness of using this intervention versus standard annual Mass Drug Administration (MDA) for eliminating LF in Léogâne. We show that the salt enterprise because of its mixed product business strategy may have already reached the production capacity for delivering sufficient quantities of edible DEC-medicated salt to bring about LF transmission in the Léogâne study setting. Due to increasing revenues obtained from the sale of DEC salt over time, expansion of its delivery in the population, and greater cumulative impact on the survival of worms leading to shorter timelines to extinction, this strategy could also represent a significantly more cost-effective option than annual DEC tablet-based MDA for accomplishing LF elimination.
A social enterprise approach can offer an innovative market-based strategy by which edible salt fortified with DEC could be distributed to communities both on a financially sustainable basis and at sufficient quantity to eliminate LF. Deployment of similarly fashioned intervention strategies would improve current efforts to successfully accomplish the goal of LF elimination, particularly in difficult-to-control settings.
| With less than three years remaining for meeting the initial 2020 target set by WHO for accomplishing the global elimination of Lymphatic Filariasis (LF), concerns are emerging regarding the feasibility of meeting this goal using the current tablet-based Mass Drug Administration strategy. Salt fortified with the antifilarial drug, diethylcarbamazine (DEC), could offer an intervention that avoids many of the barriers connected with tablet-based elimination programs. We analyzed the economic performance and cost-effectiveness of a novel DEC-salt social enterprise developed and tested in Léogâne arrondissement, Haiti, as a particularly significant strategy for accomplishing sustainable LF elimination in such complex settings. We show that because of increasing revenue from the sale of the DEC salt over time, expansion of its delivery in the population, and the adverse effect of continuous consumption of the drug on worms, the delivery of DEC through a salt enterprise can represent a significantly more cost-effective option than annual DEC tablet-based MDA for accomplishing LF elimination in settings, like Léogâne. We indicate that development of policy and research into how to deploy similarly-fashioned interventions, or work with the salt industry to increase population use of medicated salt, would improve present efforts to successfully accomplish the elimination of LF.
| Lymphatic filariasis (LF), a mosquito-borne neglected tropical disease (NTD), commonly known as elephantiasis, is one of only parasitic six diseases currently targeted for potential global eradication by 2020 using preventive mass chemotherapy [1–3]. Despite the impressive expansion of a WHO-led elimination program aimed toward the meeting of this goal in all endemic countries since 2000, stakeholders committed to global LF elimination have recognized that the current tablet-based mass drug intervention is resource-intensive, can face significant compliance issues with time, and may be difficult to implement in remote or socio-ecologically complex areas, such as urban and socio-politically unstable settings, hampering foreseen elimination goals [4–8]. These difficulties have heightened interest in investigating the impacts of either approaches aimed at scaling-up treatment strategies or inclusion of preventive activities into drug programs (such as supplemental vector control), or evaluation of novel intervention technologies, that can effectively overcome current barriers in order to accelerate parasite elimination [9–11].
Salt fortified with the anti-filarial drug, diethylcarbamazine (DEC), could offer an intervention that avoids many of the above issues connected with tablet-based elimination programs [12–18]. Indeed, DEC-salt has played a major role in the elimination of LF in a number of pilot and region-wide settings in Africa, Central America, and Asia [5,12–17]. The low dose of DEC (0.1–0.6% [w/w]) used in these studies and programs was well tolerated and rarely associated with adverse reactions. It also has the potential to be more effective than tablet-based Mass Drug Administration (MDA) programs via reduction of the durations of intervention required to interrupt parasite transmission [18]. Moreover, fortified salt can also be provided to a population without developing a dedicated public distribution system, overcoming the need for developing an effective health infrastructure capable of distributing anti-filarial drugs at the high coverages needed for achieving elimination [5,7,8].
Haiti is one of only four countries remaining in the Americas where LF is still endemic [8,19]. MDA using DEC first started in the country under the National Program to Eliminate LF (NPELF) in 2000, and by 2005 had expanded to mass treatment of some 1.6 million people at least once in 24 of the initially endemic 120 communes of the country ([8]. However, following funding, sociopolitical, and natural disaster-based challenges to further scaling up, the program realized full national coverage of all the endemic communes in the country only by 2012 [5,7,8,20]. This delay together with the technical challenge of interrupting transmission in areas of highest prevalence even with high levels of coverage using the suggested five successive years of annual MDA [5,7,8,20], indicate that it is unlikely that NPELF will meet the goal of accomplishing LF elimination in the country by the target year of 2020. In 2006, partly to overcome the above issues, a project was initiated with the collaboration of Congregation de Sainte Croix, the Notre Dame Haiti Program, and the Ministry of Public Health and the Population (MSPP), focused on the local processing and marketing of DEC-mediated salt co-fortified with potassium iodate as an alternative means to facilitate the elimination of LF (and prevention of iodine deficiency disorders) in Haiti [8]. Based on the principle of employing a social enterprise framework for providing goods and services in an entrepreneurial and innovative fashion to solve social problems [21–23], this project purchases both local and imported raw salt which are then cleaned, sized, fortified and packaged for sale in the local market as food-grade co-fortified salt, raising the potential of using a business-based approach to delivering DEC-medicated salt as a sustainable means to accomplish LF elimination in settings, such as Haiti.
Here, our major aim was to examine the economic performance and effectiveness of using the Haitian social enterprise-based framework for producing and marketing DEC-fortified salt as a sustainable, cost-effective, model for achieving the long-term elimination of LF, focusing on Léogâne arrondissement, Haiti. A cost-revenue analysis combined with a mathematical modeling-based evaluation of the cost-effectiveness of the DEC salt social enterprise compared to standard MDA was carried out to undertake this analysis. Specifically, we evaluated the economic performance and social value of the enterprise by assessing: 1) the growth in salt production, costs of resources consumed, and revenues from sales gained to determine break-even points, 2) the impact of the product-mix used for realizing the socially-relevant sale price of the salt, and 3) its cost-effectiveness compared to tablet-based MDA for accomplishing LF elimination in the study setting of Léogâne arrondissement, Haiti. We discuss the results in terms of how using a social enterprise can offer a sustainable and innovative strategy for accomplishing LF elimination in Haiti, and similarly resource-constrained settings, that face both programmatic and social difficulties in delivering long-term tablet-based LF MDAs.
We carried out a cost-revenue analysis to evaluate the production capability and financial feasibility of the developed DEC salt social enterprise via assessment of the relationship between fixed and variable costs versus the revenue received [24–27] and a modeling study centered on applying a dynamic transmission model to evaluate the cost-effectiveness of using this intervention versus standard annual MDA for eliminating LF in Léogâne arrondissement, where the salt enterprise operates [28–31]. The cost-revenue analysis was based on costs and revenue data contained in financial accounts during the production phase of the salt enterprise from 2013 to 2018, while the break-even analysis was carried out over a time horizon that ranged between 2013 to the year when the break-even point was attained. The predicted timelines, in months or years, to LF elimination along with the costs of annual MDA versus the net cost of supplying DEC-fortified salt until elimination was achieved were used to carry out the cost-effectiveness modeling study.
Costs and production of salt
Table 1 summarizes the investment, fixed and variable costs incurred in establishing and operating the Léogâne DEC salt enterprise. These costs are presented for the years between 2013–2018 when production of food-grade fortified salt began (following an experimental phase which addressed technical issues in the fortifying of salt with DEC) along with corresponding data on the quantity of the three different types of salt (industrial, coarse and fine single-fortified, coarse and fine double-fortified) produced and sold. Note that investments occurred periodically during different expansion phases (2013, 2014, and 2017), and were primarily used to acquire capital items either from the US or from within Haiti. These were recorded as fixed assets, and comprised factory items, such as different types of pumps, screens, control systems, hoppers, sealers, storage tanks, generators, and office equipment. Fixed costs, i.e., costs that remain the same whatever the level of output produced or products sold, included operating expenses, while variable costs comprised costs of items which scaled with production volume [24,25]. Examples of components of the latter two cost types incurred are given in Methods.
The data on salt production show that initial output was low (e.g., 10 metric tons of industrial salt and 230 tons of coarse double-fortified salt in 2013, compared to 272, 6, and 470 tons of industrial, single-fortified, and double-fortified salt respectively in 2014). To increase production, a three phase expansion of the project was introduced beginning in the year 2013. Phase 1 focused on installing higher capacity processing equipment (January 2013-March 2014), while Phase 2 aimed to expand processing capacity by adding bulk storage and extension of the brine-washing system (May 2014-March 2015), and Phase 3 further expanded these washing, processing and storage capacities (January 2017-June 2018). These expansion phases meant that both the fixed and variable costs of salt production varied over time (Table 1), but in general, and as expected, as production expanded the variable costs increased proportionately while the fixed costs simply reflected increases in capital investments. By contrast, the production figures show that the manufacture and sale of each type of salt increased steadily, with industrial salt dominating production particularly toward the later years followed by double-fortified (i.e., DEC and iodine medicated) and single-fortified (iodine-medicated only) coarse salts. Given the population preference for coarse edible salt in Léogâne (and Haiti in general), fine salt production and sale lagged behind in volume (Table 1). The figures, however, demonstrate that with expansion in capacity the enterprise was able to achieve significant production volumes for the double-fortified salt by year 2018 (1841 metric tons recorded for the coarse variety).
The total annual costs of salt production (= Investment + Fixed Costs + Variable Costs) and revenues attained from the sale of all three types of salt are given in Table 2. Total annual costs increased as production expanded (Table 1) from US$210,206 in 2013 to US$1,175,000 in 2018 –i.e., approximately five times–but the figures show that revenue increased even faster, up to 17 times that obtained initially in 2013 (US$61,636) to close to cost of production by 2018 (US$1,064,000). The net cash flow [24–27] figures in the Table, which represent the difference between total production cost and total revenue, although being negative for all the years from 2013–2018, capture this increasing revenue returns (as reflected by the declining negative trend in the net cash flow) towards the later years, suggesting that the project is close to achieving break-even or profitability in the near future. To estimate the exact time point when the enterprise is likely to break-even (i.e., the time point when the total program cost is equal to the total revenue), we employed a simple linear model to project forward the total project costs and revenues calculated for 2013–2018. Fig 1(A) shows that if we use the full data on annual costs and revenues obtained for the whole 2013–2018 period, the project will break-even in 2027. However, if we use the data from 2016–2018, when total salt production had reached significant levels (Table 1), to predict the time point at which the break-even point will be achieved, this will occur earlier by year 2022 (Fig 1(B)).
Fig 2 depicts the per ton revenues from sales and costs of producing the three different types of salts for the years 2013–2018. The results show that for all salt types, while initially there was a large difference between the production cost and revenue per ton (i.e., a large negative cash-flow), this difference decreased for each salt category with time. This occurred faster, however, in the case of the industrial and single-fortified salt, such that break-even was achieved by 2018. By contrast, the cash-flow from the production and sale of the double-fortified salt was still negative (lower revenue compared to production costs) at the end of the present study period of 2018. This result shows, first, that the overall break-even estimated in this study (Fig 1) for the project is due to the delay in reaching the break-even year for double-fortified salt. Second, it also highlights how the mixed product strategy of producing different types of salt targeting different market sectors can allow the more profitable products (industrial, single-fortified salt) to subsidize the sale of a product (double-fortified salt) whose cost (approximately US$200 per ton (Fig 2)) needs to be kept competitive with other edible salt sold (retailed at $US265 per ton in Haiti) in the local market.
This was evaluated by analysis of the difference in the variable costs of producing the single-fortified (which included only potassium iodate) versus the double-fortified (both potassium iodate and DEC included) salt types, given that the investment and fixed costs going into manufacturing all salt types were shared equally between each type. The variable costs per ton for producing the single-fortified (coarse and fine) and double-fortified (coarse and fine) salt for two types of bags/bales (25.0-kg bags and 12.5-kg bales) are provided in Table 3 for the years 2014–2018 when both salt products were produced (note that production of single-fortified salt began only in 2014 (Table 1)). Analysis of the difference in the variable costs for producing the single-fortified versus the double-fortified salt indicate that this was consistently about US$70 per metric ton, irrespective of which type—coarse or fine variety—or types of bags/bales were produced (Table 3). Given that the price of DEC, as delivered by Syntholab Chemicals to the project was US$21.60 per kg (James Reimer, personal communication), and DEC salt in this project was fortified with 0.32% DEC by weight or with 3.2kg of DEC/ton, it can be seen that the cost of producing one metric ton of DEC salt works to be US$ 69.12/ton (i.e., 3.2 kg DEC x US$21.60 per kg). This result indicates that the marginal cost, or difference in the variable cost, of adding DEC to single-fortified salt (US$70; Table 3) was simply due to the purchase price of DEC.
Table 4 shows the potential increase in demand for salt in Léogâne arrondissement calculated as a function of changes in population size from 2013 to 2018. It also presents the potential DEC-fortified salt coverage which may be achieved in the setting by increasing sale of the double-fortified salt produced over this period. The annual population size estimates from 2013 onwards were predicted using a growth rate of 1.28% [39], whereas the yearly population demand for edible salt was calculated by assuming that the daily salt consumption per person is 15gm (average of the reported daily per-capita consumption in Haiti of 10gm and 19gm [5,40]). Assuming that the sale of double-fortified salt is widespread in the community (i.e., not targeted towards one segment of the population), coverage of DEC salt in Léogâne can then be roughly estimated simply by dividing the quantity of salt produced over the estimated demand. The results from this calculation, listed in the last column of Table 4, shows that as production increased rapidly from 2013 to 2018, this would increase potential population coverage achieved by sales of the DEC-medicated salt from as low as 8.45% in 2013 to approximately 65% in 2018.
Fig 3 portrays the predicted timelines to LF elimination in Léogâne under each of the MDA and salt interventions investigated. For MDA, the depicted simulations indicate that it would take up to 84 months at 65% coverage, and 60 months at 80% coverage, respectively, to reach the 1% mf threshold. By contrast, the model predictions show that it will take just 1 year (12 months) at 65% coverage, 5 months at 80% coverage, and 3 years or 36 months if actual population coverage (Table 4) is used to reduce the pre-control prevalence to below this threshold via consumption of the traded DEC salt (Fig 3). This highlights the dramatic effect that daily consumption of DEC-medicated salt even at low dosages (0.32% w/w) would have compared to annual intake of higher dosages of DEC (and ALB) as provided by tablet-based MDA for eliminating LF infection in an endemic setting [12,15–18]. Note that the actual DEC salt population coverages (Table 4) used in this analysis assumed that salt supply occurred uniformly and sale was restricted to Léogâne arrondissement only. Any changes in these parameters would mean attaining lower annual population coverages than shown in Table 4; however, a sensitivity analysis using coverage values 15–20% lower than those depicted in the Table did not affect the above timelines significantly.
The comparative costs of carrying out annual MDA versus supplying DEC-medicated salt are shown in Table 5. These show that while the total cost of delivering annual MDA (here fixed at US$0.64 per person, inclusive of the cost of the drug [29,38]), simply scaled with population growth, and will continue to be substantial on a yearly basis until LF elimination is achieved, the net cost of supplying DEC-fortified salt through the social business model will decline dramatically with time as production costs decrease and revenues begin to increase over time (Fig 2). Indeed, projection forward of the net or revenue—production cost data collected during the years 2013–2018 indicates that the total and per capita net costs of DEC salt provided through the present social enterprise could potentially even become zero at the time point (2027) when the project breaks even (Fig 1(A)).
We used the total costs of implementing each strategy until LF elimination (crossing below the 1% mf threshold) is achieved to investigate the cost effectiveness of either strategy. The total costs and effectiveness of using MDA at 65% and 80% coverages and supplying DEC salt at 65%, 80% and the actual coverages given in Table 4, were evaluated and compared via calculations of the average and incremental cost effectiveness ratios [10,41–47]. With respect to DEC salt, we also conducted the analysis using three different net costs of salt production per person per year as recorded in Table 5: (i) average of the net cost of salt produced per person calculated for the years 2018–2020, i.e., US$0.57 per person/year, (ii) average of the net cost of salt per person for the years 2021–2023, i.e., US$0.3 per person/year, and (iii) average of the net cost of salt per person for the years 2025–2027, i.e., US$0.025 per person/year. This was performed to assess the sensitivity of the present results to changes in the steeply declining net cost of salt production over time observed in this study.
The results from this exercise are shown in Table 6, and indicate, principally, that irrespective of coverage, the costs of using MDA are significantly greater than those arising from using the salt strategies, primarily because of its lesser effectiveness as well as higher and stationary unit cost. Among the three salt scenarios, costs for eliminating LF, as expected, declined with decreasing net cost of production over time with scenario three showing the lowest costs. However, for all strategies, while the most effective strategy is to deliver MDA or salt at 80% coverage, the most cost-effective option (in terms of the incremental cost-effectiveness ratio (ICER)) also occurred at this coverage level for both MDA and the DEC salt strategies with incremental costs of either of these options being negative and incremental effects positive over their corresponding next-effective alternative (ICER: US$88,000 per intervention year saved by the 80% MDA strategy compared to an average CER of US$17,333 for carrying out 65% MDA using the cost of implementing annual MDA fixed at US$0.64 per person/year in Léogâne [38], and an additional saving of US$3,876 per intervention year saved by the 80% DEC salt option over the strategy delivering DEC salt at actual recorded coverages when the net cost of delivering medicated salt was fixed at US$0.57 per person (based on durations of interventions in years and total costs given in Table 6)). Note moving from delivering DEC salt at actual recorded coverages to providing salt at 65% coverage results in higher predicted total costs but also a saving of 2 years, and so this strategy is dominated by the strategy that delivers DEC salt at 80% coverage which results in extra reductions in both costs and the time needed to accomplish LF elimination (Table 6). Similar results were also obtained when the other two net costs of producing DEC salt per person were used in carrying out these calculations.
Overall, thus, these results indicate that because of: 1) decreasing net cost of DEC salt production over time, and 2) expansion of its coverage in the population leading to significantly reduced elimination timelines, the delivery of DEC through a salt enterprise may be significantly more cost-effective than annual DEC tablet-based MDA for accomplishing LF elimination in the Léogâne arrondissement setting. It is also to be appreciated that we used a conservative treatment cost of $0.64 per person for modeling the cost-effectiveness of the tablet-based MDA program in this study. This represented a best-case scenario for the MDA program implemented in Léogâne given that the actual drug costs started out higher (US$1.84 over the first 3 MDAs) before approaching the stabilized value used in our analysis as the program became more efficient [29,38]. Indeed, a recent systematic review indicated that MDA program costs can vary substantially between settings, with an average cost that could reach as high as US$1.32 [48]. Use of such values or inclusion of the actual change observed in the per person treatment cost over time in Léogâne in the present analysis would clearly further increase the cost of MDA over that presented in Table 5, which in turn would lead to an even higher cost-effectiveness ratio for MDA compared to those estimated for DEC salt in this study (Table 6). Nonetheless, depending on the cost of producing and delivering DEC-medicated salt, it is readily apparent that the cost savings to a provider of utilizing a social enterprise framework for DEC delivery can be remarkably high. For example, even at the average cost of US$0.907 per person/year (the worst-case scenario using cost data from 2015–2018 (Table 5)), delivery of DEC salt through the current enterprise for eliminating LF in Léogâne is predicted to result in total costs of US$453,386 and US$190,423 for achieving 65% and 80% coverage of the population respectively, which amounts to only 31% and 15% of the corresponding predicted total costs of using annual MDA at these coverages for achieving the same objective in this study location.
In this study, we have undertaken a performance assessment of a novel social enterprise, developed through a collaboration between Haitian and international partners engaged with LF control in the country, as a means to enhance the delivery of anti-filarial drugs to populations through the trading of salt co-fortified with DEC. Although DEC-fortified salt has been used previously in both pilot and region-wide LF intervention programs in a variety of global regions, ranging from Brazil, Tanzania, India and China to effectively control or eliminate LF [5,12–17], it is to be noted that the developed Haiti salt enterprise is the first attempt anywhere in the world to apply the principles of social entrepreneurship for delivering such an intervention. Recent work has highlighted how such social enterprises–that is, a social mission-driven organization that trades in goods or services for a social purpose–are emerging as a potentially effective supply side solution to the provision of cost-efficient public services in response to government failures, business that seek to extract maximal returns on investment, and unstable non-profit organizations [21–23]. In particular, this work has shown how these business entities can solve social problems via their potential to deliver greater responsiveness, efficiency and cost-effectiveness, through an explicit focus on meeting specific social goals while operating with the financial discipline and innovation of a private-sector business [21–23,49,50].
Although there is continuing debate as to how best to evaluate the performance of social enterprises, it is clear that at least two basic components related to the bottom line of these entities require assessment [21,23]. These primarily include: the economic-financial component for measuring overall organizational efficiency, profitability and hence sustainability, and the social effectiveness of the enterprise [21–23]. Here, we have combined the tools of financial accounting and modeling of cost-effectiveness to measure these components in order to present a first analysis of the utility of using the developed Haitian DEC salt enterprise as a sustainable and economically efficient strategy to bring about LF elimination in programmatically difficult-to-control settings, like the arrondissement of Léogâne, Haiti [10,41–47].
Our analysis of the performance of the present salt enterprise for creating social value first focused on the question of capacity to economically produce sufficient amounts of DEC-fortified salt for significantly affecting the elimination of LF in the study setting. The production figures shown in Table 1 indicate, firstly, that while initial production of all types of salt were low during the initial years of operation, by 2018, and just 5 years after processing began in 2013, the enterprise had reached high levels of both total (5,845 metric tons) and DEC-fortified salt (1,841 metric tons) production, respectively (Table 1). Our analysis of the population coverage that the sale of DEC salt could provide demonstrate that the amounts produced could have potentially resulted in a drug coverage rate of 65% by 2018 (Table 4), which our previous modeling study [18] and the present cost-effectiveness exercise (Fig 3 and Table 6) indicate is sufficient to accelerate the achievement of LF elimination in the Léogâne setting. These findings suggest that as the result of the expansion phases carried out through new capital investments (Table 1), the current DEC salt project may have reached the production capacity required to achieve its stated social mission of using a market-based approach for delivering sufficient edible DEC-medicated salt as a means to bring about efficient LF transmission interruption in the present study setting.
Assessment of the economic and financial performance of the salt enterprise carried out in this study using cost-revenue analysis and financial forecasting has provided further insights regarding the organization efforts used to reach economic equilibrium and hence trading viability. This is an important consideration for evaluating the performance of social enterprises because first and foremost these entities are enterprises, and therefore their social goals can be pursued only by ensuring economic and financial fidelity [21–23]. Our major result in this area is providing clarity regarding how the enterprise’s achieved outputs in salt production and sale may affect its potential to reach break-even points (Fig 1). Specifically, we show that given the observed trends in production costs and revenue to 2018 (Table 2), the present salt enterprise may either break-even by 2027 (if we forecast linearly using all the data from 2013 to 2018), or as early as by 2022 (if we use data collected during 2016–2018 after the project had significantly expanded capacity). This result is clearly dependent on assuming that capacity to produce the increased amount of salt to meet either break-even points is available within the enterprise without any further expansion, and demand for the produced salt in all sectors (industrial to edible salt markets) will also expand commensurately. Nonetheless, the finding that it might be possible to reach the break-even point by 2022 (i.e., over the next 4 years) is encouraging, and suggests that the enterprise is likely to be self-sustaining and could become profitable in the very near future. Indeed, analysis of trends in costs of production and revenues gained per ton of each category of salt (Fig 2) indicate that both the industrial and single-fortified salt categories may already have reached their individual break-even points in 2018, and that the delay for achieving break-even status by the social enterprise is primarily due to the lag experienced by the production and sale of the double-fortified salt. Although the per ton production cost of the latter salt declined as significantly over time as the other two salt categories (Fig 2), indicating the achievement of considerable economics of scale, the need to keep the price of the DEC salt below the marginal cost of adding the drug ($70/metric ton (see Table 3)) to compete with untreated local edible salt in the market means that either: 1) the current market price of the double-fortified DEC salt needs to be revised upwards, or 2) further economies of scale need to be found to bring down production costs, or 3) cross-subsidy from the more profitable categories of salt produced will be required in order to continue with the processing of DEC salt in this setting. While on the one hand, such a capacity to use a product mix strategy innovatively as a means to subsidize the marketing of a product for meeting a social need is a feature of using an enterprise model, note that this may be a particular effect of developing markets in settings, such as Léogâne and Haiti in general, where a strong market-based economy is only just evolving. For other LF endemic settings with stronger market economies and established salt industries, the need for such subsides may be significantly lower meaning that the sale of DEC salt could occur at nearer the true marginal cost of production, i.e., at the actual cost of purchasing the drug itself. Note also that our present forecasts do not fully consider the likely impacts of key swing factors that may significantly affect the profitability of the salt social enterprise, such as enforcement of the 2017 law requiring all food salt in Haiti to be fortified, further progress on market segmentation and the resulting product mix, and significant weather events similar to Hurricane Michael in 2016. Positive changes in the first two factors will clearly enhance the enterprise’s ability to break-even faster and hence attain profitability sooner than predicted in this work.
The cost-effectiveness modeling exercise carried out in this study showed that apart from the efficiency of the business model used for achieving economic and financial sustainability, the salt enterprise may also be more cost-effective than the standard tablet-based annual MDA program for accomplishing LF elimination in Léogâne (Table 6). This is because not only will the population coverage that can be potentially attained by sale of the DEC salt (Table 4) be sufficient to make this strategy more effective than annual MDA in reducing the number of years (3 versus 7 years) required for achieving LF elimination (as defined by reducing mf prevalence below the WHO threshold of 1% [2]), the total overall costs involved—due to both decreasing net cost of production and the need for shorter durations of control—for using the salt approach are also significantly lower than those which will be incurred in running the MDA program. Indeed, this greater social impact of using the present social salt enterprise compared to annual MDA was found to be a general outcome, irrespective of the other intervention coverages investigated (i.e., at 65% coverage–the often normal coverage obtained by MDA programs—or at the recommended optimal coverage of 80% [2]) (Table 6). These results add to our recent modeling work, which highlighted how the continuous consumption of the drug, even at low daily per capita dosages, by resulting in a cumulative impact on the survival of worms and mf which is significantly higher than that afforded by the higher-dosed annual MDA treatment, make DEC medicated interventions, even when delivered at moderate population coverages, a markedly potent strategy for interrupting LF transmission [18]. Finally, an intriguing possibility highlighted by the break-even analysis and the cost forecasting results shown in Fig 1 and Table 5 is that using a social enterprise strategy for delivering DEC through marketing of medicated salt could in principle also lead to zero disease elimination cost for a provider (viz. donor or health system) once the social business attains profitability (i.e., return a positive cash-flow). This is an important result, and demonstrates how using a social enterprise that pursues a social goal by production of services and goods whilst respecting economic efficiency may offer an effective, financially sustainable, intervention strategy in settings facing major fiscal, infrastructural and logistical barriers to carrying out tablet-based programs aiming to control or eliminate parasitic infections.
The present performance evaluation primarily focused on internal (labor, capital, income and taxes) and external (goods and services bought outside the company) expenses/resources related to the economic viability of the salt enterprise [21]. However, estimation of the full social value of a sustainable health social enterprise must also consider, apart from the social benefits accruing from reducing disease only, the wider consequences for a community [21]. Benefits here could be via the choice and use of resources that further address the community interest, such as choosing local salt suppliers to favor short supply chains, choosing socially certified suppliers, adopting a regime of decent work conditions and even giving employment to workers coming from disadvantaged backgrounds [21]. Such analysis must also include calculation of the larger social benefit associated with the potential for the double-fortified salt to additionally and simultaneously reduce the impacts of iodine deficiency in the population [5]. Note, additionally, that the present salt enterprise represents the first attempt to build industrial-scale capacity on the island for processing large volumes of salt to meet various local needs, which apart from providing a market for local raw salt producers can also act as means to significantly stabilize the price of salt sold in the local markets. These benefits, however, must be contrasted against potential adverse effects, such as domination of the market by the growing enterprise, requiring an analysis of how best to compensate for such loses. Recent developments in applying Social Returns on Investment (SROI) approaches for comparing the full monetized social costs of a program with the full monetized social benefits of achieving a health outcome (or set of outcomes) may offer a means for undertaking this fuller analysis [22,51].
We have also used rough first calculations of the population coverages that could be obtained with the expansion of salt production in the present cost-effectiveness modelling study. Field studies to assess the actual household coverage achieved through the enterprise will be critical for not only more realistically quantifying its effectiveness for accomplishing LF transmission interruption in a community, but also for identifying better marketing strategies to achieve good population coverage.
In conclusion, we have presented an economic and financial analysis of the Haitian salt social enterprise, which indicates that it may present a sustainable and socially-responsible strategy for aiding the elimination of LF via the marketing of DEC-medicated salt in settings facing fiscal, infrastructural and logistical challenges for delivering tablet-based elimination programs. Results from the break-even projections carried out in this study indicate that the strategy may even have the potential to achieve zero financial costs to a provider once it attains profitability (i.e., results in a positive cash-flow). This study further has shown that the Haitian salt enterprise may have already reached production and sales levels that could result in the coverage of the Léogâne study population at proportions sufficient enough to break LF transmission. Finally, our simulation-based cost-effectiveness study has indicated that because of: 1) increasing revenue from the sale of the DEC salt obtained over time, 2) expansion of its delivery in the population, and 3) the effect of continuous consumption of the drug, even at low daily per capita dosages, leading to a cumulative impact on the survival of worms and mf higher than that afforded by the higher-dosed annual MDA treatment [18], the delivery of DEC through the present Haiti salt enterprise may represent a dramatically more cost-effective option than annual DEC tablet-based MDA for accomplishing LF elimination. While these are encouraging first results and highlight both the economic viability and social effectiveness of using a salt enterprise in the fight against LF, it is clear that efforts to more fully quantify the social value and strategies for developing similar salt social enterprises elsewhere in other endemic settings with different market structures than those of Haiti are now required if the comparative or joint utility of the approach among the current arsenal of LF intervention strategies is to be fully appraised and understood. We note that the means by which the global iodization of edible salt has been accomplished successfully over the past two decades may offer a particularly apt model for building and sustaining the present intervention globally, and suggest that similar tactics used in that program based on introducing DEC medication into prevailing salt production and distribution systems, collaboration with the national and regional salt industries, and engagement with the government sector, civic society and the general public [52], could also make the universal deployment of DEC-medicated salt eminently possible. With less than three years remaining for meeting the initial 2020 target set by WHO for accomplishing the global elimination of LF, the present results indicate that these appraisals and development of policies and strategies for delivery of DEC-salt, either via deployment of similarly-fashioned salt enterprises, such as the present, or through mobilization of existing salt industries, perhaps along with health system-led MDA and vector-control programs, in socially-challenging environments, like Haiti, would improve our current efforts for meeting this laudable but exacting goal successfully.
|
10.1371/journal.pgen.1002646 | Sequence-Specific Targeting of Dosage Compensation in Drosophila Favors an Active Chromatin Context | The Drosophila MSL complex mediates dosage compensation by increasing transcription of the single X chromosome in males approximately two-fold. This is accomplished through recognition of the X chromosome and subsequent acetylation of histone H4K16 on X-linked genes. Initial binding to the X is thought to occur at “entry sites” that contain a consensus sequence motif (“MSL recognition element” or MRE). However, this motif is only ∼2 fold enriched on X, and only a fraction of the motifs on X are initially targeted. Here we ask whether chromatin context could distinguish between utilized and non-utilized copies of the motif, by comparing their relative enrichment for histone modifications and chromosomal proteins mapped in the modENCODE project. Through a comparative analysis of the chromatin features in male S2 cells (which contain MSL complex) and female Kc cells (which lack the complex), we find that the presence of active chromatin modifications, together with an elevated local GC content in the surrounding sequences, has strong predictive value for functional MSL entry sites, independent of MSL binding. We tested these sites for function in Kc cells by RNAi knockdown of Sxl, resulting in induction of MSL complex. We show that ectopic MSL expression in Kc cells leads to H4K16 acetylation around these sites and a relative increase in X chromosome transcription. Collectively, our results support a model in which a pre-existing active chromatin environment, coincident with H3K36me3, contributes to MSL entry site selection. The consequences of MSL targeting of the male X chromosome include increase in nucleosome lability, enrichment for H4K16 acetylation and JIL-1 kinase, and depletion of linker histone H1 on active X-linked genes. Our analysis can serve as a model for identifying chromatin and local sequence features that may contribute to selection of functional protein binding sites in the genome.
| The genomes of complex organisms encompass hundreds of millions of base pairs of DNA, and regulatory molecules must distinguish specific targets within this vast landscape. In general, regulatory factors find target genes through sequence-specific interactions with the underlying DNA. However, sequence-specific factors typically bind only a fraction of the candidate genomic regions containing their specific target sequence motif. Here we identify potential roles for chromatin environment and flanking sequence composition in helping regulatory factors find their appropriate binding sites, using targeting of the Drosophila dosage compensation complex as a model. The initial stage of dosage compensation involves binding of the Male Specific Lethal (MSL) complex to a sequence motif called the MSL recognition element [1]. Using data from a large chromatin mapping effort (the modENCODE project), we successfully identify an active chromatin environment as predictive of selective MRE binding by the MSL complex. Our study provides a framework for using genome-wide datasets to analyze and predict functional protein–DNA binding site selection.
| In Drosophila, Male Specific Lethal (MSL) complex binds to the single male X chromosome to increase transcription approximately two-fold, in order to equalize the output of both female X chromosomes [2]–[4]. We have proposed that MSL complex locates its target binding sites using a two-step mechanism [5]. First, the complex distinguishes X from autosomes by binding a subset of 200–300 sites on X known as “chromatin entry sites” (CES) [6]–[8] or “high affinity sites” (HAS) [9], [10]. Recognition of CES is a sequence-dependent step, as these sites share a GA-rich motif [8], [10] designated the “MSL recognition element”, or MRE, whose function has been demonstrated by site-directed mutagenesis [8]. In contrast, the second targeting step lacks a consensus sequence but is strongly linked to transcription [11]–[14], with the complex locating active genes on the same chromosome [15].
CES were first identified in msl3 mutant embryos, in which the initial, sequence-specific step of MSL binding occurs but the second, sequence-independent step does not [7], [8]. The MRE sequence motif was discovered based on the first 150 mapped CES (Figure 1A). CES function was tested in transgenes for the ability to attract MSL complex to autosomal insertion sites, and found to be dependent upon the intact MRE motif [8]. 150 is likely an underestimate of the total number of CES and functional MREs on X, as subsequent analysis of high occupancy MSL binding sites in wild type cells has revealed 309 peaks containing 379 MREs [8]. However, a conservative set of 150 should be sufficient to test for predictive features.
The MRE motif is only modestly enriched on the X chromosome compared to the autosomes. At a stringency where 137 of 150 CES contain the consensus motif (p-value of 10−5), there is a 1.8 fold higher MRE density on X compared to autosomes (on average 1 per 6 Kb on X, and 1 per 11 Kb on autosomes; Figure S1A), [8]. These average densities correspond to 12,481 total MREs in the genome, of which only 1 in 91 correspond to the set of CES considered here. Even if we restrict our attention to chromosome X, only 1 in 28 MREs maps to the CES set. Therefore, a key question is how functional MREs within CES are somehow recognized amongst a vast excess of un-utilized sites. That the MSL complex targets only a fraction of potential MRE sites for initial binding is a characteristic it shares with many sequence-specific binding factors whose predicted target motifs are often in vast excess to the sites actually utilized [16], [17]. Here, we investigate whether chromatin features influence binding site selection, using the MSL complex and a large compendium of genome-wide ChIP-chip profiles generated by the NHGRI modENCODE project as a model [18]. Our results support a model in which active chromatin composition and intrinsic GC content help define the initial binding sites of the MSL complex.
To search for chromatin features that can distinguish functional MREs from those that do not recruit MSL complex (i.e., non-functional MREs), we defined five classes of MREs in the Drosophila genome. The first set consists of 137 MREs that were experimentally defined by MSL complex binding [8], as discussed above. We called this set “Functional MREs”. The remaining four sets of 150 sequences each consist of the MREs that have the best consensus motif matches on either X or a control autosomal arm (2L) (“Best on X” and “Best on 2L” respectively), and 150 MREs chosen at random from either X or 2L (“Random X” and “Random 2L” respectively). We note that, in general, functional MREs display a broad range of motif binding specificity rather than being the best matches to the MRE consensus motif (Figure S1B). The result of this analysis is not affected by the choice of chromosome arm or the choice of random MREs (data not shown).
In each of the five classes of MREs, their locations are distributed along the length of the chromosome arm with no obvious clustering (Figure S1C). Within each set of sequences, we calculated the average profiles of various chromatin marks mapped by the modENCODE Drosophila Chromatin Consortium using genome-wide ChIP-chip. The average ChIP enrichment profiles for 10 kb regions centered around the motif in two male cell lines (S2 and BG3) are shown in Figure 1B–1C. We found that a number of chromatin marks associated with active transcription are strikingly enriched near functional MREs in CES, and not in the best or random MRE classes on X or autosomes. These include RNA pol II, H3K36me3, H3K9ac, and H2B-ubiquitin. In addition, functional sites are relatively depleted for core histone H4 and linker histone H1. Consistent binding of the sequence-dependent GAGA factor [19] across categories serves as an important control to demonstrate that GA-rich elements are broadly represented in each group of MREs. Another notable feature of these profiles is the enhancement of H4K16ac on the X chromosome as a whole [20]–[22], with additional enrichment of H4K16ac on true CES (Figure 1B–1C). Since enrichment of H4K16 acetylation is a known consequence of MSL targeting, we proceeded to ask whether the observed difference in chromatin context for other chromatin marks at CES might simply be a consequence of MSL binding rather than a contributing factor.
To test whether the observed difference in chromatin context at CES is a consequence of MSL binding rather than a contributing factor, we examined the profiles of the same subset of chromatin marks in female Kc cells. Interestingly, we found that the set of marks that correlates with MRE function in male S2 cells were likewise informative in female Kc cells, in the absence of MSL complex (Figure 1D). The enrichment of H4K16 acetylation on the male X is notably less pronounced in female Kc cells lacking MSL complex. Still, this mark is enriched over functional MREs, consistent with previous observations of MSL-independent H4K16 acetylation at active genes on all chromosomes in males and females [21], [22]. Most strikingly, H3K36me3 and JIL-1, are enriched in functional MREs in both male and female cells, suggesting that these marks are independent of MSL binding at CES.
We also examined MREs for intrinsic sequence composition to search for correlations with function (Figure 1E). Surprisingly, we found a marked elevation of GC content in the 10 Kb flanking region surrounding functional MREs, coupled with a decrease in GC content in the 1 Kb nearest the functional MREs. Taken together, our results support a model in which local sequence characteristics and the active chromatin context of functional MREs may facilitate their initial selection.
Next, we wanted to directly test the potential of MREs in an active chromatin environment in females to recruit MSL complex. The key female sex determination protein, SXL, represses dosage compensation by inhibiting MSL2 translation [23], [24]. Loss of SXL results in the expression, stabilization, and targeting of the MSL complex in female cells [25]. Therefore, we depleted SXL by RNA interference (RNAi) in Kc cells [26]. Upon treatment, we observed a general, MSL2-dependent increase in transcription from the female X chromosomes, consistent with a partial induction of dosage compensation (Figure 2A–2B). The increase in X-linked gene expression did not reach the maximum theoretical amount for perfect compensation (log22 = 1), consistent with observations that MSL-independent ploidy effects can also contribute to overall compensation [27], [28].
The induction of MSL complex in Kc cells allowed us to ask whether ectopic MSL complex in female cells recognized the same functional subset of MREs on X as in male cells, by examining the distribution of H4K16 acetylation as a mark of MSL function. We found that Sxl RNAi induces high levels of this modification preferentially at the same MREs that are functional in males (Figure 2C), supporting the idea that MSL complex recognizes MREs in an active chromatin context. Our findings with MSL complex parallel recent results for the heat shock transcription factor HSF [29], suggesting that sequence-specific DNA binding factors may generally utilize chromatin context to facilitate selective targeting within a complex eukaryotic genome.
Since specific chromatin marks are enriched near MREs that are utilized compared with those that are not, we next asked whether these marks provide enough information, either individually or in combination, to explain the MSL entry site binding pattern on X. To address this question, we systematically investigated the predictive power of the chromatin marks for functional MREs in Kc cells, which have the potential for MSL targeting but do not express the MSL complex. We asked whether we could build a simple prediction model based on individual or a combinations of chromatin features in Kc cells that would distinguish functional MREs from non-functional ones in male S2 and BG3 cells, where MSL complex is expressed.
We first tested whether individual chromatin features could discriminate functional MREs from non-functional ones. A chromatin feature is defined by its average ChIP enrichment within the 10 kb region surrounding each MRE. In addition, we defined two features to represent the average GC content near the MRE (center 1 kb) or in its flanking regions (10 kb excluding the center 1 kb). To test whether individual or combinations of features could distinguish the functional MREs from the non-functional ones, we used support vector machine (SVM), a classification algorithm demonstrated to have excellent performance in a wide range of problems [30], [31]. Briefly, the set of ChIP enrichment at each MRE is treated as a feature vector of that MRE. Given a set of training samples, SVM calculates an optimal hyperplane that can separate non-functional MREs from functional MREs in the feature space. Here we used a SVM with a radial basis kernel that is implemented in the R package e1071 (See Methods and Materials). To accurately estimate predictive power, and to avoid the potential bias due to using the same set of CES and non-functional MRE genomic locations in S2, BG3 and Kc cell lines, we evaluated the predictive power of a feature using 10-fold cross-validation. In this scheme, we withhold a random 10% of the MREs, build a model based on the remaining 90%, measure how well the model predicted the functionality of the withheld MREs, and repeat this process multiple times to obtain the average performance. The cross-validation result is presented using a standard measure called the Area Under Curve (AUC) of Receiver Operating Characteristic [32] curve. The ROC curve (examples are shown in Figure 3C) quantifies the sensitivity and specificity of classification by estimating true and false positive rates over all threshold values, and the AUC summarizes the curve with a single number. A random predictor receives an AUC of 0.5, and a perfect predictor achieves an AUC of 1 [32].
By comparing the AUC of each individual chromatin feature in Kc, S2 and BG3 cells, we observed that many active chromatin marks could distinguish functional from non-functional MREs (Figure 3A). Among all the features tested, H3K36me3 was the best predictor in all three cell lines (mean AUC of 0.884), followed by JIL-1 (mean AUC of 0.864). Interestingly, we had previously speculated that H3K36me3 might be involved, based on MSL affinity for this mark in active genes [8], [12], [14]. Since H3K36me3 and JIL-1 are enriched in Kc cells at predicted CES even without MSL binding, they could prime functional MREs for sequence-specific MSL binding, or be coincident with true causative factors. Nearly all putative CES in Kc cells are embedded in a chromatin environment enriched for H3K36me3 (Figure 3B), however, we previously determined that when H3K36me3 is depleted, there are still enough features for the MSL complex to distinguish MREs [12]. Since no single feature may be sufficient to drive MSL recognition, we next asked whether combinations of marks and local sequence composition might further improve predictive power for CES.
On average, the GC content of CES is similar to the random or best MREs in the 1 Kb of sequence immediately surrounding the motif, but the GC content consistently rises in flanking sequences, to produce a distinctive average profile (Figure 1E). Examination of the individual heatmaps confirms that this is a broadly consistent characteristic of CES (Figure S2). We found that the relative GC content in D. melanogaster is elevated in genes compared to intergenic sequence (44% vs. 41%) [33] so it is not surprising to see this characteristic in conjunction with the active gene clusters where CES are found. However, the distinctive shape of the profile, with low GC immediately surrounding the CES, is unexpected and not seen with autosomal MREs, even when associated with H3K36me3 and thus presumably analogous active gene clusters (Figure S2). The significance of this intrinsic feature clearly merits future experimental analysis. However, GC profile alone does not appear to provide enough information to predict functional MREs with high accuracy (Figure 3C).
To search for combinatorial marks that might distinguish functional MREs, we compared the predictive power of the SVM generated using every possible combination of features in our dataset (Figure S3A–S3B). We found that the best individual features (H3K36me3 and JIL-1) performed very well, similarly to the best combinations of features when SVM trained using Kc data, and tested on S2 or BG3 data (Figure 3C). There are many combinations of features that predict functional MREs with high accuracy. In general, the best performing combinations included the following: (1) H3K36me3 or JIL-1; (2) H2B-ubiq or H3K9ac; (3) a core histone; and (4) GC content (Figure S3A–S3B). The excellent performance of this combination is consistent with the identification of these factors as core features by feature correlation analysis (Figure S3C).
The cross-validation results are summarized in the ROC plots in Figure 3C and Figure S3D. Although not perfect, the best combination separates the functional and non-functional MREs with high accuracy (mean AUC = 0.931), as visualized in principal component space in Figure 3D. Each MRE can be considered as a point in multi-dimensions, with each axis as the enrichment level for a mark; to show the data in two-dimensions, we define new axes, called principal components, which are combinations of the original variables satisfying certain desirable properties. We can indeed observe a good separation of the functional from non-functional MREs in this view (Figure 3D).
The analyses presented so far focused on only a subset of clearly functional and non-functional MREs. This allowed us to effectively identify chromatin features that can distinguish functional MREs from non-functional ones, and to build a predictive SVM model for functional MREs. Here we extend our analysis to test whether our model could accurately select additional functional MREs genome-wide. We trained an SVM model with Kc cell chromatin features and then tested its ability to eliminate non-functional MREs on autosomes and chromosome X using S2 data. The SVM algorithm selects the decision threshold that optimally separates functional from non-functional MREs, as confirmed in Figure S3E, and the overall AUC of this prediction is about 0.84 (Figure S3F). We specifically tested different individual and combinations of features. Using the best combination of features, our SVM model can eliminate over 75% of candidate MRE sites on X, and ∼85% of candidate sites on autosomes, while retaining almost all of the functional MREs on the X chromosome (up to 94%) (Table S2). Almost 10,000 non-functional sites are eliminated using our model, with retention of approximately 1600 MREs genome-wide. Approximately half (763) of those remaining MREs map to the X chromosome but are not included in the conservative set of 137 CES.
We suspect that the large number of remaining MREs on the X chromosome indicates that there may be more true MSL binding sites than the set of 137 CES we used in this study. Therefore, we asked how many of these additional 763 MREs overlap with previously mapped MSL binding sites identified by MSL3 ChIP-seq [8]. Of the 763 SVM-predicted functional MREs on X, 503 overlapped with an MSL binding site (Figure 3E). This suggests that the actual false positive rate on the X chromosome may be as low as 260/3343 = 7.8%. This is slightly lower than the false positive rates for the autosomes (895/8144 = 11%) (Table S2), likely due to the fact that even though most strong peaks identified in the MSL3 ChIP-seq data with stringent criteria are CES [8], some may be sites to which MSL complex spreads.
In addition to enrichment for marks associated with active genes, we also saw relative depletion of histone H1 over functional MREs (Figure 1), and, from previous work, depletion of H3 and thus presumably nucleosomes themselves [8], [10]. Therefore, we examined modENCODE data from S2 and Kc cells on the release of nucleosomes following salt extraction to determine whether functional MREs are packaged into more labile chromatin when compared to non-utilized MREs [34]. We found that in male S2 cells functional MREs are depleted for histones in general, and enriched for nucleosomes that are extracted in low salt or remain in the pellet after high salt extraction (Figure 4), both fractions that were previously characterized as enriched in regulatory regions [35]. In addition, the two sets of non-functional MREs on the X chromosome appeared to have a milder, but discernable increase in nucleosome lability when compared to the MREs on autosomes, consistent with the observation that X chromosome in male S2 cells generally adopts a more open chromatin conformation [36].
In contrast, CES MREs in female Kc cells exhibit a modest average decrease in nucleosomal occupancy (Figure 4), but intriguingly, this appears to be a difference at a subset of sites rather than the entire set (Figure S4). Notably, the entire X chromosome in Kc cells does not appear to be packaged in a more open chromatin state compared to autosomes (Figure 4). We conclude that strong nucleosome depletion and a more open chromatin conformation on X are mainly consequences rather than causes of MSL binding.
Once MSL complex identifies the male X, we have proposed that it spreads to affect the active genes on the chromosome as a whole [5]. In addition to the core MSL complex consisting of proteins that are essential in males but not females (MSL1, MSL2, MSL3, MOF, and MLE), the JIL-1 kinase is known to be enriched on the male X [37]. JIL-1 is essential in both males and females and binds interband regions on all chromosomes in both sexes [38]. Since we observed that JIL-1 is enriched at functional MREs (Figure 1 and Figure 3A), we tested whether it is also associated with active gene bodies. We constructed average scaled profiles of JIL-1 binding (meta-gene profiles) of all genes greater than 2 kb in length based on the FlyBase dm3 gene annotation [39]. Gene expression in the three cell lines was determined by RNA-seq data (Figure S5) [34]. We found that JIL-1 binds active gene bodies, with a bias towards 3′ ends, on all chromosome arms (Figure 5A). On the male X, this pattern is increased in its intensity above the level on autosomes and correlates strongly with binding of the MSL complex since the increased occupancy is not seen on female X chromosomes (Figure 5B). These results are consistent with previous polytene immunostaining [37] and ChIP analyses [40].
In contrast to JIL-1, the linker histone H1 shows depletion on active genes (Figure 5C). Interestingly, in male cells this depletion is more prominent on X-linked gene bodies than on autosomal gene bodies, whereas X and autosomes show no obvious difference in female cells (Figure 5D), consistent with previous polytene immunostaining results [41]. These two results further underscore the distinct character of the Drosophila male X chromosome once MSL complex is bound to active genes.
When examining our modENCODE data, we noticed that many additional data sets are slightly enriched on the entire X chromosome compared to autosomes in male S2 cells, but not in female Kc cells (Figure S6, and Table S3). X-chromosome-wide enrichment of nucleoporins, Megator and Nup153 have also been observed in male S2 cells, but not in female Kc cells [42]. These results are unlikely to simply reflect increased access of antibodies to X chromatin in the ChIP procedure, because histone H1 shows the opposite trend. Therefore, these enrichments may be related to dosage compensation of the male X, either directly, as is the case for H4K16ac, JIL-1, and possibly RNA pol II, or indirectly, as a consequence of the more open chromatin environment created by the dosage compensation mechanism.
MSL-dependent dosage compensation is thought to be an organism-wide phenomenon in Drosophila males, excluded from the germline [43], but otherwise not restricted to particular tissues or developmental time points. Classical mitotic recombination experiments support a requirement for MSL proteins throughout development in dividing tissues [44]. In addition, comparison of gene expression in males and females demonstrates that the vast majority of X-linked genes are up-regulated in males [45]. However, stable MSL binding appears to be more restricted than its functional consequence, H4K16 acetylation [22], so we wondered whether binding favored particular types of genes. To examine this, we plotted the chromatin marks and MSL binding along active X linked genes in S2 cells (>2 Kb long), asking if clustering might define genes of particular structure, expression level, or gene ontogeny category (Figure 6). We found that MSL1 and H4K16 acetylation cover the vast majority (86%) of active X linked genes. Apparent exceptions (green cluster) are interesting as those genes also lack H3K36me3 (Figure 6) and H2B-ubiquitin (data not shown), two prominent marks of transcription [18].
While MSL1 and H4K16ac are associated with the bodies of virtually all active X linked genes in male S2 cells, the MOF H4K16 histone acetyltransferase was notably absent from a subset (dark brown cluster at bottom of Figure 6). Most of these genes still showed limited MOF enrichment around the promoter region, but not within the gene bodies as is characteristic of MSL complex. We were unable to identify any feature, such as relative expression level or chromatin state, that would distinguish these genes from the genes that showed MOF enrichment.
Since functional MSL chromatin entry sites are also preferentially associated with active gene marks, we asked where CES map relative to the set of clustered genes on X, indicating the location of CES by red dots on the gene structures in Figure 6 (H3K36me3 column). Our results demonstrate that CES are located at variable positions relative to active genes, with a bias towards their 3′-ends. Interestingly, the subset of genes lacking strong MOF binding within gene bodies also lack nearby entry sites, in agreement with the observation that MSL binding is strongest in the close vicinity of mapped chromatin entry sites [10], [46].
Genetic, genomic, and biochemical analyses in eukaryotes have revealed that DNA binding motifs alone are insufficient to explain the selective occupancy or specificity of regulatory factor function [16], [17]. The number of predicted binding sites is often vastly greater than the number of sites actually utilized. Therefore, a very important question in transcriptional regulation is how to identify additional parameters that must govern accurate binding site selection.
In this study, we considered the roles of chromatin environment and flanking sequence composition in selection of functional binding sites by a sequence-specific protein complex. It is generally not clear whether the chromatin features that are often observed at the binding sites of proteins contribute directly to binding selectivity or are simply a consequence of binding. In the dosage compensation system of the X chromosome in Drosophila, we had a unique opportunity to address this question because we can compare the chromatin environment of MSL binding sites in female cells, in the absence of the complex, to male cells, where the functional sites are bound. We also utilized binding data from an RNAi experiment in which we knocked down a component of the sex determination pathway in females to induce dosage compensation. Our bioinformatic analysis of a large number of profiles from the modENCODE project suggests that a pre-existing active chromatin context plays a critical role in establishing the initial binding of the MSL complex on the X. We also made the surprising discovery that GC content in the DNA surrounding functional binding sites has a characteristic profile.
In summary, our results strongly support a model in which an active chromatin composition helps define the initial entry sites selected by the MSL complex (Figure 7). Functional MSL binding results in increased lability of local nucleosomal composition, and H4K16 acetylation and JIL-1 binding along the bodies of virtually all active X-linked genes. Our work provides key insights into the order of events leading to dosage compensation in Drosophila, and can also serve as a model for using genome-wide data sets to understand how sequence-specific factors find their ultimate targets.
The majority of the ChIP-chip data are from the modENOCODE project [18]. Genomic DNA Tiling Arrays v2.0 (Affymetrix) were used to hybridize both ChIP and input DNA. We obtained the log-intensity ratio values (M-values) for all perfect match (PM) probes: M = log2(ChIP intensity)−log2(input intensity), and performed a whole-genome baseline shift so that the mean of M in each microarray is equal to 0. The smoothed log intensity ratios were calculated using LOWESS with a smoothing span corresponding to 500 bp, combining normalized data from two replicate experiments. All data are publicly accessible online through the modENCODE project (URL listed in Table S1). Data analysis was performed in R statistical programming environment (http://www.r-project.org). For the visualization of the heatmap (e.g., Figure 1), the +/−5 kb region surrounding each MRE was separated into non-overlapping bins of 200 bp. The smoothed probe value within each bin is averaged to obtain the enrichment value for that bin.
The GC content around each nucleotide is defined as the proportion of G or C in the closest 101 bp (ie, the target nucleotide, 50 bp upstream and 50 bp downstream). Similar to the ChIP-chip data, we separated the +/−5 kb region surrounding each MRE into non-overlapping bins of 200 bp. The average GC content in each bin represents the average of the GC content of the 200 bp within that bin.
We generated double-stranded RNA (dsRNA) to target GFP (negative control) or Sxl transcripts (amplicons designed by the Drosophila RNAi Screening Center (www.flyrnai.org), as described previously [22]. The following primer sets were utilized to amplify PCR products to template dsRNA synthesis:
GFP
F 5′-TAATACGACTCACTATAGGGAGAGGTGAGCAAGGGCGAGGAGCT-3′
R 5′-TAATACGACTCACTATAGGGAGATCTTGAAGTTCACCTTGATGCCG-3′
Sxl (DRSC21490)
F 5′-TAATACGACTCACTATAGGGAGAGATCACAGCCGCTGTCC-3′
R 5′-TAATACGACTCACTATAGGGAGATACCGAATTAAGAGCAAATAATAA-3′
Sxl (DRSC28896)
F 5′-TAATACGACTCACTATAGGGAGACCCTATTCAGAGCCATTGGA-3′
R 5′-TAATACGACTCACTATAGGGAGAGTTATGGTACGCGGCAGATT-3′
For expression analyses, GFP and Sxl DRSC21490 RNAi was performed in Kc cells using 6-well plates as described [47].
For ChIP-chip, RNAi using GFP, Sxl DRSC 21490, and Sxl DRSC28896 amplicons in Kc cells was scaled up to T225 flasks and chromatin preparation, and H4K16ac ChIP was performed using anti-H4K16ac antibody (Millipore, 07-329) and custom Nimblegen tiling arrays as described [22].
Kc GFP H4K16ac ChIP-chip datasets were published previously [22]: Gene Expression Omnibus accession numbers: GSM372470 (replicate #1) and GSM372471 (replicate #2).
We used 10-fold cross-validation to estimate the predictive power of a classification model based on a training dataset (e.g., chromatin feature in Kc cells) and a test dataset (e.g., chromatin feature in S2 cells). Each sample in a dataset is an MRE, a feature is a histone modification (e.g., H3K36me3) or a chromatin binding protein (RNA Pol II), and the label for each sample is either “Functional MRE” (positive class) or “Non-functional MRE” (negative class). The aim is to train a prediction model that can distinguish functional from non-functional MREs based on the chromatin features. In a 10-fold cross-validation, the training data are randomly divided into 10 equal-sized portions in which the same proportion of positive and negative samples are preserved in each portion. In each of the 10 iterations, the data from nine portions are used to train a predictive model, while the remaining one portion is used to test the performance of the prediction model. Performance is measured by true positive rate (sensitivity) and false positive rate (1-specificity). The tradeoff between true and false positive rates are often represented by a receiver operation characteristic curve, and the Area Under the ROC Curve (AUC) is a measure of the prediction accuracy that takes into account both sensitivity and specificity of the prediction model. A random predictor receives an AUC of 0.5, and a perfect predictor achieves an AUC of 1. Using 10-fold cross-validation, an AUC is calculated for each fold (one iteration), and the mean and standard deviation of the 10 AUC values are recorded. Calculation and visualization of the ROC curves were performed by the ROCR package [32].
SVM is a supervised classification algorithm that separates two classes of data based on a set of features [30], [31]. In this study, the set of ChIP enrichment at each MRE is treated as a feature vector of that MRE. Given a set of feature vectors from functional MREs and a set of feature vector from non-functional MREs, the SVM algorithm calculates an optimally-separating hyperplane by maximizing the distance (called margin) between the hyperplane and the nearest points from the two classes. This hyperplane effectively divides the space of feature vectors into two regions, one for each class, with the idea that the larger margin lowers the error of the classifier. To make a prediction, the feature vector corresponding to a MRE from the test set is compared to this hyperplane to determine on which side of the separating boundary this sample is located. We used the SVM implementation in the R package e1071, which is optimized for the radial basis function kernel and uses an Sequential Minimal Optimization-type algorithm [48] using default parameters for training the SVM.
We used the MSL3-TAP ChIP-seq data from Alekseyenko et al [8]. The raw sequence reads were aligned to the Drosophila melanogaster genome assembly dm3 using bowtie with default options [49]. We only allowed uniquely mapped reads to be reported. This procedure resulted in 2.8 and 2.4 million mapped reads for ChIP and input DNA samples. The aligned reads were then analyzed with SPP [50] to identify ChIP-enriched regions (FDR threshold of 0.05).
Gene expression level estimates in S2, BG3 and Kc cells were obtained from the modENCODE project [34]. The expression of each gene is quantified in terms of RPKM (reads per million reads per kilobase). The distribution of gene expression in each cell line was assessed and a cut-off of RPKM = 3 was determined to be a good threshold to separate active vs. inactive genes (Figure S5A). This definition of active vs. inactive genes was used in the construction of meta-gene profiles.
We used the gene annotation from FlyBase [39] to define transcription start and end sites (TSS and TES respectively). We only included genes with a minimum length of 2 kb (7,231 of 15,186 genes) to exclude short genes from our analysis. The ChIP enrichment in the 2 kb region centered on the TSS and TES, as well as the ChIP enrichment within the gene body scaled to 1 kb, were calculated and averaged for the active and inactive genes in X and autosomes. The definitions of active vs. inactive genes were defined by RNA-seq data.
All ChIP-chip and RNA-seq data are available from modENCODE, and the URL for individual datasets is listed in Table S1. The ChIP-chip and microarray gene expression data pertaining to the Sxl RNAi experiments are accessible from GEO (Accession number: GSE34859).
|
10.1371/journal.ppat.1002719 | Cellular Levels and Binding of c-di-GMP Control Subcellular Localization and Activity of the Vibrio cholerae Transcriptional Regulator VpsT | The second messenger, cyclic diguanylate (c-di-GMP), regulates diverse cellular processes in bacteria. C-di-GMP is produced by diguanylate cyclases (DGCs), degraded by phosphodiesterases (PDEs), and receptors couple c-di-GMP production to cellular responses. In many bacteria, including Vibrio cholerae, multiple DGCs and PDEs contribute to c-di-GMP signaling, and it is currently unclear whether the compartmentalization of c-di-GMP signaling components is required to mediate c-di-GMP signal transduction. In this study we show that the transcriptional regulator, VpsT, requires c-di-GMP binding for subcellular localization and activity. Only the additive deletion of five DGCs markedly decreases the localization of VpsT, while single deletions of each DGC do not impact VpsT localization. Moreover, mutations in residues required for c-di-GMP binding, c-di-GMP-stabilized dimerization and DNA binding of VpsT abrogate wild type localization and activity. VpsT does not co-localize or interact with DGCs suggesting that c-di-GMP from these DGCs diffuses to VpsT, supporting a model in which c-di-GMP acts at a distance. Furthermore, VpsT localization in a heterologous host, Escherichia coli, requires a catalytically active DGC and is enhanced by the presence of VpsT-target sequences. Our data show that c-di-GMP signaling can be executed through an additive cellular c-di-GMP level from multiple DGCs affecting the localization and activity of a c-di-GMP receptor and furthers our understanding of the mechanisms of second messenger signaling.
| C-di-GMP is a ubiquitous intracellular signaling molecule in bacteria and controls diverse cellular processes including biofilm formation, motility and virulence. The genomes of many bacteria often contain numerous genes encoding proteins predicted to produce and degrade c-di-GMP. However, it is currently unclear how a bacterial cell orchestrates multiple c-di-GMP signaling proteins to elicit a specific cellular response. The microbial pathogen, Vibrio cholerae, contains a multitude of c-di-GMP proteins and c-di-GMP signaling is likely important for the bacterium to cause the deadly diarrheal disease called cholera. In the present study, we define the requirements for c-di-GMP signal transduction in V. cholerae. We identify specific c-di-GMP proteins that additively stimulate the subcellular localization and activity of the c-di-GMP binding protein and transcriptional regulator, VpsT. We further show that c-di-GMP signaling does not require the interaction of c-di-GMP signaling components. However, a common cellular level of c-di-GMP contributes to VpsT localization and activity. This is the first account of the subcellular localization of a transcriptional regulator modulated by c-di-GMP binding. Furthermore, this study establishes that c-di-GMP signal transduction can act at a distance through a common cellular level of c-di-GMP and defines how an intracellular second messenger can regulate cellular processes in bacteria.
| Second messengers are small diffusible signaling molecules that are produced or degraded in response to external stimuli and relay information to a receptor to elicit a specific cellular response [1]. The cyclic nucleotide cyclic diguanylate (c-di-GMP) is a bacterial second messenger that controls the transition between a free living and biofilm lifestyle [2], [3]. C-di-GMP is produced by diguanylate cyclases (DGCs), containing GGDEF domains, and degraded by phosphodiesterases (PDEs), containing EAL or HD-GYP domains. Cellular c-di-GMP is sensed by receptors that interact with downstream targets to affect cellular functions. C-di-GMP signaling often involves numerous GGDEF, EAL or HD-GYP domain containing proteins and receptors [4], and previous reports suggest that the compartmentalization of c-di-GMP signaling components could facilitate the activation of specific cellular processes [3], [5], [6]. However, it is currently unclear whether compartmentalization is required to mediate c-di-GMP signal transduction in bacteria.
Recent advances in the identification of c-di-GMP receptors have helped define the mechanisms by which c-di-GMP mediates downstream processes. These receptors include riboswitches [7] and proteins that contain various binding domains. PilZ domains are known to bind c-di-GMP and proteins harboring these domains modulate cellular processes such as motility through protein-protein interactions with the flagellar motor complex [8]–[10]. Proteins containing degenerate GGDEF or EAL domains, which have lost their enzymatic activity, are also known to be c-di-GMP receptor proteins. In Pseudomonas fluorescens, LapD binds c-di-GMP through a degenerate EAL domain and modulates the cell surface association of an adhesin through direct interactions with a periplasmic protease [11]–[13]. The degenerate GGDEF domain containing protein CdgG was shown to regulate biofilm formation in Vibrio cholerae [14]. C-di-GMP can also regulate gene expression by binding transcriptional regulators such as the Crp homolog Clp [15] or FleQ [16]. Although the identities of many c-di-GMP receptor proteins are known, the mechanisms of c-di-GMP-mediated signal transduction and gene regulation are not fully understood.
In V. cholerae, the bacterial pathogen that causes the disease cholera, c-di-GMP regulates biofilm formation, motility and virulence [17]–[19]. The V. cholerae genome contains 31 genes encoding proteins with GGDEF domains, 11 genes encoding proteins with EAL domains, 10 genes encoding proteins with both GGDEF and EAL domains and 9 genes encoding proteins with HD-GYP domains [14], [20]. Recently, we characterized VpsT, which is a key c-di-GMP receptor known to regulate biofilm formation in V. cholerae [21]. Biofilm formation in V. cholerae requires the biosynthesis of Vibrio polysaccharide (VPS) [22], [23], and VpsT activates vps expression through direct binding of the vpsL promoter [21], [24]. VpsT is a novel member of the FixJ, LuxR and CsgD family of transcriptional regulators that contains a c-di-GMP binding motif and a 6th alpha helix at its N-terminal receiver domain that stabilizes homodimerization [21]. These features make VpsT unique compared to other response regulators and c-di-GMP binding proteins.
In this study, we report that VpsT requires c-di-GMP binding and subcellular localization to regulate gene expression. The wild-type VpsT localization pattern is dependent on c-di-GMP binding, c-di-GMP-stabilized dimerization, and the VpsT DNA binding domain. We also show that VpsT does not co-localize or interact with DGCs. Instead, multiple DGCs contribute additively to a cellular c-di-GMP concentration, which affects the localization and activity of the c-di-GMP receptor protein, VpsT.
We hypothesized that the c-di-GMP receptor protein, VpsT, is subcellularly localized, and this localization facilitates c-di-GMP signal transduction. To determine whether VpsT is subcellularly localized, we constructed an N-terminal tagged green fluorescent protein (GFP)-VpsT fusion. Expression of gfp-vpsT recovered vpsL expression in a ΔvpsT strain (Figure 1A). vpsL is the first gene in the vps-II operon and VpsT directly binds to the upstream regulatory region of this gene [21], [22]. Expression of vpsL was similar between strains expressing gfp-vpsT or vpsT alone indicating that our fusion protein is functional. When observed by fluorescence microscopy, GFP-VpsT formed a pattern of localization within the cell (Figure 1B), while a strain expressing GFP exhibited homogenous fluorescence throughout the cytoplasm. We confirmed that this localization was not due to different cellular protein concentrations, as levels of GFP-VpsT were similar to levels of GFP alone (Figure S1). A census of more than 150 cells per treatment showed that cells expressing GFP-VpsT contained more spots per cell when compared to cells expressing GFP alone when quantified using MicrobeTracker software (Figure 1C) [25]. GFP-VpsT localization also exhibited a higher ratio of maximum to average fluorescence intensity across the length of individual cells when compared to cells expressing GFP alone (Figure S1). These results indicate that GFP-VpsT is subcellularly localized.
The striking number of GGDEF, EAL and HD-GYP domain containing proteins present in many bacteria is thought to generate flexibility in signal transduction, allowing multiple sensory inputs to be fed through a single diffusible signaling molecule [4]. Since VpsT is a c-di-GMP binding protein and is subcellularly localized, we wondered whether specific DGCs or PDEs are important for this localization pattern. We therefore measured expression of the vpsL promoter in wild-type V. cholerae and 52 strains containing in-frame deletions of each gene in the V. cholerae genome encoding proteins with GGDEF, EAL or GGDEF and EAL domains. Of the strains examined, 5 DGC deletion strains showed a 2-fold or greater decrease in expression of vpsL (Figure 1D and S2), namely the previously characterized genes encoding DGCs cdgA (VCA0074), cdgH (VC1067), cdgK (VC1104) and cdgL (VC2285) [14], [26], [27], and a predicted DGC, VC1376, which we name here, cdgM. Furthermore, c-di-GMP levels decreased between 86% and 54% in each single DGC deletion strain when compared to wild type (Figure 1E). These results show that multiple DGCs are involved in vps regulation and thus identified likely candidate DGCs important for VpsT localization.
We then observed VpsT localization in strains lacking each of the 5 DGCs important for vpsL expression. VpsT localization was not markedly altered in any strain containing a single DGC deletion (Figure S1). We then reasoned that VpsT localization may not be dependent on a single DGC, but instead, multiple DGCs contribute additively to VpsT localization. Therefore, we created a strain where all 5 DGCs are deleted in combination, designated Δ5DGC. Δ5DGC exhibited a lower vpsL expression than any single DGC mutant strain (Figure 1D). Moreover, c-di-GMP levels were significantly decreased (17%) in the Δ5DGC strain compared to wild type (Figure 1E). In the Δ5DGC strain, GFP-VpsT localization was reduced and the number of spots per cell and ratio of maximum to average fluorescence intensity were markedly lower compared to wild type expressing the same fluorescent fusion protein (Figure 1B, 1C and S1). This change in GFP-VpsT localization was not due to different cellular protein concentrations, as GFP-VpsT levels were similar to levels of GFP alone in the Δ5DGC strain (Figure S1). These results indicate that no single DGC is sufficient to cause VpsT mis-localization, and instead, multiple DGCs additively impact the GFP-VpsT localization pattern. The number of spots per cell in the Δ5DGC strain was not completely diminished, and we speculate that a low level of c-di-GMP is still present in the cell due to remaining DGCs, which facilitate VpsT localization. Alternatively, a range of VpsT target promoters that differ in their affinities for the active regulator could cause this localization pattern. Above observations of VpsT localization and activity suggest that VpsT function is dependent on reaching a critical cellular c-di-GMP threshold. Thus we wondered whether a single DGC could rescue vpsL expression in the Δ5DGC strain. When cdgA was expressed on a plasmid in the Δ5DGC mutant, vpsL expression was recovered in the Δ5DGC strain when compared to the Δ5DGC mutant harboring the vector alone (Figure S3). These results suggest that one DGC can rescue a cellular level of c-di-GMP for the activation of vpsL expression in the Δ5DGC strain.
In our survey of DGC and PDE mutants, we also observed multiple PDEs to be negative regulators of vps expression (Figure S2), consistent with previous work [26], [28]–[30]. However, strains harboring deletions of three of these genes encoding PDEs, mbaA, rocS and cdgC individually or in combination, exhibited no significant alteration in GFP-VpsT localization pattern (Figure S4). Therefore, an upper c-di-GMP concentration limit may exist, after which, further VpsT localization is not observable. Alternatively, the experimental system might be saturated, and no further localization can be observed.
VpsT as a response regulator is not unique in its capacity to subcellularly localize in response to specific stimuli or modification. CsgD from Salmonella enterica was shown to form foci associated with the membrane in a subpopulation of cells in response to cell aging [31]. WspR from Pseudomonas aeruginosa was shown to localize to foci in response to phosphorylation [32]. OmpR from Escherichia coli subcellularly localizes in response to the presence and activity of its cognate histidine kinase, EnvZ [33]. Whereas typical response regulators, such as OmpR, are activated by a single major cognate histidine kinase [34], VpsT localization and activity is modulated in response to c-di-GMP produced by multiple DGCs. These results are consistent in the context of second messenger signaling, where multiple independent inputs can be fed through a single diffusible signaling molecule to elicit a specific cellular response [1].
It is proposed that the subcellular compartmentalization of c-di-GMP signaling components might allow c-di-GMP to act locally on specific cellular processes such as motility or biofilm formation [5], [35]. C-di-GMP signaling proteins could exert their effects by participating in complexes that include signal producers (DGC), removers (PDE), receptors, and/or targets [3], [6]. To determine if co-localization occurs between DGCs activating VpsT and the c-di-GMP receptor, VpsT, we analyzed their subcellular localization. We chose CdgA and CdgH, two DGCs that affect vps expression (Figure 1D) and have demonstrated DGC activity (Shikuma and Yildiz, unpublished data) [14]. To observe the subcellular localization of CdgA and CdgH we constructed C-terminal tagged CdgA-GFP and CdgH-GFP fusions. Both cdgA-gfp and cdgH-gfp were able to complement in-frame deletions of cdgA and cdgH, respectively (Figure S5), indicating that our fusion proteins are functional. When observed by fluorescence microscopy, CdgA-GFP and CdgH-GFP both appeared to localize to the cell membrane (Figure 2A). Consistent with these results, both CdgA and CdgH are predicted to contain 2 and 1 transmembrane domains, respectively [36].
To corroborate these results, we performed cellular fractionation and western blot analysis to identify the subcellular location of VpsT, CdgA and CdgH. We therefore created strains containing an N-terminal HA tagged vpsT, a C-terminal HA tagged cdgA or a C-terminal HA tagged cdgH in their native chromosomal loci. Strains containing each fusion protein exhibited similar vpsL expression when compared to wild type (Figure S5). Both CdgA-HA and CdgH-HA localized to the total membrane fraction, as predicted (Figure 2B). In contrast, HA-VpsT localized mostly to the cytoplasmic fraction, but a lower level also consistently appeared in the total membrane fraction. To determine whether VpsT localization is dependent on the presence of specific DGCs or c-di-GMP levels, we performed a cellular fractionation of wild-type and Δ5DGC strains and probed for HA-VpsT levels. HA-VpsT localization was not different between wild-type and Δ5DGC strains (Figure S6), suggesting that the 5 DGCs or c-di-GMP levels are not important for the localization of VpsT to specific cellular fractions.
Although VpsT resides mainly in a different subcellular region of the cell when compared to CdgA or CdgH, it is possible that transient interactions between these proteins contribute to specificity in c-di-GMP signaling. To address whether VpsT can interact with CdgA or CdgH directly, we performed a bacterial two-hybrid analysis using a system suited to identify protein-protein interactions, even under the condition where one or both proteins are membrane bound [37]. Using bacterial two-hybrid vectors, VpsT, CdgA and CdgH were fused to the T18 or T25 complementary fragments of Bordetella pertussis adenylate cyclase (CyaA). Interaction between co-expressed proteins of interest in E. coli reconstitute a functional CyaA, leading to cAMP production [38]. As expected, a signal indicative of interaction of VpsT with itself was observed by colorimetric blue production on LB agar containing bromo-chloro-indolyl-galactopyranoside (X-gal), as well as quantitatively using β-galactosidase assays (Figure 2C and S7). Interaction of CdgA with itself and CdgH with itself was also observed (Figure 2C and S7). These results were expected as DGCs from other bacteria were shown previously to catalyze c-di-GMP production as dimers [39], [40]. Interestingly, E. coli containing CdgA and CdgH on complementary plasmids exhibited increased β-galactosidase production, suggesting that these DGCs might interact, however the physiological relevance of this observation is unclear at this point. Importantly, strains expressing both VpsT and CdgA or VpsT and CdgH did not exhibit increased cAMP production, even when the reciprocal exchange of fusion domains was performed (Figure 2C and S7). These results suggest that VpsT does not interact directly with CdgA or CdgH.
We next wondered whether VpsT localization is dependent on specific domains and/or interfaces important for VpsT function. Mutations in residues required for c-di-GMP binding (VpsTR134A) or c-di-GMP-stabilized dimerization (VpsTI141E) were unable to complement a ΔvpsT mutation (Figure 3A), consistent with our previous findings [21]. When observed by fluorescence microscopy, both GFP-VpsTR134A and GFP-VpsTI141E mutants exhibited a homogenous fluorescence throughout the cytoplasm, possessed almost no spots per cell, and showed a significantly lower maximum to average fluorescence intensity ratio when compared to strains expressing a wild-type GFP-VpsT fusion (Figure 3B, 3C and S8). VpsT contains a C-terminal helix-turn-helix (HTH) DNA binding domain and H193 of VpsT aligned with other histidine residues in the LuxR/FixJ superfamily shown previously to be required for DNA binding (Figure S8) [41], [42]. A strain harboring GFP-VpsTH193A was unable to induce vps expression (Figure 3A) and appeared to localize to foci that were more dispersed throughout the cell when compared to wild type GFP-VpsT (Figure 3B). The number of spots per cell and the ratio of maximum to average fluorescence intensity of the GFP-VpsTH193A expressing strain were decreased compared to wild-type GFP-VpsT (Figure 3C and S8). Therefore, VpsT localization, albeit different than that of the wild-type localization pattern, can still occur in the absence of DNA binding. The subcellular localization patterns were not due to differential protein levels, as cellular concentrations of wild-type GFP-VpsT were similar to GFP-VpsT with R134A, I141E or H193A point mutations (Figure S8). Taken together, our results indicate that the wild-type VpsT localization pattern is dependent on c-di-GMP binding and DNA binding. These results suggest that VpsT forms oligomers on DNA binding sites distributed on the V. cholerae chromosomes and the localization pattern is due to binding of VpsT to its target sequences on the genome.
To determine whether there are other factors responsible for VpsT localization in V. cholerae, we expressed GFP-VpsT in E. coli. GFP-VpsT was surprisingly homogenous throughout the cytoplasm when expressed in E. coli in contrast to the same construct expressed in V. cholerae (data not shown), suggesting that the localization of VpsT requires cellular components or a cellular environment provided by the V. cholerae cell. We then hypothesized that the localization of VpsT might either require increased levels of c-di-GMP or specifically require a DGC important for biofilm formation in V. cholerae. A compatible plasmid that expresses cdgA from an IPTG inducible promoter was therefore introduced into E. coli containing GFP-VpsT. Strains expressing CdgA showed a marked decrease in motility when compared to strains containing vector alone (Figure 4C), indicating that CdgA is functional as a DGC in E. coli. When observed by fluorescence microscopy, GFP-VpsT formed foci in the presence of CdgA in E. coli (Figure 4A). This strain exhibited an increase in the number of spots per cell and a significantly increased ratio of maximum to average fluorescence intensity compared to a strain with GFP-VpsT and an empty compatible plasmid (Figure 4A, 4B and S9). To determine whether VpsT localization is dependent on the catalytic activity of CdgA, we also expressed CdgA containing a point mutation converting the conserved GGDEF motif to GADEF (cdgAG287A) in cells also expressing GFP or GFP-VpsT. Expression of CdgAG287A in E. coli was not able to recover VpsT localization, in contrast to wild type CdgA (Figure 4A, 4B and S9). Furthermore, the motility zone diameter of a strain expressing CdgAG287A was similar to that of a strain with vector alone (Figure 4C). As expected, strains expressing GFP alone with the same compatible plasmids showed no localization pattern (Figure 4A, 4B and S9). These results suggest that the catalytic activity of CdgA as a DGC is required for VpsT localization.
We show above that the wild-type VpsT localization in V. cholerae is dependent on an intact DNA binding domain. To test whether VpsT requires DNA binding in E. coli, we expressed GFP-VpsT with a plasmid harboring the vpsL promoter (vpsLp). However, this strain did not exhibit a VpsT localization pattern (Figure 4A, 4B and S9). To determine whether VpsT requires both CdgA and a native DNA binding region, we expressed GFP-VpsT in cells containing a plasmid with both cdgA and the vpsL promoter. In this strain, GFP-VpsT appeared to form a more discrete pattern, exhibited an increased number of spots per cell, and a higher maximum to average intensity ratio when compared to GFP-VpsT cells co-expressing only CdgA (Figure 4A, 4B and S9). E. coli co-expressing GFP alone with the same compatible plasmids showed no localization pattern (Figure 4A, 4B and S9). To further determine whether GFP-VpsT activity requires c-di-GMP in E. coli, we quantified the expression of vpsL in the presence and absence of CdgA. Only E. coli co-expressing GFP-VpsT and CdgA activated vpsL expression while a strain expressing only GFP-VpsT did not show vpsL activation (Figure 5). These results suggest that VpsT localization is enhanced by DNA binding and requires elevated c-di-GMP levels to activate gene expression.
We then wondered whether CdgA, as a V. cholerae DGC, is required for VpsT localization or if a heterologous DGC could induce VpsT to localize. We therefore expressed adrA, a previously characterized gene encoding a DGC from Salmonella typhimurium [43], in strains also containing GFP or GFP-VpsT. Strains expressing AdrA showed a marked decrease in motility (Figure 4C), indicating that AdrA is functional in E. coli. In E. coli, AdrA caused GFP-VpsT to localize to foci, similar to foci induced by CdgA (Figure 4A, 4B and S9). As expected, co-expression of GFP and AdrA showed no localization. These results indicate that VpsT localization depends on the cellular level of c-di-GMP, and not on the presence of a specific V. cholerae DGC. Altogether, our results suggest that a direct interaction is not required for c-di-GMP signal transduction between DGCs and c-di-GMP receptors.
Recently, the subcellular localization of other c-di-GMP receptors was found to be dependent on c-di-GMP binding. C-di-GMP controls the subcellular localization of the PilZ domain containing c-di-GMP receptor YcgR in E. coli, where interaction of a YcgR-c-di-GMP complex with the flagellar motor leads to decreased motility and counter-clockwise rotational bias [8]–[10]. Moreover, multiple DGCs were shown to contribute additively to these motility phenotypes [8]. In Caulobacter crescentus, c-di-GMP binding to a conserved I-site of PopA mediates the sequestration of this protein to the cell pole, where PopA facilitates cell cycle progression [44]. No single deletion of a GGDEF or EAL domain containing protein was sufficient to alter PopA localization [44]. However, the combined activity of two DGCs, PleD and DgcB, was shown to alter cell cycle dynamics [45]. The subcellular localization of YcgR and PopA appears to be modulated by the additive activity of multiple DGCs in combination, similar to our findings with VpsT.
This study is the first account of the subcellular localization of a c-di-GMP binding transcriptional regulator. Results presented here suggest that adequate levels of c-di-GMP contributed by multiple DGCs modulate VpsT activity and not a physical interaction or compartmentalization of c-di-GMP signaling components (Figure 6). This study identifies the requirements for signal transduction, localization and activity of a c-di-GMP receptor protein and furthers our understanding of the mechanisms of second messenger signaling.
The bacterial strains and plasmids used in this study are listed in Table S1. In-frame deletion, chromosomal fusion and point mutation strains were generated according to previously published protocols [46]. All V. cholerae and E. coli strains were grown aerobically, at 30°C and 37°C, respectively, unless otherwise noted. Growth medium consisted of LB media (1% Tryptone, 0.5% Yeast Extract, 1% NaCl), pH 7.5. LB-agar and soft agar plates contained 1.5% and 0.3% (wt/vol) granulated agar (Difco), respectively. Concentrations of antibiotics used, where appropriate, were as follows: ampicillin (100 µg/ml), rifampicin (100 µg/ml), chloramphenicol (E. coli 20 µg/ml, V. cholerae 5 µg/ml), kanamycin (50 µg/ml) and gentamicin (30 µg/ml).
All strains were verified by PCR. Plasmid sequences were verified by DNA sequencing by Sequetech Corporation (Mountain View, CA). Primers used in the present study were purchased from Bioneer Corporation (Alameda, CA) and sequences are available upon request.
V. cholerae cells harboring the indicated plasmid were grown overnight (15 to 17 h) aerobically in LB medium supplemented with ampicillin. Cells were then diluted 1∶1000 in fresh LB medium and grown aerobically for 2 h, at which point arabinose was added at a final concentration of 0.05% and cells were harvested at exponential phase 2 h later (optical density at 600 nm (OD600 nm) of 0.2 to 0.4). E. coli cells containing the indicated plasmid were grown overnight in LB medium containing 2% glucose, kanamycin and ampicillin. Cells were then diluted 1∶50 in fresh LB medium containing 0.1% arabinose and 100 µM IPTG and cells were harvested 3 h later. Cell culture was spotted onto 1% agarose pads prepared with phosphate-buffered saline (PBS), pH 7.4. Images were acquired using a Zeiss Axiovert 200 microscope equipped with a 63× Plan-Apochromat objective (numerical aperture, 1.4), and were recorded with a Cool-Snap HQ2 camera (Photometrics). Images were minimally processed using Adobe Photoshop 11.0 and ImageJNIH software. MicrobeTracker [25] was employed, using the alg4ecoli parameter to identify cell outlines, the spotFinderZ tool to determine the number of spots per cell and the intprofile tool to determine the maximum and average fluorescence intensities of single cells. Data were acquired from at least 3 independent experiments and quantification was performed on at least 150 cells per treatment. All statistics were calculated using Graphpad Prism 4.
Overnight cultures were diluted 1∶200, grown to an OD600 nm of 0.3 to 0.4, and diluted again 1∶200. Cells were harvested at an OD600 nm of 0.3 to 0.4 by centrifugation (10,000× g) and fractionation was carried out as described previously [47]. Protein levels were quantified using a bicinchoninic acid (BCA) kit (Thermo Fisher Scientific Inc.) and normalized between fractions. Proteins were separated on a 12% SDS-polyacrylamide gel and electroblotted onto a nitrocellulose membrane with a Mini Trans-Blot Cell (Bio-Rad) as described previously [47]. Rabbit polyclonal antiserum against V. cholerae OmpU (provided by K. Klose) was used at a dilution of 1∶100,000. Mouse monoclonal antibody against GFP (Santa Cruz Biotechnology) and rabbit polyclonal antibody against the HA epitope (Santa Cruz Biotechnology) were used according to the manufacturer's instructions. Horseradish peroxidase-conjugated goat anti-rabbit secondary antibody (Santa Cruz Biotechnology) or goat anti-mouse secondary antibody (Santa Cruz Biotechnology) was used according to the manufacturer's instructions. Immunoblot analyses were conducted with at least three biological replicates.
β-galactosidase assays were performed and Miller units calculated as described previously [48]. The assays were repeated with three biological replicates and six technical replicates.
V. cholerae or E. coli cells harboring the indicated plasmid were grown overnight (15 to 17 h) aerobically in LB medium supplemented with the appropriate antibiotics. Cells were then diluted 1∶1000 in fresh LB medium and harvested at exponential phase at an OD600 nm of 0.3 to 0.4. E. coli were grown in the presence of 0.1% arabinose and 100 µM IPTG for protein expression. Luminescence was measured using a Victor3 Multilabel Counter (PerkinElmer) and Lux expression is reported as counts min−1 ml−1/OD600 nm. Assays were repeated with at least three biological replicates and four technical replicates.
Cellular c-di-GMP levels were measured in the indicated strains grown to exponential phase in LB medium. Protein concentration was determined using a BCA kit according to the manufacturer's instructions. C-di-GMP extraction, analysis by high-performance liquid chromatography-tandem mass spectrometry (HPLC-MS/MS) and c-di-GMP standard curve generation were carried out as described previously [26]. C-di-GMP quantification was performed with at least three biological replicates.
Bacterial two-hybrid assays were performed as described previously [38]. Translational fusions were created with proteins of interest and T18 or T25 fragments of B. pertussis adenylate cyclase (CyaA). All constructs were confirmed by DNA sequencing. Plasmids pKT25-zip and pUT18C-zip, each containing translational fusions to the leucine zipper of GCN4, were used as positive controls. Production of cAMP by reconstituted CyaA was observed in the E. coli strain BTH101, lacking a native cyaA gene. Protein-protein interactions were observed by growing cells for 48 to 72 h at 30°C on LB agar containing ampicillin (100 µg/ml), kanamycin (50 µg/ml), X-gal (40 µg/ml) and IPTG (10 to 500 µM), or quantified by performing β-galactosidase assays with cells grown overnight at 30°C in LB medium containing ampicillin (100 µg/ml), kanamycin (50 µg/ml) and IPTG (10 µM).
GenBank accession numbers are as follows: VpsT, NP_233336.1; VpsL, NP_230581.1; CdgA, NP_232475.1; CdgH, NP_230712.1; CdgK, NP_230749.1; CdgL, NP_231916.1; CdgM, NP_231020.1; MbaA, NP_230352.1; RocS, NP_230302.1; CdgC, NP_233171.1.
|
10.1371/journal.pbio.1000193 | Negative Regulation of Active Zone Assembly by a Newly Identified SR Protein Kinase | Presynaptic, electron-dense, cytoplasmic protrusions such as the T-bar (Drosophila) or ribbon (vertebrates) are believed to facilitate vesicle movement to the active zone (AZ) of synapses throughout the nervous system. The molecular composition of these structures including the T-bar and ribbon are largely unknown, as are the mechanisms that specify their synapse-specific assembly and distribution. In a large-scale, forward genetic screen, we have identified a mutation termed air traffic controller (atc) that causes T-bar–like protein aggregates to form abnormally in motoneuron axons. This mutation disrupts a gene that encodes for a serine-arginine protein kinase (SRPK79D). This mutant phenotype is specific to SRPK79D and is not secondary to impaired kinesin-dependent axonal transport. The srpk79D gene is neuronally expressed, and transgenic rescue experiments are consistent with SRPK79D kinase activity being necessary in neurons. The SRPK79D protein colocalizes with the T-bar-associated protein Bruchpilot (Brp) in both the axon and synapse. We propose that SRPK79D is a novel T-bar-associated protein kinase that represses T-bar assembly in peripheral axons, and that SRPK79D-dependent repression must be relieved to facilitate site-specific AZ assembly. Consistent with this model, overexpression of SRPK79D disrupts AZ-specific Brp organization and significantly impairs presynaptic neurotransmitter release. These data identify a novel AZ-associated protein kinase and reveal a new mechanism of negative regulation involved in AZ assembly. This mechanism could contribute to the speed and specificity with which AZs are assembled throughout the nervous system.
| Neurons communicate with each other through electrochemical impulses transmitted primarily at specialized intercellular junctions termed synapses. At each synapse, the primary site of synaptic vesicle fusion occurs at the active zone, an electron-dense presynaptic membrane with associated fibrillary matrix. Many active zones also possess one or more electron-dense cytosolic projections that are believed to facilitate vesicle mobilization to the active zone membrane and are required for normal synaptic transmission. These electron-dense projections are referred to as T-bars in Drosophila or ribbons in vertebrates. The molecular composition of these structures remains poorly characterized, and very little is known about how these structures are specifically assembled and stabilized at the presynaptic membrane. Here, we identify in Drosophila a neuronally expressed serine-arginine kinase called SRPK79D that localizes to the presynaptic active zone and that through its kinase activity appears to repress T-bar formation within peripheral axons. Our study thus provides evidence for kinase-dependent repression of active zone assembly, with implications for the development and growth of synaptic connections throughout the nervous system.
| The majority of stimulus-dependent synaptic vesicle fusion occurs at presynaptic specializations called active zones (AZs). Ultrastructurally, AZs consist of at least two components; 1) a presynaptic membrane of high electron density, reflecting the presence of proteins such as Ca2+ channels, t-SNAREs, and cell adhesion molecules and 2) a fibrillary cytomatrix (CAZ) that includes cytoskeletal elements, scaffolding proteins, and AZ-specific molecules such as Piccolo/Aczonin, Bassoon, Unc-13/Dunc-13/Munc-13, RIM, and ELKS/Brp/ERC [1]. Many synapses that are characterized by a high release probability also include an electron-dense cytosolic projection that is believed to facilitate synaptic vesicle movement to the AZ. These projections are referred to as ribbons in the mammalian retina, dense bodies at the mammalian neuromuscular junction (NMJ), and T-bars at the Drosophila NMJ [1]–[3]. To date, five proteins have been found within the presynaptic ribbon at synapses in the vertebrate retina, including Piccolo, Kif3A, RIM, CtBP1, and RIBEYE/CtBP2 [1]–[3].
In Drosophila, there are no clear homologs of RIBEYE or Piccolo, and it remains unknown whether RIM or Kif3A associate with the Drosophila T-bar. The protein currently known to localize at the T-bar is the Drosophila homolog of ELKS/ERC, called Bruchpilot (Brp) [4],[5]. Recently, it was demonstrated that mutations in the brp gene eliminate T-bars and severely impair synaptic vesicle release, consistent with the conclusion that T-bars and Brp are essential components of the presynaptic AZ [4],[5].
T-bars and ribbons are large, macromolecular structures. In Drosophila, T-bars are first assembled at late embryonic stages as the nascent neuromuscular synapse begins to mature [6]–[8]. The appearance of T-bars and T-bar-associated antigens correlates with the ability of the neuromuscular junction to support larval movement. T-bars are formed only at the presynaptic face of the AZ and are not found at other sites, implying the existence of mechanisms that ensure site-specific assembly of these large, cytoplasmic structures. However, virtually nothing is known about how T-bar and ribbon structures are assembled and positioned at the AZ.
One model for AZ formation that could be extended to ribbon/T-bar assembly is based upon the existence of transport vesicles that contain AZ components, including calcium channels as well as the Piccolo and Bassoon proteins. It has been suggested that these transport vesicles fuse at sites of nascent synapse formation to deliver protein constituents of the AZ in a site-specific manner [9]–[11]. Although transport vesicles have not been isolated in Drosophila motoneurons, it was recently demonstrated that mutation of a Kinesin 3 (immaculate connections; imac) prevents the transport of synaptic vesicle proteins to the developing synapse, and in this mutant background, both AZ and T-bar formation are significantly impaired [8]. These data suggest that a critical component of AZ and T-bar assembly is contributed by Imac-dependent axonal transport. Although transport vesicles could represent a mechanism to deliver transmembrane and membrane-associated proteins to the AZ, there presumably exist other mechanisms to control the site-specific assembly of cytoplasmic proteins into a T-bar.
Here, we describe a previously uncharacterized gene in Drosophila that encodes a serine-threonine kinase that we have termed serine-arginine protein kinase at 79D (srpk79D). The SRPK79D protein is a member of the serine-arginine protein kinase family previously shown to be involved in mRNA splicing and processing [12]. This gene was identified in a large-scale forward genetic screen for genes involved in the development, maturation, and stabilization of the Drosophila NMJ. In this study, we present evidence that SRPK79D is a T-bar-associated protein kinase that is necessary to prevent premature T-bar assembly in peripheral axons. We also present evidence that SRPK79D activity must be overcome within the NMJ for normal AZ assembly and neurotransmission. As such, our data identify a new T-bar-associated antigen and indicate that synapse-specific assembly of the presynaptic T-bar may be achieved, in part, through suppression of T-bar assembly at nonsynaptic sites including the axon.
In an ongoing screen to identify genes involved in the formation and stabilization of the Drosophila NMJ, we identified a P-element insertion (P10036) in which the peripheral nerves contain numerous large accumulations of the AZ associated protein Brp (Figure 1E–1G). These large, aberrant Brp accumulations ranged from roughly spherical to grossly elongated in appearance (Figure 1E–1G). In wild-type animals, by contrast, axons within the peripheral nerves showed virtually no anti-Brp staining and the Brp puncta that did appear in these axons were small and spherical in appearance (Figure 1B–1D).
This phenotype is very unusual, based upon the results of our ongoing genetic screen. In this forward genetic screen, we have analyzed over 2,000 independent transposon insertion lines, including PiggyBac lines on chromosomes 2 and 3 from the Exelixis collection and an independent collection of P{GAL4} lines [13]. In each mutant background, we have stained three to five larvae with anti-Brp and anti-Discs Large (Dlg) antibodies and examined both the peripheral nerves and the neuromuscular synapse for defects. P10036 is the only mutation identified to date that causes the observed accumulation of anti-Brp staining in peripheral axons. The P10036 transposon resides within an intron of the previously uncharacterized gene CG11489, which resides at chromosomal position 79D and is predicted to encode a member of the SRPK family (Figure 1A and see below). Due to the dramatic effect on Bruchpilot (German for crash pilot) protein accumulation in peripheral axons, we named this mutant air traffic controller (atc), and we refer to P10036 as srpk79Datc throughout this article.
We next developed quantitative measures of the axonal Brp accumulations to further characterize and analyze the srpk79Datc mutant phenotype (see Materials and Methods). In all cases, genetic controls were dissected, processed, stained, and imaged identically and in parallel with srpk79Datc mutants. We found a statistically significant increase in total nerve Brp fluorescence in srpk79Datc mutants compared to wild-type and heterozygous controls (p<0.001, Student t-test; Figure 1K). We also found a highly significant increase in the average puncta fluorescence intensity compared to wild-type and heterozygous controls. Indeed, the entire distribution of puncta intensities was shifted toward larger values (p<0.001, Mann-Whitney U Test; Figure 1L). Finally, we estimate that the frequency of these aberrant accumulations corresponds to 0.03 accumulations per micron of individual motor axon length. From these data, we conclude that Brp-positive puncta in srpk79Datc mutant axons represent larger, abnormal, protein aggregates compared to observations made in wild-type axons.
Next, we assayed synaptic Brp staining intensity and NMJ morphology in the srpk79Datc mutant. We found that synaptic Brp staining intensity is significantly decreased compared to wild-type animals, assayed as both total Brp fluorescence (p<0.001, Student t-test; Figure 2A–2E) and as the distribution of individual puncta intensities (p<0.001, Mann-Whitney U Test; Figure 2A–2D and 2F). This effect occurs at NMJ throughout the animal, and there is no evidence for a strong anterior–posterior gradient of this phenotype (Figure S1). Our data suggest that the accumulation of Brp aggregates in the axon of the srpk79Datc mutant depletes Brp protein from the presynaptic nerve terminal. Consistent with this conclusion, we found that total Brp protein levels, assayed by western blot, are unaltered in the srpk79Datc mutant background despite the dramatic increase in nerve Brp (see below).
We also determined whether the decrease in total Brp fluorescence causes a decrease in total Brp puncta number, which would be indicative of a change in AZ number. We found, however, that Brp puncta density within srpk79Datc mutant NMJs is identical to wild type and that total bouton numbers are wild type in the srpk79Datc mutant background (Figure 2G and 2H). Moreover, anti-Dlg, anti-Synaptotagmin 1 (Syt), and anti-Cysteine String Protein (CSP) staining at srpk79Datc mutant synapses are not different compared to wild type (unpublished data). Thus, synapse growth, morphology, and AZ number appear normal in the srpk79Datc mutant.
Consistent with the observed lack of morphological change, we found no change in neurotransmitter release in the srpk79Datc mutant background. We assayed neurotransmission by recording from the third-instar NMJ of homozygous sprk79Datc mutants, as well as homozygous srpk79Datc mutants lacking one copy of the brp gene (brp69/+;srpk79Datc) [5]. In all cases, evoked excitatory junctional potential (EJP) amplitude and spontaneous miniature EJP (mEJP) amplitudes were wild type (wild-type average EJP = 34.28±1.59 mV compared to srpk79Datc = 34.27±1.28 mV; n = 10, p>0.3; wild type average mEJP = 0.97±0.04 mV compared to srpk79Datc = 0.99±0.03 mV; n = 10, p>0.3). There was also no difference in the ability of the NMJ to sustain high-frequency (10 Hz) stimulation in high extracellular calcium saline (2 mM) (unpublished data; see below for additional electrophysiological analyses). Thus, the srpk79Datc mutant causes inappropriate axonal accumulations of Brp protein, resulting in a depletion of this synaptic protein from the presynaptic AZ. However, the amount of depletion of Brp from the NMJ does not cause a defect in synaptic function over the time course of 4 d of larval development.
We have used our quantitative assays to confirm that the phenotype of axonal Brp accumulation is caused by disruption of the srpk79D gene and to determine the nature of this genetic disruption. First, we demonstrated that the axonal Brp accumulation and synaptic Brp deficit phenotypes in the homozygous srpk79Datc mutant are statistically identical to those observed when the srpk79Datc mutation is placed in trans to a deficiency chromosome that uncovers the srpk79D gene locus, Df(3L)Exel6138 (Figures 1E–1L, 2E, and 2F). Furthermore, an independently identified molecular null allele of srpk79D (srpk79DVN100; Eric Buchner, personal communication), has axonal and synaptic Brp phenotypes that are statistically identical to those observed in homozygous sprk79Datc (Figure S2A–S2H). These data are consistent with the conclusion that the srpk79Datc transposon insertion is a strong loss-of-function or null mutation in the srpk79D gene. Interestingly, we found that the heterozygous srpk79Datc/+ mutant axons also have a slight, but statistically significant, increase in Brp fluorescence compared to wild type. These data indicate that srpk79D is partially haploinsufficient for the regulation of axonal Brp accumulation.
Next, we determined the expression pattern of the srpk79D gene. In situ hybridizations performed on wild-type Drosophila embryos targeting an exon common to all known srpk79D transcripts (see Materials and Methods) detected high levels of srpk79D mRNA in the embryonic ventral nerve cord with lower expression present outside of the nervous system (Figure 3A and 3B). This expression pattern is consistent with a function of srpk79D gene products in neurons, but does not rule out a possible function in other tissues including peripheral glia.
To confirm that loss of srpk79D is responsible for the phenotype of axonal Brp accumulation, and to determine where srpk79D is required for normal Brp targeting, we employed a srpk79D RNA interference (RNAi) transgene (UAS-srpk79DRNAi; Vienna Drosophila RNAi Collection). We found that expression of UAS-srpk79DRNAi in neurons phenocopies the srpk79Datc mutation (Figure 3C–3F), whereas expression of UAS- srpk79DRNAi in glia (also present in peripheral nerve) does not cause formation of axonal Brp aggregates. These data indicate that srpk79D function is required in neurons, consistent with enriched expression in the central nervous system (CNS).
We also performed a genetic rescue experiment by expressing a Venus-tagged, full-length srpk79D transgene (UAS-v-srpk79D-rd*) in neurons in the homozygous srpk79Datc mutant background. In this experiment, neuronal expression of UAS-v- srpk79D-rd* significantly rescued the srpk79Datc mutant phenotype toward wild-type levels (Figure 3G–3J). The presence of axonal Brp accumulations was reduced (Figure 3G and 3H), and there was a correlated increase in synaptic Brp fluorescence in the rescue animals compared to the mutation (unpublished data). Taken together, our data are consistent with the conclusion that loss of srpk79D, in neurons, is responsible for the abnormal accumulation of Brp in peripheral nerves.
Finally, we noted that the srpk79D gene resides just downstream of the gene encoding CSP. In mammals, CSP was recently shown to suppress axonal protein aggregation [14]. Therefore, we pursued additional experiments to determine whether disruption of the Csp gene might participate in the phenotype of Brp axonal accumulation. In these experiments, we took advantage of a strong hypomorphic Csp allele in which the 5′ region of the Csp gene is deleted and the srpk79D locus is intact (CspX1, Figure 1A) [15]. When srpk79Datc was placed in trans to the CspX1 mutation, we found a modest increase in Brp fluorescence and Brp puncta intensity compared to wild type, but not compared to the srpk79Datc/+ heterozygous mutant (Figure 1K and 1L). On the basis of these data, we conclude that Csp is not directly involved in the phenotype of increased axonal Brp puncta staining observed in the srpk79D mutant.
To date, the formation of axonal protein aggregates has been documented in mutations that disrupt both retrograde and anterograde axonal transport [16]–[19]. For example, mutations in kinesin heavy chain and disruption of the Dynein/Dynactin protein complex cause large axonal aggregates composed of diverse synaptic proteins and organelles including, but not limited to, Syt, CSP, Dap160/Intersectin (Dap160) mitochondria, and Brp [16],[17],[19]. Thus, we considered the possibility that the srpk79Datc mutation disrupts axonal transport by asking whether additional synaptic proteins accumulate with Brp in the srpk79Datc mutant axons. We found, however, that the distribution of Syt, CSP, mitochondria, Dap160, and Liprin-alpha were all unchanged relative to wild type in the srpk79Datc mutants (Figure 4A–4L). We also find overexpressed EGFP-CaV2.1 is wild type in the srpk79Datc mutants (unpublished data) [20]. Thus, the srpk79Datc mutation seems to specifically disrupt the transport or aggregation of the Brp protein in peripheral axons without affecting the transport of synaptic vesicles or other AZ constituent proteins.
We next explored the possibility that SRPK79D participates in the specific transport of Brp protein. In recent years, proteins have been identified that are specifically required for the anterograde transport of synaptic proteins or other cellular organelles such as mitochondria [8],[17],[21]. SRPK79D is not strictly required for axonal transport of Brp because Brp protein is present at the NMJ in the srpk79Datc mutant. However, it is possible that SRPK79D facilitates the anterograde transport of Brp/T-bars. Therefore, we pursued genetic interactions between srpk79D and either kinesin heavy chain (Khc) or kinesin 3 (immaculate connections; imac) [8],[16]. Larval nerves that are heterozygous for the amorphic Khc8 allele contain rare “axonal swellings” that contain Syt, CSP, Brp, Dap160, and KHC [16] (E. L. Johnson and G. W. Davis, unpublished data). Importantly, these swellings can be clearly distinguished from the axonal Brp accumulations observed in srpk79D mutants because Brp accumulations in srpk79Datc do not contain any other known synaptic protein. Therefore, we are able to assess whether the presence of a heterozygous Khc or imac mutation would enhance the srpk79Datc mutant phenotype by increasing the abundance of Brp-specific protein aggregates in animals colabeled with anti-Brp and an additional synaptic protein. If SRPK79D facilitates axonal transport of Brp, then reducing KHC or Imac protein levels should enhance the srpk79Datc mutant phenotype (Brp-specific protein aggregates). However, we found that placing a heterozygous Khc8/+ or imac170/+ mutation in an srpk79Datc homozygous mutant background (Khc8/+; srpk79Datc or imac170/+; srpk79Datc) affected neither the frequency nor the severity of the Brp-specific axon aggregates characteristic of the srpk79Datc mutant, nor was there any difference in the axonal swellings characteristic of the Khc mutant (multiprotein aggregates) (Figure 4M and 4N). We then repeated this experiment using numerous additional mutations in the Khc gene as well as other genes implicated in axonal transport including: 1) the antimorphic Khc16 allele [16], 2) Df(3L)34ex5, which deletes the kinesin light chain locus [22], and 3) the amorphic dynein heavy chain at 64C allele, Dhc64C4-19 [23]. We also analyzed double mutants for srpk79D and liprin-alpha (lip-α), an AZ protein shown to play a role in axon transport [24] (lip-αR60/lip-αF3ex5; srpk79Datc). None of these perturbations had any effect on Brp-specific protein accumulations in axons (unpublished data). Finally, through direct observation, we find that small Brp puncta continue to be transported along axons in the srpk79Datc mutant larval nerves, whereas large aggregates appear to be stalled (Figure S3). Taken together, our genetic and live imaging data support the conclusion that Brp accumulations observed in srpk79D mutants are not due to a general defect in axonal transport.
Finally, we asked whether T-bars might be preassembled structures that are trafficked to the NMJ and inserted at the AZ. In some mutant backgrounds, T-bars have been observed to dislodge from the synapse and reside in the cytoplasm [25]. However, we have never observed the appearance of T-bar–like structures in wild-type Drosophila axons at the ultrastructural level (R. D. Fetter and G. W. Davis, unpublished data). This suggests that T-bars are normally assembled at the presynaptic AZ. To examine this question further, we analyzed the size and intensity of anti-Brp puncta in wild-type axons and synapses. At the light level, the vast majority of Brp puncta in wild-type axons are smaller and less intense than the puncta observed within the wild-type presynaptic nerve terminal, suggesting that synaptic T-bars are assembled at the synapse from constituent proteins, including Brp, that are transported down the axon to the synapse (Figure 5). By contrast, Brp puncta observed in srpk79D mutants stained more intensely and were much larger than Brp puncta found in wild-type axons. These Brp puncta were also often larger than the T-bar-associated Brp puncta observed at wild-type NMJ (Figure 5). Thus, the large Brp accumulations found in srpk79D mutant axons could represent superassemblies of T-bar-related proteins, including Brp. To address this possibility, we examined srpk79D mutant axons ultrastructurally.
Mutations that cause focal accumulation of synaptic proteins in Drosophila nerves have been described previously and ultrastructural analyses have been carried out for three of these mutants. In Khc and Dhc64C mutants, axons become dramatically enlarged and are filled with an array of membrane-bound organelles, including multivesicular bodies, prelysosomal vacuoles, and mitochondria [16],[17]. In contrast, lip-α mutant axons have normal diameters and contain organelle accumulations composed predominantly of clear-core vesicles [18].
In the srpk79D mutant, we found that axon diameters were not different from wild type (Figure 6A and 6B). Remarkably, and in contrast to all three of the mutants described above, we found that srpk79D mutant motor axons contained highly organized, electron-dense structures that were not surrounded by a vesicular or intracellular membrane compartment (Figure 6A–6C). Often, these electron-dense structures appeared strikingly similar to T-bars that had been joined at their “bases” into a large T-bar aggregate (Figure 6C and 6D). We have never observed a similar structure in wild-type axons. In this study, we performed electron microscopy on five wild-type animals, analyzing 150 sections from the segmental nerves. None of these sections showed evidence of electron-dense aggregates. We performed electron microscopy on nine srpk79D mutant animals, analyzing 325 sections from segmental nerves. Sections from every mutant animal showed evidence of electron-dense plaques. Nearly every section from an individual mutant showed evidence of electron-dense plaques, consistent with the highly penetrant phenotype observed at the light level. The dimensions of these electron-dense structures, the prevalence of these structures in our electron microscopy sections and the similarity of their shape to T-bars present at the AZ strongly suggest that these structures represent the large Brp aggregates (superassemblies) that we observe at the light level in the srpk79D mutant background. Finally, similar to T-bars found at AZs, these electron-dense structures were surrounded by a filamentous matrix (Figure 6A–6D). Although vesicles were also observed in these areas, we believe that they are molecularly distinct from synaptic vesicles because synaptic vesicle markers do not colocalize with Brp in the srpk79D mutant axons (Figure 4A–4L). In contrast, synaptic ultrastructure in srpk79D mutants is identical to wild type (unpublished data). Thus, loss of srpk79D leads to the formation of T-bar–like superassemblies in axons. Since we have never observed T-bar–like structures in wild-type axons, we propose that SRPK79D is required as part of a mechanism that normally suppresses premature T-bar assembly in the axon.
To gain insight into the subcellular distribution of SRPK79D, we generated Venus-tagged srpk79D transgenes under UAS control (UAS-v-srpk79D) and expressed these transgenes in Drosophila neurons. We found that neuronally expressed Venus-SRPK79D-RD*, which rescues axonal Brp accumulations (Figure 3G–3J), precisely colocalizes with Brp in both the nerve and at each presynaptic AZ (see below). The voltage-gated calcium channel Cacophony is among very few other proteins that have been demonstrated to colocalize with Brp at the presynaptic AZ [20]. Furthermore, Venus-SRPK79D-RD* is highly unusual in that this protein has been shown to colocalize with Brp in a wild-type axon. These data suggest that SRPK79D closely associates with Brp during axonal transport of the Brp protein to the presynaptic nerve terminal. It is possible that the distribution of the tagged-SRPK79D protein does not reflect the wild-type SRPK79D protein distribution. However, the observation that Venus-SRPK79D-RD* shows a very restricted distribution, colocalizing with Brp in at least two cellular compartments, argues that this protein reflects, at least in part, the normal protein distribution.
Given that SRPK79D colocalizes with Brp, we considered two hypotheses for SRPK79D function. First, we considered the hypothesis that SRPK79D somehow influences total Brp protein levels in the cell, perhaps by influencing Brp stability or turnover. However, when we assayed total Brp protein levels by western blot, we found no change in the srpk79D mutant compared to wild type and no evidence of altered protein degradation (Figure S4B and S4C). Although western blots fail to measure Brp protein levels exclusively in motoneurons, this is consistent with our prior observation that axon Brp fluorescence increases while synaptic Brp decreases in the srpk79D mutant, leaving total Brp protein levels constant in the cell. To further examine this possibility, we overexpressed a GFP-tagged brp transgene in otherwise wild-type motoneurons using the GAL4-UAS expression system [4]. Although this resulted in the accumulation of Brp protein within axons, GFP-Brp overexpression did not precisely phenocopy the srpk79D mutant. GFP-Brp expression simultaneously increased synaptic and axonal fluorescence, whereas the srpk79D mutation causes increased axonal Brp and a correlated decrease in synaptic Brp (Figures 1, 2, and 7A–7I). Thus, although Brp overexpression is sufficient to cause axonal aggregates, it seems unlikely that this is the cause of the defect in the srpk79D mutant. Consistent with this conclusion, co-overexpression of BRP and SRPK79D-RD* does not reduce the severity of the axonal accumulations caused by BRP overexpression alone. Similarly, axonal accumulations caused by BRP overexpression are not dramatically enhanced by mutating one copy of the srpk79D gene (Figure 7J). To further address this issue, we asked whether Brp aggregates form in homozygous srpk79Datc mutants in which we decrease total Brp levels by removing one copy of the brp gene (brp69/+;srpk79Datc) [5]. We found that large axonal Brp aggregates persisted even when one copy of the brp gene is removed in the background of the srpk79Datc homozygous mutant (Figure 7D, 7H, and 7I). Taken together, these experiments indicate that an SRPK79D-dependent elevation in Brp protein is not the direct cause of premature T-bar–like assembly formation in the axon. We therefore favor an alternative model based upon the observation that SRPK79D colocalizes with Brp and speculate that SRPK79D could sequester or inhibit the function of axonal T-bar proteins and thereby prevent the formation of axonal T-bar–like superassemblies.
As mentioned above, sequence analysis of the predicted srpk79D gene products reveals similarity to a group of serine-threonine kinases called SRPKs (Figure 8B). Members of this protein kinase family share a characteristic split serine-threonine kinase domain [26]. We therefore performed experiments to determine whether the kinase domain is required for SRPK79D activity. Transgenically expressed full-length SRPK79D (SRPK79D-RD*) colocalizes with Brp and rescues the srpk79D mutant phenotype (Figure 8C–8K). In contrast, expression of an SRPK79D isoform with a truncated kinase domain (SRPK79D-RD) colocalized with Brp, but failed to rescue the srpk79D mutant phenotype (Figure 8B and 8K). These data indicate that the SRPK79D kinase domain is involved in preventing axonal superassemblies of Brp but that it is not important for colocalization with Brp. To further test the importance of SRPK79D kinase activity, we generated a kinase dead srpk79D transgene by introducing a missense mutation into SRPK79D-RD* that is predicted to disrupt the ATP binding pocket of the kinase domain and thereby inhibit kinase activity (SRPK79D-RD*KD, Figure 8B). A similar strategy has been used previously to eliminate kinase activity in other SRPKs ranging from yeast to human [27]–[29]. Like SRPK79D-RD*, SRPK79D-RD*KD colocalized with Brp. However, even when expressed at higher levels than SRPK79D-RD*, SRPK79D-RD*KD failed to rescue the srpk79D mutant phenotype (Figure 8K and unpublished data).
We next asked which domains might be required for SRPK79D protein trafficking and/or localization. We have found that an SRPK79D transgene that possesses an alternative SRPK79D N-terminal region but is otherwise identical to SRPK79D-RD* (SRPK79D-RB) failed to be efficiently trafficked out of the neuronal soma, was not found to colocalize with Brp, and failed to rescue the srpk79D mutant phenotype (Figure 8B and 8K and unpublished data). This suggests that the common N-terminal domain of SRPK79D-RD, SRPK79D-RD*, and SRPK79D-RD*KD is required for the axonal transport of SRPK79D and its colocalization with Brp.
Our data are consistent with a model in which SRPK79D prevents premature assembly of T-bars within axons. This model also suggests that SRPK79D activity must be inhibited locally, at the AZ, in order for synaptic T-bar assembly to proceed. We reasoned that overexpressing SRPK79D might overwhelm the synaptic machinery that disrupts SRPK79D activity and thereby reveal a role for SRPK79D during T-bar assembly or synaptic function. Here, we show that SRPK79D overexpression disrupts the punctate, highly organized appearance of synaptic Brp immunoreactivity (Figures 9A–9D and S5A–S5D). For example, we observed regions where Brp was diffusely organized near the synaptic membrane. These regions encompass areas that would normally contain several individual Brp puncta. We hypothesize that these regions of diffuse Brp reflect either failed T-bar assembly or severely perturbed AZ organization. In addition, we found that SRPK79D overexpression also led to a decrease in total synaptic Brp fluorescence (Figure 9E). This might be consistent with perturbed AZ formation but is also similar to that found in homozygous srpk79D mutants (srpk79Datc), mutants heterozygous for a null mutation in brp (brp69/+), and mutants heterozygous for the brp null mutation and homozygous for the srpk79Datc allele (brp69/+; srpk79Datc; Figures 2 and 7, and unpublished data). It should be noted, however, that the diffuse synaptic Brp staining caused by SRPK79D overexpression is not observed in any of these srpk79D or brp loss-of-function paradigms. It should be further noted that SRPK79D levels in this overexpression experiment are higher than the SRPK79D levels that are sufficient to rescue the srpk79D mutation (Figures 3G–3J, 9, and S5). SRPK localization was determined in rescue animals expressing relatively low levels of transgene-derived SRPK79D, and we believe that this is why we observe normal synaptic architecture and SRPK79D localization in those experiments (Figure 8C–8F). Finally, overexpression of SRPK79D-RD*KD (kinase dead) or SRPK79D-RD (truncated kinase domain) did not cause diffuse Brp staining (unpublished data), indicating that the kinase domain is required for this phenotype.
If SRPK overexpression perturbs T-bar assembly or organization, then we might expect a disruption of presynaptic vesicle release. When we assayed synaptic function in larvae overexpressing SRPK79D, we found a dramatic (∼50%) decrease in excitatory postsynaptic potential (EPSP) amplitude along with a trend toward an increase in the average amplitude of spontaneous miniature events (minis, Figure 9F and 9G). Estimating the average number of vesicles released per action potential (quantal content; calculated according to the average EPSP/average mEPSP per NMJ), we found that quantal content was severely perturbed. Since synapse function is intact in srpk79Datc homozygous animals, brp69/+ heterozygous animals, and brp69/+; srpk79Datc double-mutant larvae (see above), the defects caused by SRPK79D overexpression are likely a consequence of excess SRPK79D activity at AZs. In addition, overexpression of SRPK79D-RD*KD (kinase dead) or SRPK79D-RD (truncated kinase domain) did not cause a defect in synaptic function (unpublished data) indicating that the kinase domain is required for this overexpression phenotype. Finally, it is worth noting that the defects in synaptic function caused by SRPK79D-RD* overexpression are similar to those found in brp null mutants, which lack T-bars [5].
Here, we present the identification and characterization of a novel serine-threonine kinase termed Serine-Arginine Protein Kinase at 79D (SRPK79D) that colocalizes with the T-bar-associated protein Brp in both the axon and at the mature AZ. SRPK79D is one of very few proteins known to localize to T-bars or ribbon-like structures at the AZ and is the only known kinase to localize to this site [1]–[5] (Figure 8). We further provide genetic evidence that SRPK79D functions to represses the premature assembly of T-bars in axons. In particular, we show that loss-of-function mutations in srpk79D cause the appearance of T-bar–like protein aggregates throughout peripheral axons, and we are able to rule out the possibility that this is an indirect consequence of impaired axonal transport (Figures 4 and 6). The appearance of ectopic T-bars is highly specific since numerous other synaptic proteins and mitochondria are normally distributed in the neuron and are normally trafficked to the presynaptic nerve terminal in the srpk79D mutant background (Figure 4A–4L). Thus, SRPK79D appears to have a specific function in repressing T-bar assembly prior to the AZ, consistent with the strong colocalization of SRPK79D protein with Brp and T-bar structures.
Finally, we also uncover a potential function for SRPK79D at the AZ where it is observed to colocalize with Brp. SRPK79D loss-of-function mutations do not alter the number, density, or organization of Brp puncta at the synapse and do not alter synaptic function (Figure 2 and unpublished data). This is consistent with a negative regulatory role for SRPK79D during T-bar assembly and indicates that once SRPK79D-dependent repression of T-bar assembly is relieved, AZ assembly proceeds normally. Overexpression of SRPK79D, however, severely disrupts neurotransmission. The defect in presynaptic release is correlated with a disruption of Brp puncta organization and integrity. These phenotypes are consistent with a function for SRPK79D as a negative regulator of T-bar assembly and AZ maturation.
SRPK79D is a member of the SRPK family of constitutively active cytoplasmic serine-threonine kinases that target serine-arginine–rich domains of SR proteins [26]–[29]. Thus, it is interesting to postulate what the relevant kinase target might be. Given that SRPK79D and Brp colocalize, an obvious candidate is the Brp protein itself. However, the Brp protein does not have a consensus SR domain, and decreasing the genetic dosage of srpk79D does not potentiate axonal Brp accumulations that appear upon Brp overexpression [4] (Figure 7J). As such, Brp may not be the direct target of SRPK79D kinase activity. We hypothesize, therefore, that SRPK79D colocalizes with Brp and another putative SR protein that is the direct target of SRPK79D kinase activity.
The best-characterized role for SRPKs is in controlling the subcellular localization of SR proteins, thereby regulating their nuclear pre-mRNA splicing activity [12]. More recently, SR protein involvement in several cytoplasmic mRNA regulatory roles has been reported [30],[31]. In particular, a phosphorylation-dependent role for SR proteins has been reported in both Drosophila and mammalian cell culture [32],[33].
It is interesting to speculate that the function of SRPK79D to prevent premature T-bar assembly might be related to the established function of SRPKs and SR-domain-containing proteins during RNA binding, processing, and translation [12],[30]. One interesting possibility is that RNA species are resident at the T-bar. In such a scenario, SRPK79D-dependent repression of RNA translation could prevent T-bar assembly in the axon, and relief of this repression would enable T-bar assembly at the AZ. The continued association of SRPK79D with the AZ could allow regulated control of further T-bar assembly during development, aging, and possibly as a mechanism of long-term synaptic plasticity. Several results provide evidence in support of such a possibility. First, local translation has been proposed to control local protein concentration within a navigating growth cone [34],[35]. There is also increasing evidence in support of local translation in dendrites and for the presence of Golgi outposts that could support local protein maturation [36],[37]. A specific role for RNA binding proteins at the presynaptic AZ is supported by the prior identification of the RIBEYE protein, which is a constituent of the vertebrate ribbon structure. RIBEYE contains a CtBP domain previously shown to bind RNA [2]. The discovery of a different RNA binding protein (CtBP1) at the ribbon and our description of a putative RNA regulatory protein at the Drosophila T-bar further suggest that RNA processing might be involved in the formation or function of these presynaptic electron dense structures [3].
In light of these data, we explored the possibility that SRPK79D might participate in translational control related to T-bar assembly. We, therefore, examined mutations in genes that could represent SRPK79D-dependent negative regulators of translation, such as aret (bru), cup, pum, nos, and sqd [38]–[44], reasoning that the loss of such a translational inhibitor might result in the ectopic synthesis of AZ proteins, ultimately leading to a phenotype similar to that observed in srpk79D mutants. We also generated genomic deletions for bru2 and bru3. However, we did not find evidence of axonal Brp aggregation in any of these mutants. Next, we assayed mutations previously shown to be required for mRNA transport and local protein synthesis. If necessary for T-bar assembly, these mutations might disrupt synaptic Brp-dependent T-bar formation. These mutations, including orb, vas, and stau, have phenotypes at earlier stages of development, but show no defect in synaptic Brp staining [38],[45]–[48]. Thus, although these experiments do not rule out a function for SRPK79D in local translation, we have examined mutations in several additional candidates and failed to uncover evidence in support of this model.
Another possibility is that SRPK79D inhibits T-bar assembly through the constitutive phosphorylation-dependent control of a putative SR protein that colocalizes with SRPK79D and Brp within a nascent T-bar protein complex. Upon arrival of this nascent T-bar protein complex at the presynaptic nerve terminal, T-bar assembly could be initiated in a site-specific manner through the action of a phosphatase that is concentrated at a newly forming synapse. There are several examples of phosphatases that can be localized to sites of intercellular adhesion, some of which have been implicated in the mechanisms of synapse formation and remodeling [49]. This model, therefore, proposes that negative regulation of T-bar assembly, via SRPK79D, is a critical process required for the rapid and site-specific assembly of the presynaptic AZ-associated T-bar structure. Finally, we can not rule out the possibility that SRPK79D normally functions to prevent T-bar superassembly as opposed to T-bar assembly per se. Consistent with this idea is the observation of T-bar aggregates in axons and prior observation that detached ribbon structures coalesce into large assemblies in vertebrate neurons [50].
Synapse assembly is a remarkably rapid event. There is evidence that the initial stages of synapse assembly can occur in minutes to hours, followed by a more protracted period of synapse maturation [11],[51]–[53]. Synapses are also assembled at specific sites. In motoneurons and some central neurons, synapses are assembled when the growth cone reaches its muscle or neuron target [53],[54]. However, many central neurons form en passant synapses that are rapidly assembled at sites within the growing axon, behind the advancing growth cone [53],[54]. Current evidence supports the conclusion that intercellular signaling events mediated by cell adhesion and transmembrane signaling specify the position of the nascent synapse [54]–[56]. The subsequent steps of presynaptic AZ assembly remain less clear. Calcium channels and other transmembrane and membrane-associated proteins appear to be delivered to the nascent synaptic site via transport vesicles that fuse at the site of synapse assembly [9]–[11]. It has been proposed that cytoplasmic scaffolding molecules then gradually assemble at the nascent synapse by linking to the proteins that have been deposited previously [11]. This model assumes, however, that the protein–protein interactions between the numerous scaffolding molecules that comprise the presynaptic particle web do not randomly or spontaneously occur in the cytoplasm prior to synapse assembly. What prevents these scaffolds from spontaneously assembling in the small volume of an axon, prior to synapse formation at the nerve terminal and between individual en passant synapses? Currently, nothing is known about how premature scaffold assembly is prevented. We propose that our studies of srpk79D identify one such mechanism of negative regulation that prevents premature, inappropriate assembly of a presynaptic protein complex. We further propose that such a mechanism of negative regulation, when relieved at a site of synapse assembly, could contribute to the speed with which presynaptic specializations are observed to assemble.
The listed strains were obtained from the following sources: srpk79D[atc] (c00270), f00171, d09582, f05463, and d09837 from the Exelixis collection at Harvard Medical School; v47544 (UAS-srpk79DRNAi) from the Vienna Drosophila RNAi Collection; P{GawB}elavC155 (C155), P{GawB}sca109-68 (Sca), P{GawB}OK371 (OK371), P{GAL4}repo (Repo), Khc8, Khc16, Df(3L)34ex5; dhc64C4-19, Df(3L)Exel6138, UAS-mitoGFP, cup1, sqdj4b4, pum13, nosL7, vasRJ36, orbdec, stau1, and staury9 from the Bloomington Stock Center; CspX1 was a generous gift from Konrad Zinsmaier; srpk79DVN100 was a generous gift from Erich Buchner; imac170 was a generous gift from Thomas Schwarz; aretPA, aretPD, and aretQB were generous gifts from Paul MacDonald; and UAS-gfp-brp (UAS-g-brp) and brp69 were generous gifts from Stephan Sigrist.
Wandering third-instar larvae were dissected in calcium-free saline and fixed with either 4% paraformaldehyde/PBS (15 min) or 100% Bouin's Solution (2 min). Excess fixative was removed by extensive washing in PBS+0.1% Triton-X (PBT). Dissected larvae were then incubated overnight at 4°C in PBT with one or more primary antibodies, washed in PBT, incubated either overnight (4°C) or for 1 h (22°C) in PBT with one or more fluorescent-conjugated secondary antibodies, and washed again before being mounted on a slide for imaging analysis. Primary antibodies: NC82 (anti-Brp; Developmental Studies Hybridoma Bank) 1∶100; 3H2 2D7 (anti-Syt; Developmental Studies Hybridoma Bank) 1∶25; anti-Liprin-alpha (a generous gift from David Van Vactor) 1∶1,000; 1G12 (anti-DCSP-3; Developmental Studies Hybridoma Bank) 1∶25; and anti-Dap160 (Marie et al., 2004 [57]) 1∶100. Fluorescent-conjugated secondary antibodies: goat-anti-mouse Alexa 488 (Invitrogen) 1∶500; goat-anti-mouse Alexa 555 (Invitrogen) 1∶500; and goat-anti-rabbit Alexa 488 (Invitrogen) 1∶500. Where applicable, anti-HRP-Cy3 (Jackson Immunoresearch) 1∶200; anti-HRP-FITC 1∶100 or anti-HRP-Cy5 1∶50 were used at the same step as secondary antibody incubation. Genotypes being directly compared were grouped together during all of the above procedures.
Images were digitally captured using a cooled CoolSnapHQ CCD camera mounted on a Zeiss Axiovert 200 M microscope. Images were acquired and analyzed using Slidebook software (Intelligent Imaging Innovations). Individual nerves/synapses were optically sectioned at 0.5 µm (11–27 sections per nerve) using a piezoelectric-driven z-drive controlling the position of a Zeiss 100× oil immersion objective (numerical aperture [NA] = 1.4). The intensity of anti-BRP immunostaining was quantified as follows: Each series of 0.5-µm optical nerve sections was deconvolved (nearest-neighbors; Intelligent Imaging Innovations). Two-dimensional projections of the maximum pixel intensity were then generated, and the total Brp fluorescence and the maximum fluorescence intensity of each Brp punctum within the nerve/synapse area were determined for each resulting image using a semiautomated procedure as described previously [58],[59]. For all quantifications, the nerve/synapse area was defined as that delimited by anti-HRP staining.
Live imaging was carried out as previously described [60]. In brief, wandering third-instar larvae were dissected in HL3 saline (0.4 mM Ca2+) on a glass coverslip and held in place using pressure pins. Images were digitally captured using a Photometrics Cascade 512B camera mounted on an upright Zeiss Axioskop 2 microscope using a 100× water immersion (NA = 1.0) objective and a GFP filter set (Chroma). Time-lapse images were collected and analyzed using Slidebook software (Intelligent Imaging Innovations).
srpk79D mRNA was detected using a protocol based upon the “96-well plate RNA in situ protocol” available at the Berkeley Drosophila Genome Project (BDGP) Web site (http://www.fruitfly.org). In short, mixed-stage embryos were collected, fixed in 3.7% formaldehyde/1×PBS, and prepared for incubation with SP6 or T7 polymerase generated digoxigenin (DIG)-labeled nucleotide probes. To generate probes, a 954-base pair (bp) fragment of the srpk79D gene was amplified by PCR from cDNA AT02150, obtained from the Berkeley Drosophila Genome Project using primers with the sequence 5′-ttacccggattcgtccgac-3′ and 5′-gcagtgattttcttctccgttcgg-3′. This fragment was TA cloned into the pGEM-T Easy vector (Promega). The resulting product was used as a template for T7/SP6 DIG-labeled RNA probe synthesis (Roche). After incubation and removal of excess probe, embryos were incubated with alkaline-phosphatase-conjugated anti-DIG Fab fragments (Roche). Excess Fab fragments were removed by washing, and a NBT/BCIP developing reaction was performed (Roche).
Adult heads were removed by freezing at −70°C, followed by agitation. Heads were isolated using mesh filters. RNA was extracted using TRIzol reagent and standard molecular biology techniques. DIG-labeled RNA probes were generated by amplifying an 800-bp fragment of the brp gene from cDNA IP09541 obtained from the Berkeley Drosophila Genome Project using primers with the sequence 5′-gcaatgggcagtccatactacc-3′ and 5′-cccattcccttggcctgc-3′ and the 738-bp insert from rp49 cDNA RE59709 obtained from the Berkeley Drosophila Genome Project and 5′-cggcaaggtatgtgcg-3′ and 5′-actaaaagtccggtatattaacgtttac-3′ and TA cloning into pGEM-T Easy (Promega). The resulting product was used as a template for T7/SP6 DIG-labeled RNA probe synthesis (Roche). Northern blot analysis was carried out using Ambion NorthernMax-Gly protocols and reagents. Probe detection was carried out using alkaline phosphatase-conjugated anti-DIG Fab fragments (Roche) in conjunction with the DIG Wash and Block Kit and CSPD Ready-to-Use.
Third-instar larval brains were pulverized in 2× Laemmli sample buffer. Proteins were separated by SDS-PAGE and transferred to PVDU membrane. The membrane was blocked in 2% milk powder in 1×TBS-Tween, and then incubated for 1 h at room temperature with an anti-Brp monoclonal antibody (Developmental Studies Hybridoma Bank NC82, 1∶100) or anti-GFP monoclonal antibody (Invitrogen 3E6, 1∶100). As a protein loading control, the membrane was co-incubated with an anti-β-tubulin monoclonal antibody (Developmental Studies Hybridoma Bank E7, 1∶1,000). After washing in 1×TBS-Tween, the membrane was incubated for 1 h at room temperature with horseradish peroxidase-conjugated anti-mouse secondary antibody (1∶20,000), washed again and an electrogenerated chemiluminescence (ECL) detection reaction (Amersham) was performed.
Mutant and wild-type third-instar larvae were prepared for electron microscopy as follows. Larvae were filleted in physiological saline and fixed with 2% glutaraldehyde in 0.12 M Na-cacodylate buffer (pH 7.4, 10 min). The fixed larvae were then transferred to vials containing fresh fixative and fixed for a total of 2 h with rotation. Larvae were rinsed with 0.12 M Na-cacodylate buffer and postfixed with 1% osmium tetroxide in 0.12 M Na-cacodylate buffer for 1 h. Specimens were then rinsed with 0.12 M Na-cacodylate buffer, followed by water, and then stained en bloc with 1% aqueous uranyl acetate for 1 h. After water rinse, dehydration, and embedding in Eponate 12 resin, sections were cut with a Leica Ultracut E microtome, collected on Pioloform-coated slot grids, and stained with uranyl acetate and Sato's lead. Sections were photographed with a Tecnai spirit operated at 120 kV equipped with a Gatan 4 k×4 k camera.
Recordings were taken in HL3 saline (Ca2+ 0.4 mM, Mg2+ 10 mM) from muscle 6 in abdominal segments 2 and 3 of third-instar larvae as previously described [61]. Only recordings with resting membrane potentials more negative than −60 mV and input resistances greater then 7 MΩ were used for analysis. Measurements of EPSP and spontaneous miniature release event amplitudes were made using MiniAnalysis software (Synapsoft).
|
10.1371/journal.pcbi.1002317 | Switches, Excitable Responses and Oscillations in the Ring1B/Bmi1 Ubiquitination System | In an active, self-ubiquitinated state, the Ring1B ligase monoubiquitinates histone H2A playing a critical role in Polycomb-mediated gene silencing. Following ubiquitination by external ligases, Ring1B is targeted for proteosomal degradation. Using biochemical data and computational modeling, we show that the Ring1B ligase can exhibit abrupt switches, overshoot transitions and self-perpetuating oscillations between its distinct ubiquitination and activity states. These different Ring1B states display canonical or multiply branched, atypical polyubiquitin chains and involve association with the Polycomb-group protein Bmi1. Bistable switches and oscillations may lead to all-or-none histone H2A monoubiquitination rates and result in discrete periods of gene (in)activity. Switches, overshoots and oscillations in Ring1B catalytic activity and proteosomal degradation are controlled by the abundances of Bmi1 and Ring1B, and the activities and abundances of external ligases and deubiquitinases, such as E6-AP and USP7.
| The generation of polyubiquitin chains on target proteins as a degradation signal was a landmark discovery rewarded by the 2004 Nobel Prize in Chemistry. However, emerging evidence suggests that protein ubiquitination is more versatile. Different types of ubiquitin chains serve numerous non-proteolytic functions, among them regulation of the biological activities of target proteins. Here we demonstrate a flexible role of ubiquitination in the dynamic control of Ring1B, a ubiquitin ligase that monoubiquitinates histone H2A, which in turn silences gene expression. Remarkably, Ring1B increases its own activity by self-ubiquitination. A binding partner of Ring1B, Bmi1, facilitates Ring1B self-ubiquitination and protects both proteins from rapid degradation. We use computational modeling to show that the Ring1B/Bmi1 system can act as analog-digital converter, generating abrupt switches, multistable dynamics, oscillations and excitable overshoots. For instance, an increase in Bmi1 abundance brings about an abrupt “On” switch of Ring1B monoubiquitinating activity and downregulation of H2A-controlled genes, while a decrease in Bmi1 leads to an “Off” switch. These digital responses can display hysteresis, creating the biological memory. Distinct types of Ring1B activity responses (oscillatory, bistable and excitable) facilitate signal discrimination and allow the Ring1B/Bmi1/H2A system to distinctly affect gene silencing and potentially trigger different cell fates.
| Recent discoveries have revolutionized our perception of the role of protein ubiquitination in signaling networks. Although initially ubiquitination was considered as a signal for proteasomal degradation, emerging evidence suggests that different types of ubiquitin chains may have non-proteolytic roles and can dramatically alter the biological activities of a target protein [1]. Early work showed that a polyubiquitin chain consisting of at least four ubiquitin molecules, which are linked through Lys48 (K48) initiates the rapid degradation of a target protein by the ubiquitin-proteasome system (UPS) [2]. Later, non-degradative roles of K63-linked oligo- and polyubiquitin chains were found for proteins involved in the DNA-damage response, the JNK, p38 MAPK and NF-κB signaling pathways, and endocytic trafficking [1]. Recently, atypical, branched ubiquitin chains that involve K6/K27/K48 ubiquitin linkages were discovered on the E3 ligase Ring1B that monoubiquitinates histone H2A. Interestingly, these atypical ubiquitin chains were generated only by Ring1B self-ubiquitination (also referred to as auto-ubiquitination), whereas Ring1B ubiquitination by the E3 ligase E6-AP (E6-associated protein) resulted in canonical K48 linkages [3].
The E3 ligase Ring1B is a RING finger protein, which interacts with another RING finger protein, Bmi1. Together with Polyhomeotic 1 and Chromobox protein homologue 4, Ring1B and Bmi1 form the core human Polycomb transcriptional Repressive Complex 1 (PRC1), which plays a critical regulatory role in the control of genes during development, ageing and cancer [4], [5], [6]. Owing to Ring1B catalytic activity, PRC1 is a major E3 ligase of histone H2A in vivo. Monoubiquitinated H2A (uH2A) represses transcriptional initiation and elongation [7], [8], [9], leading to gene silencing that was implicated in tumorigenesis and stem cell development [4], [6], [10], [11]. Increased monoubiquitination of H2A was observed upon UV radiation in mammalian cells, implying a role of uH2A in the DNA damage response and/or DNA repair-induced chromatin remodelling [12], [13]. By contrast, uH2A deubiquitination was found to facilitate cell cycle progression where repressive histone marks are removed during G0-G1/S transition to allow S-phase gene expression [14], [15]. Thus, Ring1B-induced H2A monoubiquitination (and subsequent deubiquitination) plays an essential role in regulating gene expression.
Both Ring1B and Bmi1 are short-lived proteins, which are degraded by UPS. It has long been understood that self-ubiquitination of RING finger-containing E3 ligases targets them for UPS-mediated destruction [16], [17]. Surprisingly, recent work reveals that degradation of Ring1B is independent of its self-ubiquitinating activity [18]. Self-ubiquitination of Ring1B generates branched K6/K27 ubiquitin chains, and this is required for efficient in vitro monoubiquitination of histone H2A, whereas canonical K48-linked chains, generated by other ligases target Ring1B for degradation. The presence of Bmi1 greatly facilitates Ring1B monoubiquitinating activity with respect to H2A, and the association between Ring1B and Bmi1 protects these proteins from rapid degradation [18].
Similar to protein phosphorylation/dephosphorylation, ubiquitination is reversed by the opposing process of deubiquitination. Ubiquitination chains with distinct linkages and structures are recognized and deubiquitinated by different deubiquitinases (DUBs) that feature specialized ubiquitin-binding domains. Studies of protein phosphorylation have shown that upon small changes in input kinase or phosphatase activities, target proteins can abruptly switch between distinct phosphorylation states, a phenomenon termed “ultrasensititvity” [19]. Moreover, phosphorylation on two or more residues not only increases ultrasensitivity, but potentially leads to bistability or multistability where under the same input conditions, a target protein can reside in any of two or more stable stationary states with different phosphorylation levels [20], [21]. Recently, it has been shown that intermolecular auto-phosphorylation, a salient feature of activation of many protein kinases [22], [23], [24] can bring about the intricate dynamic behavior that involves abrupt activity switches, bistability and hysteresis [25]. Interestingly, when the phosphorylation dynamics are bistable, multiplicity of deactivation routes can result in sustained, pulsatory oscillations in kinase activities [25].
Ubiquitination reaction circuitry is more complex than (de)phosphorylation cycles. Two enzymes, ubiquitin-activating (E1) and ubiquitin-conjugating (E2) enzymes, are involved in every ubiquitin molecule transfer to a target protein by ligase E3, and ubiquitin molecules can form polymeric chains of different structures, which is not the case for phosphorylation. Yet, there is a striking similarity between intermolecular auto-phosphorylation of protein kinases and auto-ubiquitination that results in self-activation or self-inhibition of E3 ligases. For instance, auto-ubiquitination of Itch ligase was shown to be an intermolecular reaction, which generated K63-linkages, rather than the K48-linked chains that target Itch for proteasomal degradation [26]. Likewise, self-ubiquitination of the HECT-type E3 ligase Nedd4 leads to better recognition and higher rate of monoubiquitination of Eps15 by Nedd4 in the EGFR internalisation and degradation pathway [27], whereas auto-ubiquitination of DIAP1(Drosophila inhibitor of apoptosis protein 1), an E3 ligase responsible for cell death regulation in Drosophila, serves to attenuate DIAP1 ligase activity towards its substrates (such as the proapoptotic protein Rpr) via formation of K63-linkages rather than K48-based polyubiquitin chains [28]. In vitro data suggest that two Ring1B molecules can self-dimerize via their C-terminal domains [29], reiterating the possibility of intermolecular auto-ubiquitination.
The present paper shows that extremely complex Ring1B activity and degradation dynamics (that results in the intricate temporal control of H2A monoubiquitination) can be brought about by the Ring1B - Bmi1 interaction circuitry, intermolecular Ring1B auto-ubiquitination, and Ring1B (de)ubiquitination by external ligases and DUBs. Using computational modelling to elucidate these dynamics, we demonstrate that the Ring1B/Bmi1/H2A network can display oscillatory, bistable and excitable behaviors. We show that overexpression (or mutation) of Bmi1 and a deubiquitinating enzyme USP7 do not merely change the amplitude of Ring1B degradation and catalytic rates, but dramatically transform their response dynamics. For instance, an increase in Bmi1 abundance can bring about bistable, all-or-none Ring1B monoubiquitination activity and bistable, all-or-none expression of H2A-controlled genes. Under the proper conditions, which include the upregulation of USP7 and Bmi1, self-perpertuating oscillatory responses of Ring1B monoubiquitination are facilitated. In the proximity of oscillatory regimes, the Ring1B/Bmi1 system displays an excitable behavior where a transient perturbation causes Ring1B activity and degradation rate to overshoot before returning to the basal level. Our findings unveil the intrinsic complexity of the dynamics of Ring1B activity and H2A monoubiquitination and allow for direct experimental testing.
We developed a computational model that encapsulates key molecular interactions and reveals intricate dynamic behaviors of the Ring1B/Bmi1/H2A network. The model is based on a careful examination of all available biological data and accounts for distinct, Bmi1-dependent and independent modes of Ring1B self-induced ubiquitination, Ring1B and Bmi1 ubiquitination by external ligases and monoubiquitination of histone H2A by catalytically active forms of Ring1B (Fig. 1). We modelled the system on two different timescales: (i) a short timescale (<1 hour) where degradation reactions occur but can be neglected, as the protein abundances in the system have not yet changed much; and (ii) a long timescale (>1 hour), where protein synthesis and degradation are explicitly considered in the model. We then show that the inclusion of protein synthesis and degradation rates does not practically change the network dynamics observed on short timescales, where the key biochemical modifications take place that precede and in part determine changes in protein degradation.
Below, we describe underlying core biochemical mechanisms and key biological observations which led to the assumptions built into the model.
A commonly observed behavior of protein ubiquitination is a gradual approach to a (quasi-) stationary state featuring specific ubiquitination levels of targets. This behavior can also be observed for the Ring1B/Bmi1/H2A network that displays single stable steady states for a wide array of kinetic parameter values. For instance, when the Ring1B/Bmi1 ubiquitination assays have been carried out in a cell-free reconstituted system, the system quickly relaxes to (quasi-) steady state conditions [18]. Yet, in certain parameter ranges, the Ring1B system exhibits intricate dynamic behavior that may be exploited by cells to efficiently control H2A monoubiquitination and gene expression.
Although the characteristic timescales of the (de)activation of the Ring1B/Bmi1 system in live cells have not been well documented, the available data suggest that the post-translational, (de)ubiquitination dynamics is much faster than the rates of protein synthesis and degradation. (De)ubiquitination typically occurs on the timescale of seconds to minutes [41], [42], whereas Ring1B/Bmi1 degradation and in vivo synthesis evolve on the timescale of several hours [18]. Yet, it is still unknown what the specific signaling inputs to the Ring1B/Bmi1/H2A system in vivo are, and whether these inputs are transient or sustained. Therefore, our model largely focuses on the post-translational modification dynamics of the Ring1B/Bmi1/H2A system on the relatively short timescale within an hour after abrupt changes in the input. As we will see below, these input changes can result in digital “on” or “off” H2A monoubiquitination responses. We subsequently examine the Ring1B/Bmi1/H2A dynamics on the long timescale taking into account protein synthesis and degradation and show that during the first hour, a long timescale model behaves practically indistinguishable from a short-scale model that neglects protein synthesis and degradation.
Ring1B-dependent monoubiquitination of histone H2A, enhanced by Bmi1, is an essential mechanism of Polycomb-mediated gene silencing. Monoubiquitinated histone H2A is involved in the initiation and maintenance of the silenced state of PRC1 target genes. To understand the temporal dynamics of H2A-directed gene silencing, it is crucial to understand dynamics of the Ring1B/Bmi1 system and H2A monoubiquitination. The present paper shows that complex dynamic behaviors can be brought about by the intrinsic circuitry of Ring1B-Bmi1 interactions, auto-ubiquitination of Ring1B and its ubiquitination and deubiquitination by other ubiquitin ligases and DUBs. Using computational modelling to elucidate these dynamic properties, we demonstrate that the Ring1B/Bmi1 system can act as analog-digital converter, generating abrupt switches, multistable dynamics, oscillations and overshoots. Distinct types of responses facilitate signal discrimination and allow the Ring1B/Bmi1 system to differentially affect gene silencing, which may trigger different cell fates.
We show that overexpression or mutation of Bmi, other ubiquitin ligases and DUBs do not merely change the amplitude of responses to external stimuli, but can dramatically transform the response dynamics. For instance, upregulation of Bmi1 abundance can move the system from a monostable regime prevailing at low Bmi1 levels into more complex regimes exhibiting self-perpertuating oscillatory or bistable responses. Under conditions where the concentrations of Bmi1 and DUBs deubiquitinating Ring1B (e.g. USP7) are high, oscillatory responses are promoted; whereas underexpression of USP7 confers the system more prone to bistable responses.
We show that these complex dynamics of the Ring1B system arise from positive feedback loops brought about by intermolecular self-induced ubiquitination of Ring1B combined with the saturable kinetics of the Ring1B deubiquitinating reactions. Interestingly, additional mechanisms exist in the Ring1B/Bmi1 system, which can give rise to bistable and oscillatory responses. For example, three or more ubiquitination/deubiquitination cycles occurring during the formation of polymeric ubiquitin chains, can bring about bistable behavior provided different ubiquitination forms compete for the same ligase or DUB [21]. Although we confined our analysis to bistable behavior, relaxing the assumption about the first order DUB kinetics brings about multistable behavior. In fact, when both steps 7 and 10 are saturable, multistable steady states can occur in the Ring1B/Bmi1 system. In this case, up to three stable states can be observed for a given total Bmi1 concentration (Fig. S6). However, if positive feedback loops are absent, only abrupt switches, but not bistability can be observed (Fig. S7).
When protein synthesis and degradation rates are explicitly taken into account by our long-timescale model, the intricate dynamic features of the Ring1B/Bmi1 system discussed above, including bistable and oscillatory behaviors, remain the same on short timescales (Fig. S8). As can be seen in Fig. S8, on the timescale up to one hour, a model that describes the long-term dynamics behaves practically indistinguishable from a short-timescale model that neglects protein synthesis and degradation, and both bistable (Figs. S8a) and oscillatory (Figs. S8b) responses are observed. However, on the long timescale (>>1 hr), complex dynamics such as bistability and oscillations might not be exhibited due to the effect of protein synthesis and degradation, and the system would approach a stable steady state after more than 10 hrs (at selected parameter values). The timescale of the experiments exploring the post-translational, (de)ubiquitination dynamics of the Ring1b/Bmi1 system was much less than an hour, which is significantly shorter that the timescale associated with Ring1B/Bmi1 synthesis and degradation (about 3 hour half-life for Ring1B and Bmi1 and ∼7–8 hour half-life for the Ring1B-Bmi1 complex) [18]. Accordingly, we have mainly focused on the one hour timescale to account for the data on different ubiquitinated forms of Ring1B and Bmi1 that relaxed to (quasi-) stationary concentrations [18]. Further experimental testing of the model requires the kinetic monitoring of different forms of ubiquitin chains that control Ring1B activity or target it for degradation. However, at present, there are no reagents that can discern between these different linkage types. We were able to obtain experimental data that support some modelling predictions, but are not fully conclusive to claim that the predicted dynamics (switches, oscillations or excitable response) are realized in the tested cells (Supplementary Text S1, section 6, and Figs. S12, S13, S14). Although we are still carrying out further experiments, the more complete verification of model predictions requires extensive time and effort that goes beyond the scope of the current paper. Because this is the first mathematical model of the Ring1B/Bmi1 ubiquitination system, our main objective is to draw attention to a rich repertoire of dynamical behaviors that the ubiquitination system can exhibit.
The results of the present study shed light on recent experimental findings related to concentrations of Bmi1 and ubiquitinated histone H2A in stem cells, tumour cells and cells undergoing differentiation. Among the PRC1 component proteins, Bmi1 has been demonstrated to be strongly involved in multiple biological processes including tumorigenesis, stem cells self-renewal, and differentiation [6]. Overexpression of Bmi1 is frequently observed in various types of human cancers, including lung cancer, ovarian cancer, acute myeloid leukemia, nasopharyngeal carcinoma, and neuroblastoma [56], [57], [58], [59], [60]. This oncogenic property of Bmi1 has been linked to its ability to protect cells from apoptosis through suppressing the expression of tumour suppressor and pro-apoptotic genes. For examples, in Bmi1-deficient mice the number of lymphocytes is significantly reduced due to increased apoptosis [61]. Expression of Bmi1 in stem cells leads to the silencing of the tumour suppressor locus CDKN2A, which encodes INK4A and ARF [6]. These observations are consistent with our model predictions that overexpression of Bmi1 upregulates H2A monoubiquitination, which facilitates gene silencing.
Bmi1 is highly expressed in adult and fetal mouse and adult human hematopoietic stem cells (HSCs) [62], suggesting its important roles in maintaining the stem cell pool. Indeed, hematopoietic capacity is markedly reduced in Bmi1-knock out mice because of defective self-renewal ability of HSCs [62]. Downregulation of Bmi1 also results in decreased proliferation and self-renewal ability both in vitro and in vivo of neural stem cells [63]. Interestingly, Hosen et al. [64] showed that Bmi1 expression is high in HSCs and that Bmi1 is downregulated once the HSCs have differentiated into a particular lineage. Such differentiation in HSC cells may be enabled by the irreversible toggle switch characteristics of the Ring1B/Bmi1 system revealed in this paper.
It has been previously shown that shortly after murine erythroleukemia (MEL) cells were exposed to inducers of differentiation, significant increase of histone H2A ubiquitination occurred before returning to control levels [65]. Such transient pulse of monoubiquitinated H2A, which appears essential for MEL differentiation, can be explained by H2A excitability that follows gradual activation by some upstream regulators (e.g. Ring1B, Bmi1, external ligases or DUBs). The elevation of histone H2A monoubiqutination may initiate silencing of inhibitors of differentiation genes and thereby instigate differentiation. This implies that the Ring1B/Bmi1 excitable behavior can be implicated in cell-fate decision processes.
Our model can also helps explaining in vivo data involving knockout of E6-AP, the main exogenous ligase modifying Ring1B for degradation. E6-AP knockout mice display an elevated level of Ring1B and ubiquitinated histone H2A in various tissues, including cerebellar Purkinje neurons and liver.
When the number of Ring1B/Bmi1 molecules is not very high, random variation (noise) in these protein numbers influence signaling dynamics [48]. Within the bistable regime, depending on the stimulus history, intrinsic or extrinsic noise can lead to abrupt, random switches between low and high states of ubiquitinated H2A, potentially resulting in switching between On and Off states of H2A-silencing genes, respectively. This might partly contribute to transcriptional bursts of expression of these genes [66].
In summary, this paper presents a computational model of the Ring1B/Bmi1 ubiquitination system that reveals a high intrinsic complexity of the dynamics of Ring1B activity and H2A monoubiquitination and allows for direct experimental testing. Our findings provide a new perspective on (de)ubiquitination networks, which can display remarkably rich and complex dynamic behaviors.
|
10.1371/journal.pgen.1000852 | Proteasome Nuclear Activity Affects Chromosome Stability by Controlling the Turnover of Mms22, a Protein Important for DNA Repair | To expand the known spectrum of genes that maintain genome stability, we screened a recently released collection of temperature sensitive (Ts) yeast mutants for a chromosome instability (CIN) phenotype. Proteasome subunit genes represented a major functional group, and subsequent analysis demonstrated an evolutionarily conserved role in CIN. Analysis of individual proteasome core and lid subunit mutations showed that the CIN phenotype at semi-permissive temperature is associated with failure of subunit localization to the nucleus. The resultant proteasome dysfunction affects chromosome stability by impairing the kinetics of double strand break (DSB) repair. We show that the DNA repair protein Mms22 is required for DSB repair, and recruited to chromatin in a ubiquitin-dependent manner as a result of DNA damage. Moreover, subsequent proteasome-mediated degradation of Mms22 is necessary and sufficient for cell cycle progression through the G2/M arrest induced by DNA damage. Our results demonstrate for the first time that a double strand break repair protein is a proteasome target, and thus link nuclear proteasomal activity and DSB repair.
| Chromosome Instability (CIN) is a genome phenotype that involves changes in chromosome number or structure, and accounts for most malignancies. In this paper, we describe a screen to identify a set of novel CIN genes and find that proteasomal subunits represent a major functional group. We show that proteasome dysfunction affects CIN by impairing DNA double strand break (DSB) repair. Previous studies speculated that the proteasome is required to degrade one or more components of the DSB repair machinery; however, until now, no such target has been identified. Here we identify the previously described CIN gene MMS22 as a proteasomal target. We found that, as a result of DNA damage, Mms22 is ubiquitinated and recruited to chromatin. Mms22 then undergoes polyubiquitination and subsequent proteasome-mediated degradation. We also provide evidence that the degradation of Mms22 is important for the normal course of DNA repair and for exit from the G2/M arrest induced by DNA damage. Our results demonstrate for the first time that a DSB repair protein is a proteasome target, linking nuclear proteasomal activity and DSB repair. The mechanism of regulation of Mms22 may serve as a paradigm to understand how these additional proteins are regulated by the proteasome.
| Genomic instability is recognized as being an important predisposing condition that contributes to the development of cancer [1]. A major class of genome instability is Chromosome Instability (CIN), a phenotype that involves changes in chromosome number and structure. Studies in yeast have shown that multiple overlapping pathways contribute to genomic stability [2]. The current view is that most spontaneous chromosomal rearrangements result from DSBs created mainly during DNA replication as a result of broken, stalled or collapsed replication forks [3]. In eukaryotes, DSBs are repaired either by Homologous Recombination (HR) or by Non-Homologous End Joining (NHEJ) mechanisms. Defects in either repair pathway result in high frequencies of genomic instability [4]. The HR pathway utilizes a homologous sequence to faithfully restore the DNA continuity at the DSB [5]. In contrast, NHEJ is a mechanism able to join DNA ends with no or minimal homology [6]. Recent studies suggest a role for the proteasome in DSB repair pathways: The Sem1/DSS1 protein is a newly identified subunit of the 19S proteasome in both yeast and human cells. In yeast, Sem1 is recruited to DSB sites with the 19S and 20S proteasome particles, and is required for efficient repair of DSBs by HR and NHEJ [7]. Human DSS1 physically binds to the breast cancer susceptibility protein BRCA2, that plays an integral role in the repair of DSBs, and is required for its stability and function and consequently for efficient formation of RAD51 nucleofilaments [8],[9].
The Ubiquitin-Proteasome System (UPS) is the supramolecular machinery that mediates the ubiquitin-mediated proteolysis of damaged or misfolded proteins, or of short-lived regulatory proteins. The 26S proteasome comprises the 20S core particle (CP) and the 19S regulatory particle (RP), which represent the base and lid substructures, respectively [10]. Nuclear targets that are degraded by the proteasome include proteins involved in pathways critical for chromosome integrity. For example, degradation of polyubiquitinated mitotic cyclin and of the anaphase inhibitor Pds1/securin allow sister chromatids to dissociate at the onset of anaphase [for a review see [11]]. The protein levels of the tumor suppressor protein p53 are also subtly controlled by ubiquitin-mediated degradation [12].
Previous studies suggest that the amino-terminal ubiquitin-like (Ubl) domain of Rad23 protein can recruit the proteasome for a stimulatory role during nucleotide excision repair (NER) in S. cerevisiae. It has also been shown that the 19S regulatory complex of the yeast proteasome can affect nucleotide excision repair independently of Rad23 protein [13]. Other studies suggested a model for the regulation of Xeroderma Pigmentosum protein C (XPC), which plays a role in the primary DNA damage sensing in mammalian global genome NER. According to this model the ubiquitin-proteasome pathway has a positive regulatory role for optimal NER in mammalian cells, and appears to act by facilitating the recruitment of XPC to DNA damage sites [13]–[18].
A putative role for the proteasome at DSB sites could be to degrade components of the DNA damage response after their function is completed. However, so far no protein involved in DSB repair has been described as a direct target of the proteasome.
In this paper, we describe a systematic screen of a recently released collection of temperature- sensitive (Ts) yeast alleles [19], to find a set of novel CIN genes. The screen and subsequent analysis of individual mutants revealed that proteasomal subunits represent a major functional group, with an evolutionarily conserved role in CIN. We found that the CIN phenotype is associated with a failure of proteasomes to localize to the nucleus in viable cells, and show that proteasome dysfunction affects chromosome stability by impairing the kinetics of DSB repair. We also identify the DNA repair protein Mms22 as a proteasome target, and demonstrate that the impaired DNA repair phenotype can be attributed to a failure in the recruitment and subsequent degradation of ubiquitinated chromatin-bound Mms22.
In this study we expanded a recent screen for mutants affecting chromosome stability [19], by assessing the chromosome transmission fidelity (Ctf) phenotype (for details see Materials and Methods) in an additional 208 Ts strains. The functional distribution of the identified genes reveals that proteasome subunits are highly represented (Figure 1A and Table S1), we therefore decided to examine the mechanisms by which mutations in proteasome subunits cause CIN.
To test whether the CIN phenotype associated with proteasome dysfunction is evolutionarily conserved, we examined whether diminished proteasome subunit levels would cause a CIN phenotype in human cell lines. Small interfering RNAs (siRNAi) were used to target two human proteasome core (PSMA6 and PSMA4) and two lid subunits (PSMD4 and PSMD12) in the HCT116 cell line. To reduce the off-target effect, each experiment was performed with the two most effective siRNA duplexes (pointed by black arrows in Figure S1A). As shown in Figure 1B, relative to the controls, knockdown of Psma6, Psma4, Psmd4 and Psmd12 resulted in an increase in the frequency of cells with DNA contents greater than that of G2/M cells. Chromosome spreads after targeted knockdown of PSMA6 and PSMD12 established that the increase in DNA content is due to a dramatic increase in the number of cells with a total chromosome number above 46 (Figure 1C). Taken together, these results suggest that the proteasome lid and core components have a role in chromosome stability maintenance.
Previously it was established that the 26S proteasome localizes to the nucleus [20]. Here we confirmed the nuclear localization of the proteasomal lid and core subunits both in yeast and human cells (Figure 2A and 2B). The CIN phenotype caused by Ts alleles of proteasome subunits suggests that a nuclear function of the proteasome is impaired in the mutants. Sequence analysis of the rpn5 Ts allele reveals a single base pair insertion that introduces a premature stop codon, resulting in truncation of 39 amino acids at the C-terminus (Figure 2C). To analyze the localization of this truncated form, termed rpn5ΔC, GFP was fused in frame at its N-terminus. As a control, an identical N-terminal GFP fusion was constructed for the wt RPN5 gene (both expressed from a galactose-inducible promoter). The results show that whereas the control GFP-Rpn5 protein localizes predominantly to the nucleus, GFP-Rpn5ΔC localizes predominantly to the cytoplasm (Figure 2D). Similar nuclear mislocalization results were obtained for the mutated core subunit, Pup2Ts-GFP (Figure 2D). The mislocalization of the rpn5ΔC mutant protein indicates that the C-terminal domain (CTD) is important for Rpn5 nuclear localization in yeast.
Next we wanted to address the underlying defect in proteasome function that results in CIN. First we examined the proteasomal CIN mutants for sensitivity to Bleomycin (bleo) [21], and to hydroxyurea (HU)[22]. Mutants involved in DSB repair are usually sensitive to both drugs [23],[24]. We show that at semi-restrictive temperatures all proteasome mutants display varying degrees of sensitivity to these drugs (Figure 3A and Figure S1B). These results support a previous study showing that other proteasome mutants show sensitivity to DNA damaging agents [7]. Moreover, Ts alleles of rpn5ΔC and pup2, display a synthetic growth defect when either one is combined with rad52, a key factor in the DSB repair pathway [25] (for details see Figure 3B).
In support of a role for the yeast proteasome in DSB repair, previous ChIP experiments have provided evidence for the recruitment of the proteasome to DSB sites [7]. To test whether this phenomenon is conserved in mammalian cells, we performed Indirect ImmunoFluorescent (IIF) on Hela cells treated with Bleo, to look at the association of the RP subunit Psmd4 with DSB sites, represented by 53BP1 large foci (Figure 3D). In 183/200 53BP1 large foci counted, the Psmd4 focus was peripherally associated with the DSB site (Figure 3D). As a control we analyzed a similar number of unchallenged cells; in this case a significantly lower number of 53BP1 foci (44) could be detected, 21 of which were associated with a Psmd4 focus. As was previously shown [26],[27], these foci likely represent spontaneous DNA DSBs generated during DNA replication. In addition, we quantitated the signals of 50 Psmd4 foci that were associated with 53BP1 as a result of Bleo treatment; in 95% of the cases this signal was 5–10 times more intense than the average signal representing the Psmd4 foci not associated with 53BP1. These results provide evidence for an association of the proteasome with DSBs sites in human cells.
In order to examine the nature of the observed difference in DSB repair under proteasome dysfunction we studied the effect of the well-characterized proteasome inhibitor, MG132 [28] on the repair kinetics of a single defined chromosomal break in the yeast genome using the strain MK203 [29] (for more details see Figure 4A, and Materials and Methods). The strain used carried a mutation in the PDR5 gene, to prevent the cells from pumping the drug out of the cell [30].
As previously described [29], the control and MG132-treated cells arrest at G2/M three hrs after DSB induction. However, while the control cells exited from the arrest after 8 hrs, MG132-treated cells remained arrested even 10 hrs after DSB induction (Figure 4B). Southern blot analysis detected complete repair of the broken chromosome by 5 hrs following induction in control cells. In contrast, MG132-treated cells exhibited only partial repair of the DSB. Nine hours after transfer to galactose (which induces DSB repair), more than 30% of the cells still carry a broken chromosome (Figure 4C). At later times this proportion is reduced, probably due to outgrowth of cells with a repaired chromosome V.
We next examined the kinetics of formation of the gene conversion (GC) repair product (Figure 4D). In the control cells, GC can be detected 3.5 hrs after DSB induction, and the whole cell population was completely repaired by 6.5 hrs. In contrast, in MG132-treated cells only 70% of the cells exhibited repair 10 hrs after DSB induction (Figure 4D). Taken together, these results demonstrate that inhibition of proteasome activity affects the ability of yeast cells to carry out repair of a DSB, resulting in a prolonged cell cycle arrest. Moreover, MG132-treated cells also exhibit a higher level of CIN, measured using the a-faker-like (ALF) genome instability test [31] (Figure 2SA).
One possible explanation for the requirement of an active proteasome to complete the DSB repair is that the proteasome could be required to degrade one or more components of the DSB repair machinery. We looked for potential proteasome targets with a role in DSB repair. Such a target is expected to exhibit phenotypes that include both CIN (similar to that of proteasomal mutants), and sensitivity to DNA damaging agents, such as ionizing radiation or radiomimetic drugs such as methyl methanesulfonate (MMS). We recently used the Sacharomyces cerevisiae deletion collection to systematically screen for mutants exhibiting a CIN phenotype [31]. The mms22 mutant, which shows sensitivity to several DNA damaging agents that cause DSBs [32],[33] was among the mutants exhibiting the strongest CIN phenotype. To test whether Mms22 is a substrate of the proteasome, a strain carrying an inducible tagged protein (GAL1-HA-Mms22) was subjected to a promoter shutoff experiment. Figure 4E shows that under these conditions in wt cells Mms22p is degraded; in contrast, in the presence of MG132, the level of Mms22 protein stays high, and appears to be degraded to a lesser degree.
To further assess MMS22 function, we conducted a two-hybrid screen using Mms22 as the bait. This approach identified Rtt101/Cul8 as a protein that interacts with Mms22. We confirmed this interaction by IP. Consistent with a recent study [34], we concluded that Mms22 and Rtt101 proteins interact in vivo (Figure S2B and S2C). Rtt101 is one of four cullins in S. cerevisiae, with demonstrable ubiquitin ligase activity in vitro, but as yet no known substrate in vivo [35]. Based on the physical interactions seen between Mms22p and Rtt101, it has been suggested that Mms22 is a functional subunit of the Rtt101-based ubiquitin ligase [34]. Our results show that Mms22 is targeted by the proteasome; we therefore hypothesized that the turnover of Mms22 could be mediated by the Rtt101 E3 ubiquitin ligase complex. A promoter shut-off chase was used again to analyze the stability of the Mms22 protein in the presence or absence of the Rtt101 cullin. Figure 4F shows that Mms22p accumulated to a higher level during the induction period in rtt101 mutants in comparison to wt cells. To rule out the possibility that only the overexpressed proteins were being degraded by the proteasome, and to show that similar results can be observed in the context of endogenous levels of Mms22, we have performed cyclohexamide chase experiments in cells expressing Mms22-HA (Figure S2D). Western blot analysis revealed that, as observed in the GAL-driven overexpression experiments, Mms22 is also degraded in wt cells. Notably, Mms22 accumulated to higher levels in the presence of MG132, or in a Δrtt101 background. These results clearly demonstrate that the turnover of Mms22 is regulated by the Ubiquitin-Proteasome System (UPS), and mediated by the Rtt101 cullin.
mms22 cells show sensitivity to several DNA damaging agents that cause DSBs [32],[33]. To directly examine the kinetics of DSB repair in mms22 mutants, we used the MK203 system again, to compare repair kinetics of wt vs. mms22 cells following induction of a DSB. As seen in cells under proteasome inhibition, mms22 mutants show a delay in the disappearance of the broken chromosome compared to wt cells (Figure 4G). Additionally, and also similarly to MG132-treated cells, mms22 cells show a difference in gene conversion kinetics compared to wt cells (Figure S2E).
Substrates are usually targeted for degradation by the proteasome by polyubiquitination [10]. To test the ubiquitination levels of Mms22, we performed the experiments described in Figure 5A and Figure S2F. IP of Mms22-HA followed by immunoblotting reveals an additional band which migrates slower than Mms22-HA. This band most probably represents the ubiquitinated form of Mms22, as revealed by successive immunoblotting with an anti-Ubi antibody (Figure 5A). Additional proof that this band represents ubiquitinated Mms22 was obtained by successive immunoblotting with an anti-Myc antibody in a strain carrying a myc-tagged version of the ubiquitin protein (Figure 2SF). The results also show that Mms22 ubiquitination is RTT101 dependent. Finally, a 4 -fold increase in Mms22 ubiquitination was observed when cells were exposed to DNA damage, suggesting that the ubiquitination of Mms22 plays a functional role in DNA repair (Figure 5A).
To provide a rigorous in vivo demonstration that the ubiquitinated proteins observed by Western blotting were indeed a series of polyubiquitinated forms of Mms22, we performed the following experiment in which the 3HA and 6HA tagged versions of Mms22 were used in parallel. Cells were grown in the presence of MMS, and subjected to IP followed by immunoblotting with anti-HA. As expected from the results shown in Figure 5A, treatment with MMS led to an additional band which migrated more slowly than the band representing Mms22. Importantly, this band changed its electrophoretic mobility upon switching the tag on Mms22 from 3HA to 6HA, demonstrating unequivocally that it represents a specific in vivo modification of Mms22 (Figure 5B, left). This modification is indeed the specific ubiquitination of Mms22 as revealed by a similar electrophoretic shift of the bands that appeared following a successive immunoblotting with an anti-Ubiquitin antibody (Figure 5B, right). Alignment of the anti-HA and anti-Ubi antibody membranes, and the observed co-alignment of the electrophoretic shifts characteristic of the differentially tagged Mms22 protein species, indicates that the additional band that migrates slower than Mms22-HA is the mono-ubiquitinated form of Mms22.
Genome-wide genetic interaction results have shown that MMS22 clusters with RTT109, and ASF1 [36], two proteins required for histone H3 modification [37]. We therefore hypothesized that the ubiquitinatation of Mms22 may facilitate its recruitment to chromatin upon DNA damage. To test this idea we separated whole cell extracts (WCE) into soluble (SU) and chromatin-bound (CHR) fractions. Fractions were then subjected to immunoblotting using anti HA (Mms22-HA). The results clearly show that in unchallenged cells Mms22 is mainly present at the SU fraction (Figure 5C top). Treatment with MMS, however, leads to an enrichment of ubiquitinated Mms22 on the chromatin-bound fraction (Figure 5C middle). This enrichment was significantly reduced in the absence of RTT101 or RTT109 (Figure 5C bottom and Figure S3A).
Taken together, our results suggest that the ubiquitinated form of Mms22 on chromatin plays a functional role in dealing with DNA damage. A similar experimental approach was used to show chromatin enrichment of the proteasomal lid subunit Rpn5 upon exposure to MMS (Figure 3SB), which is consistent with ChIP analysis of proteasomal subunits at induced DSBs sites [7].
DNA damage induces the recruitment of both Mms22 and the Proteasome to chromatin, predicting that Mms22 degradation by the proteasome plays an important role in performing DNA repair. To test this idea we performed the experiment described in Figure 5D. MMS treatment resulted in G2/M arrest, and a chromatin fractionation assay revealed that Mms22 was recruited to chromatin in the presence, or in the absence, of MG132 (Figure 5D-2 and 5D-3). These results indicate that the recruitment of Mms22 to chromatin is proteasome-independent. The recruitment to chromatin is not cell cycle dependent, since a similar recruitment of Mms22 to chromatin was detected even when cells were kept in G1 during MMS treatment (data not shown). In contrast, the exit from the G2/M arrest following the removal of MMS was proteasome dependent, since only removal of MG132 from the medium led to the degradation of Mms22 from chromatin, which was associated with the exit from the G2/M arrest (compare 6D-5 vs. 6D-6 and 6D-7). To rule out the possibility that the prolonged exposure to MG132 and not MMS treatment led to the G2/M accumulation, we performed the control experiment described in Figure S3C. We show that samples released from the G1 arrest and constantly exposed to MG132 continued cycling normally in contrast to samples from a similar time point exposed to MMS+MG132 (compare Figure 5D-5 to Figure S3C).
Next we wanted to test whether the correlation between the accumulation of Mms22 on chromatin, and the failure to recover from cell cycle arrest upon DNA damage can be attributed (among other factors) to the specific accumulation of Mms22 in cells with defective proteasome activity. We therefore tested whether overexpression (OE) of Mms22 (which simulates the accumulation of Mms22 in proteasome mutants) also results in impaired recovery from DNA damage induced by MMS. While wt cells start to recover from the G2/M arrest 80 min after the removal of MMS from the medium (Figure 5E top), cells that overexpress Mms22 were still arrested even after 110 min (Figure 5E middle). Importantly, the removal of MMS together with Mms22 promoter shutoff (leading to the degradation of Mms22, data not shown), led to enhanced recovery from the G2/M arrest, when compared to cells still overexpressing Mms22 (Figure 5E middle versus bottom).
Having shown that degradation of Mms22 can promote exit from the G2/M arrest, we tested whether the specific degradation of Mms22 is sufficient for the exit. We created yeast strains carrying an allele of Mms22 (Mms22-T) that is expressed from its endogenous promoter and can be cleaved by the tobacco etch virus (TEV) protease [38],[39]. This protease can be conditionally expressed (for details see Figure 6A). Induction of the TEV protease leads to cleavage and inactivation of the Mms22-T protein (Figure 6A and 6B). When we conditionally expressed the protease in the presence of MG132 in cells arrested in G2/M as a result of MMS treatment, the cleavage of Mms22-T resulted in a clear release from the DNA damage-induced G2/M arrest (Figure 6C, compare top and bottom panels). Thus, degradation of Mms22 is essential for release from the cell cycle arrest induced by DNA damage.
We have shown that the degradation of Mms22 is essential for release from damage-induced cell cycle arrest. Next, we tested whether the prolonged G2/M arrest is also associated with impaired response to DNA damage. Using yeast strains tagged with fluorescent versions of three major components of the DNA DSB repair machinery: Mre11, Ddc2, Rad52, we conducted temporal analysis of focus formation following DNA damage (Figure 6D). Consistent with previous data [40] we show that Mre11 (a member of the MRX complex) is the earliest protein to form foci. Mre11 foci formation is followed by later recruitment of Ddc2 (the yeast orthologue of human ATR-interacting protein ATRIP) and the repair protein Rad52. We show (Figure 6D, left) that in wt cells, as Rad52 and Ddc2 were recruited, Mre11 foci disassembled. This disassembly was circumvented when cells were exposed to a proteasome inhibitor, and led to delayed and reduced focus formation of Rad52 (Figure 6D, middle). Remarkably, the specific degradation of Mms22 resulted in a clear disassembly of Mre11 foci and recovery of Rad52 foci (Figure 6D, right). Our results demonstrate that degradation of Mms22 is essential for the normal course of DNA DSB repair, and for the release from the cell cycle arrest induced by DNA damage.
We describe the first systematic screen of a recently released resource (still under development) consisting of Ts mutants of all essential yeast genes for which no Ts-allele had previously been isolated [19]. Among the 40 genes identified, 8 encoded proteasomal subunits. Genetic and biochemical analysis showed that CIN was associated with the failure of proteosomal subunits to localize to the nucleus, impaired kinetics of DSB repair, and failure to turnover the DNA repair protein Mms22 targeted for degradation by the proteasome.
Recent studies have suggested a role for the proteasome in the repair of DSB in yeast [7], and mammalian cells [41],[42]. In our current work, we show that mutations in the proteasome subunits rpn5ΔC and pup2, which cause nuclear mislocalization, are associated with impaired DSB repair. All other proteasomal Ts mutants tested were sensitive to drugs inducing DSBs, implying that the proteolytic activity of the proteasome is required for DNA repair. By examining the kinetics of DSB repair in cells treated with the proteasome inhibitor MG132, we obtained evidence for delayed kinetics of repair (Figure 4C). We showed that both the disappearance of the break as well as the kinetics of formation of the gene conversion product were delayed in treated cells compared to untreated cells (Figure 4D). As MG132-treated cells arrest in G2/M similarly to untreated cells, it is evident that checkpoint regulation due to DSB is not impaired in the treated cells. The delay in DSB repair suggests that proteasome activity might be required for the regulation of the DNA repair machinery.
A potential role for regulation of DSB repair by the proteasome in mammalian cells is supported by a recent study showing that proteasome inhibition affected the choice of HR repair pathways [41]. A different study showed that proteasome-dependent protein degradation substantially contributes to HR but not NHEJ [42]. It is tempting to speculate that the proteasome accumulates at sites of DSB, and that its proteolytic activity is required to degrade one or more components of the DSB repair machinery, or DNA damage response/repair proteins.
To date no protein involved in DSB repair has previously been described as a direct target of the proteasome. In this study, we identify Mms22, a protein required for efficient repair of DSBs (Figure 4G and Figure S2E), as a direct target of the proteasome degradation pathway (Figure 4E and 4F and Figure 2SD). Recently, Zaidi and colleagues [34] showed that Mms22 physically interacts with Rtt101, and suggested that Mms22 is a functional component of the SCFrtt101 ligase, perhaps as a substrate specificity factor. Although our studies do not address whether Mms22 is a subunit of SCFrtt101, we show clear evidence that Mms22 is a substrate of the SCFrtt101 and that proteasome-mediated turnover of Mms22 is important for the process of DNA repair.
The effect of MMS22 accumulation on the course of DSB repair (Figure 6D) suggests that Mms22 activity facilitates the recruitment of the HR machinery to DSBs. Histone modification occurs readily at sites of DSB or UV damage [43],[44] and it is becoming increasingly clear that proper chromatin handling is essential for successful repair. Indeed, we show that DNA damage results in Mms22 recruitment to the chromatin bound fraction (Figure 5C). Importantly, our results also show that recruitment of Mms22 to chromatin is not sufficient for the normal course of DNA repair, and that an essential step is a proteasome-mediated degradation of Mms22. These results thus identify for the first time a proteasome target that links proteasomal nuclear activity and DNA double strand break repair.
We propose the following model for the mechanism by which nuclear activity of the proteasome contributes to repair of DSBs. DNA damage results in a SCFrtt101 E3 ubiquitin ligase-dependent accumulation of the ubiquitinated form of Mms22 on chromatin that, as suggested above, plays a role in dealing with DNA damage. Subsequent degradation of ubiquitinated Mms22 by the proteasome is an important step in completion of the DNA repair process. Once Mms22 executes its function in DNA repair it becomes a target for degradation by the UPS, and is removed from chromatin. Failure to degrade Mms22 results in impaired DNA repair and prolonged cell cycle arrest. In support of our model, we show that reactivation of an inhibited proteasome results in degradation of the accumulated chromatin-bound Mms22, and in recovery from the G2 arrest induced by DNA damage (Figure 5D).
The synthetic genetic interaction that we describe for the proteasome and the rad52 mutant points to additional roles of the proteasome in DNA repair. Given its central role in protein degradation, it is indeed very likely that, in addition to Mms22, the proteasome regulates additional proteins involved in DNA repair. In this regard, proteasome inhibition in combination with DNA damage probably results in the accumulation of many proteins besides Mms22, which altogether may lead to the impaired recovery from the cell cycle arrest. We show, however, that specific accumulation of Mms22 by overexpression causes defects in recovery from DNA damage-induced G2/M arrest, whereas turnover of Mms22 after promoter shutoff allows recovery to occur (Figure 5E). Moreover, we also show that the specific degradation of Mms22 in the presence of proteasome inhibitor is sufficient for the exit from the DNA damage-induced G2/M arrest (Figure 6C). The arrest by itself was not affected by proteasome inhibition, which is consistent with the normal kinetics of foci formation of the checkpoint protein Ddc2. In contrast, proteasome inhibition affected the disassembly of Mre11, which in turn impaired the recruitment of the repair machinery, as demonstrated by the kinetics of Rad52 foci formation (Figure 6D). A similar phenotype was previously reported for Δsae2 mutants, supporting the notion that Sae2 is required for the transition from Mre11 binding to the recombinational repair function carried out by Rad52 [40],[45]. Our results suggest that degradation of Mms22 occurs at the same transition stage and that when this transition is impaired, cells can no longer proceed with the normal course of DNA repair. Suggestions about the possible activity of Mms22 at this transition stage comes from genetic interaction data [36]. In these studies, MMS22 clusters with RTT109, and ASF1, which are required for histone H3 acetylation [37]. These results suggest that Mms22, in association with Rtt109, and Asf1 are recruited to the sites of DNA lesions to modify their chromatin structure, perhaps facilitating DNA resection and recruitment of downstream-acting repair proteins such as Rad52. Recruitment of Rad52 in the form of foci depends on the removal Mms22 from DNA by the proteasome.
Taken together, we show that Mms22 is a proteasome target that links nuclear proteasomal activity and DSB repair. We believe that the CIN phenotype and impaired DNA repair caused by proteasome dysfunction can, in part, be attributed to the specific accumulation of Mms22. This idea is further supported by the observation that accumulation of Mms22 sensitizes the cells to DNA damaging agents, and results in CIN (Figure 2SA and Figure 3SD–3SF), and by previous studies showing that mutants in genes that play roles in DSB repair cause CIN phenotype in yeast and mammalian cells [31],[46]. It is likely that additional proteasomal targets important for genome stability await discovery. The mechanism of regulation of Mms22 may serve as a paradigm to understand how these additional proteins are regulated by the proteasome.
Yeast strains that were used for the CTF screen are the result of backcrossing the haploid Ts strain (MATa ura3Δ0 leu2Δ0 his3Δ1 lys2Δ0 (or LYS2) met15Δ0 (or MET15) can1Δ::LEU2-MFA1pr::His3 yfeg-ts::URA3), to the Donor strain SB1 (MATα ade2-101::NAT his3 ura3 lys2 can1Δ mfa1Δ::MFA1pr-HIS3 CFVII(RAD2.d)::LYS2), and selecting for a Lys+ spore clone (indicating the presence of CFVII(RAD2.d)::LYS2) resistant to ClonNAT (thus carrying the ade2-101 ochre mutation) and Ura+ (yfeg-ts::URA3) [for more details see [19]]. Other strains that were used in this study are listed in Table S2. The following strains were generated by crossing the indicated strains (in brackets), and selecting for the appropriate spores: SB162-(SB158xTs944), SB163-(SB160xTs944), SB220-(SB158xTs670), SB223-(SB160xTs670), SB258-(SB256xSB148), SB259-(SB256xSB147), SB175-(Ts602Xsb132). Yeast strains used for DSB repair assay are isogenic derivatives of strain MK203 (MATa-inc ura3::HOcs lys2::ura3::HOcs-inc ade3::GALHO ade2-1 leu2-3,112 his3-11,15 trp1-1 can1-100) [29],[47], a derivative of W303. In SB276, a TEV protease consensus cleavage site was introduced at position 2801 of Mms22 in the strain K9127 [39] by a two-step gene replacement.
Were preformed as previously described [48].
Was performed as previously described [49]. Cells were grown to O.D600-0.5 in 50 ml culture. Samples were spun down in 50 ml conical tubes for 5 min, resuspended in 3 ml of 100 mM PIPES/KOH pH 9.4, 10 mM DTT, 0.1% Na-Azide, and incubated for 10 min at RT. Samples were then spun for 2 min. Supernatant was aspirated off, and samples were resuspended in 2 ml of 50 mM KPi, pH 7.4, 0.6M Sorbitol, 10 mM DTT, and transferred to 2 ml microfuge tubes. 10 ul aliquot was then diluted in 990 ul H2O in a cuvette. 4 ul of 20 mg/ml Zymolase T-100 was added for 10 min, in 37°C water bath (tubes were gently inverted every 2–3 minutes). After about 1 min, 10 ul aliquot was used to measure the O.D600 (for hypotonic lysis). The O.D of the 1∶100 dilutions after spheroplasting was less the 10% of the value before. From this point on everything was done in a cold room. Tubes were spun for 1 min, cells were then washed with 1 ml of 50 mM HEPES/KOH pH 7.5, 100 mM KCl, 2.5 mM MgCl2, 0.4M sorbitol. Tubes were spun for 1 min, and resuspended in equal pellet volume EB (around 80 ul). 1/40 volume 10% Triton X-100 (0.25% final, e.g. 4 ul for 160 ul suspension), was added and cells were incubated for 3 min for lysis on ice, (vortexed occasionally). This sample represents the whole cell extract (WCE). 20 ul sample was removed and 20 ul of SDS loading buffer was added (WCE). 100 ul EBX-S was prepared in separate microfuge tubes. 100 ul of whole cell extracts were laid onto the EBX-S, and microfuge tubes were spun for 10 min. The resulted fractions represent a white chromatin pellet (CHR), the clear sucrose layer, and above a yellow supernatant fraction (SUP). 20 ul of SDS loading buffer was added to 20 ul of the SUP fraction (SUP). The rest of supernatant and sucrose buffer were then aspirated. The chromatin pellet was resuspended in 100 ul EBX, and spun for 5 min. Supernatant was aspirated off, and chromatin pellet was resuspended again in 100 ul EBX. 20 ul sample was then removed and added to 20 ul of SDS loading buffer (CHR). EB: 50 mM HEPES/KOH pH 7.5, 100 mM KCl, 2.5 mM MgCl2, 1 mM DTT, 20 ug/ml leupeptin, 2 mM benzamidine, 2 ug/ml aprotinin, 0.2 mg/ml bacitracin, 2 ug/ml pepstatin A, 1 mM PMSF (add it just before use). EBX: EB +0.25% Triton X-100. EBX-S: EBX +30% Sucrose.
Cells were plated onto sterilized glass coverslips so that they were 50% to 80% confluent on the following day. Subsequent to fixation for 5 min at 25°C with fresh 4.0% paraformaldehyde, cells were permeabilized with phosphate-buffered saline (PBS; pH 7.5) containing 0.5% Triton X-100 for 5 min. Cells were washed twice with PBS and subjected to sequential series of 30-min incubations with appropriate primary and secondary antibodies. Wash steps consisted of a single wash with PBS containing 0.1% Triton X-100 and two washes with PBS. The following primary antibodies were used: anti Psmd4 (Abcam ab20239), Psma1 (Abcam ab3325), anti 53BP1 (Abcam ab21083) and anti γH2AX (Abcam ab18311). Primary antibodies were recognized with appropriate mouse or rabbit secondary antibodies conjugated with either Alexa-fluor 488 or Cyanin-3 (Cy-3) (MolecularProbes, and the Jackson ImmunoResearch Laboratories respectively). Coverslips were mounted onto slides containing approximately 10 µl of a 90% glycerol-PBS–based medium containing 1 mg/mL parapheylenediamine and 0.5 µg/ml DAPI. Image acquisition and processing was preformed as detailed previously [50] using a Zeiss Axioplan 2 digital imaging microscope equipped with a ×63 (1.3 numerical aperture) and a x100 (1.4 numerical aperture) plan-apochromat oil-immersion lens, a Coolsnap HQ cooled charge-coupled device camera (Roper Scientific), and Metamorph imaging software (Universal Imaging Corp).
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.